essaysondevelopmentpolicyandthe …936631/...the results indicate that contracts with lower interest...

Essays on Development Policy and thePolitical Economy of Conflict

Miri Stryjan

ii

© Miri Stryjan, Stockholm, 2016

ISBN 978-91-7649-451-6ISSN 0346-6892

Cover Picture: Rain coming in over Moroto town© Miri Stryjan, 2015.Portrait photo taken by Anna Sandberg, 2016.

Printed in Sweden by Holmbergs, Malmö 2016

Distributor: Institute for International Economic Studies

iii

Doctoral DissertationDepartment of EconomicsStockholm University

AbstractElectoral Rules and Leader Selection: Experimental Evidence from Ugan-dan Community Groups. This paper studies leadership selection in communitygroups. Despite a large body of work documenting how electoral systems affect pol-icy outcomes, less is known about their impact on leader selection. We compare twotypes of participatory decision making in Ugandan community savings groups: voteby secret ballot and open discussion with consensus. Random assignment of electoralrules allows us to estimate the causal impact of the rules on leader types and socialservice delivery. We find that vote groups elect leaders more similar to the averagemember while discussion group leaders are positively selected on socio-economic char-acteristics. Further, the dropout rates are significantly higher in discussion groups,particularly for poorer members. After 3.5 years, vote groups are larger in size andtheir members save less and get smaller loans. We conclude that the secret ballot votecreates more inclusive groups while open discussion groups are more exclusive andfavor the economically successful. The appropriate method for leader selection thusultimately depends on the objective and target group of the program. Our findings offerimportant contributions to the literature on leader selection and to the understandingof public service delivery in developing countries.

Preparing for Genocide: Community Meetings in Rwanda. How do po-litical elites prepare the civilian population for participation in violent conflict? Weempirically investigate this question using sector-level data from the Rwandan Geno-cide in 1994. Every Saturday before 1994, Rwandan villagers had to meet to workon community infrastructure, a practice called Umuganda. The practice was highlypoliticized and, according to anecdotal evidence, regularly used by the local politicalelites for spreading propaganda in the years before the genocide. This paper presentsthe first quantitative evidence of this abuse of the Umuganda community meetings. Toestablish causality, we exploit cross-sectional variation in meeting intensity induced byexogenous weather fluctuations. We find that an additional rainy Saturday resulted ina five percent lower civilian participation rate in genocide violence. We find no resultsfor other weekdays. Moreover, this result is entirely driven by places under the controlof the pro-genocide Hutu parties. In the few places with the pro-Tutsi minority in

iv

power the effects are reversed. The results pass a number of indirect tests regardingthe exclusion restriction as well as other robustness checks and placebo tests.

Selection into Borrowing: Survey Evidence from Uganda. In this paper,I study how modifications to the standard credit contract affect loan demand andselection into borrowing, using a representative sample of micro enterprises in urbanUganda. Despite widespread enthusiasm about microfinance as a tool in alleviatingpoverty, recent evaluations of microfinance-initiatives have found the long run impacton firm growth and the welfare of borrower-households to be limited. Existing studiesfocus on present or previous borrowers, and can therefore provide only limited insightinto how contractual changes would affect credit demand and investment behaviorthrough changes in the composition of the borrower pool. I study loan attitudes in arepresentative sample of 925 entrepreneurs, most with no experience of borrowing, incore sectors within both retail and manufacturing. Hypothetical loan demand ques-tions are used to test whether firm owners respond to changes in loans’ contractualterms and whether take-up varies by firms’ risk type and firm owner characteristics.The results indicate that contracts with lower interest rates and less stringent col-lateral requirements are likely to attract less risky borrowers. This is especially trueamong manufacturing sectors. The findings are robust across different ways of definingriskiness and suggest that there is scope for improvement of standard financial contractterms.

Credit Contract Structure and Firm Growth: Evidence from a Random-ized Control Trial.We study the effects of credit contract structure on firm outcomesamong small- and medium- sized firms. A randomized control trial was carried out todistinguish between some of the key constraints to efficient credit use connected tofirms’ business environment and production function, namely (i) backloaded returns;(ii) uncertain returns; and (iii) indivisible fixed costs. Each firm was followed for theone-year loan cycle. We describe the experiment and present preliminary results fromthe first 754 out of 2,340 firms to have completed their loan cycle. Firms offered agrace period on their repayments early in the loan cycle have higher profits and higherhousehold income than firms receiving a grace period later on as well as the controlgroup. They also increased the number of paid employees and reduced the number ofunpaid employees, an effect also found among firms that received a cash subsidy atthe beginning of the loan cycle. We discuss potential mechanisms behind these effects.

v

AcknowledgmentsHow come it takes so long to do a PhD? I mean, how hard can it be? Countless

times in the past years I have been asked various flavors of that question. I have oftenfelt intimidated by it despite understanding that the person asking usually has no wayof knowing the struggles involved in completing a dissertation. To make matters worse,I have been asking myself the same question often enough. So why, indeed, does it takeso long to complete a PhD? In my case, the coursework and rigidity of the first yearof the doctoral program killed most of my inspiration and self-esteem. Finding themagain, and learning to ask relevant questions and sticking to ideas and projects hasbeen a long process. My PhD years have also involved living, working and studyingon three different continents - something which required much learning and digestionalso of issues not directly related to my research. I am very grateful to everyone whohas accompanied and helped me on this journey.

First and foremost, I want to thank my main advisor Jakob Svensson, who intro-duced me to the field of development economics and without whom the PhD wouldprobably have taken a very different turn. Jakob’s support and input during the workwith the chapters of the dissertation has been invaluable, especially during the lastyear of the PhD including the job market process. It is also thanks to Jakob that Igot the opportunity to work in Uganda and get my first experiences of practical fieldwork. Secondly I want to thank my co-advisor Andreas Madestam who invited me onboard on another project in Uganda and thereby gave me the opportunity to spendmore time in the field, which gave rise to additional projects. Living and working for12 months in Uganda and East Africa has shaped all the projects that are included inthis dissertation. Andreas has both been my advisor and my coauthor in the past fewyears and has taught me a lot about research, paper writing and how to stay optimisticin a very stressful environment.

Two people had a decisive impact on my choice to pursue a research career, andthose are Jan Pettersson and Lena Nekby, who supervised my bachelor and mastertheses, respectively. Thank you Janne for all your support and encouragement duringthe thesis writing and the first steps of the PhD. To Lena, who very sadly passed awaytwo years ago, I am also deeply thankful for believing in me and my ideas, and forsupporting me as a mentor during the first half of my PhD.

Over the past four years I have had the privilege to work with a number ofco-authors and I want to thank Evelina Bonnier, Erika Deserranno, Selim Gulesci,Francesco Loiacono, Andreas Madestam, Jonas Poulsen, Munshi Sulaiman and Thorsten

vi

Rogall for fruitful collaborations. A special thanks to Selim, who has taught me a lotover the past three years of working together. I have also been fortunate to work withthe NGO BRAC in Uganda, where I have had my office on and off for almost one yearand learned a lot about practical development work in the process, and I am thankfulto everyone who made this possible. I am also grateful to Ted Miguel who invited meto visit one year at the economics department of UC Berkeley.

For most of the PhD I have been based at Stockholm University where a numberof people have had an extra impact on my life and research. I am immensely thankfulto Anna Tompsett for her supportiveness and many concrete suggestions around myjob market process and paper. At the IIES I am also extra grateful to Ingvild Almås,Konrad Burchardi, Tom Cunningham and Masayuki Kudamatsu for many helpful andinteresting discussions and to Jon de Quidt and Kurt Mitman for encouragement andhelp during my last year at the IIES and during the job market, to Anna Sandbergfor great feedback on two of my papers and to Torsten Persson for his useful input,advice and support on the job market. I want to send a big thank you to the admin-istrative staff of the IIES: Annika Andreasson, Viktoria Garvare, Christina Lönnblad,Åsa Storm and Hanna Weitz, and of the Department of Economics: Ingela Arvids-son, Anne Jensen, Anita Karlsson, and Audrone Mozuraitiene. A special thanks toAnita for providing a compass in the arbitrariness surrounding the rights of a PhDstudent, Christina for language work on the thesis, Annika for her assistance and en-couragement during the job application process and Viktoria for help with the thesislayout.

The PhD is definitely a lonely experience, but nevertheless my fellow PhD col-leagues during these years have meant a lot. Sara Fogelberg and Manja Gärtner whostarted the doctoral program together with me have been especially important. Sara,thank you for your endless support, loyalty and great sense of humor - and for all ourgreat times at the staff gym. Manja, the discussions with you about economics andlife, and your wonderful sarcasms have lifted me up countless times. During the firstyear, the work in the "coal mine" was made more bearable by the great company ofSara and Manja as well as of PO Robling, Laurence Malafry, Erik Prawitz, ShuheiKitamura, Yangzhou Yang and Daniel Hedblom, and during the last year, the jobmarket period was made easier and even fun thanks to the company of Audi Bal-trunaite, Mounir Karadja, Shuhei, Niels-Jakob Harbo Hansen and Andrea Guariso.Other colleagues in Stockholm and Uppsala have also helped put a silver lining onthe PhD years: Lotta Boström, Linnea Wickström Östervall, Emma von Essen, Maria

vii

Cheung, Alex Schmitt, Abdulaziz Behiru Shifa, Hanna Mühlrad, Erik Lundin, EvelinaBonnier, Anna Aevarsdottir, Mathias Iwanowsky and Karl Harmenberg, to name onlya few. Jonas Poulsen, our interesting discussions about development and academiahave cheered me up many times. During my long stays in Uganda I was especiallyhappy for the company of Benedetta Lerva who showed me the real Kampala, andtaught me to bargain with the boda drivers, Vittorio Bassi with whom I had manyrewarding conversations about our parallel research projects, and Mozammel Huq whoinvited me to share his Kampala home and made me feel less like an outsider.

My years as a doctoral student have also been shaped by the people who haveshared my everyday life outside of work and university. Some of the friends from beforethe PhD journey started are surprisingly still around. I am grateful to Elsa, Palmina,Jocke and Itay for being there throughout, despite my sometimes very distant mood.I am so fortunate to have you as my friends. I also want to thank all my lovelyhousemates from Stavsund on Ekerö where I have lived for the past four years, withwhom I have cried and laughed and who have provided their perspectives from outsideof economics and academia. Especially to Virlani, Daniel, Gabriella, Sofie, Mario, Lisaand Clara. Yotam, who has accompanied me and made me happy during the last crazyperiod of the PhD also deserves a very special thank you.

Finally, I want to thank my family. To my dear sister Noa, thank you for alwaysbeing there and for your wise words that have helped me so many times when the PhDjourney seemed endless. And to my parents, Signe and Yohanan Stryjan - thank youfor encouraging me to be curious and open to impressions and knowledge, for teachingme languages to understand the world and independence to dare to take it on. I knowthat without you I would not have been where I am today.

Stockholm, July 2016Miri Stryjan

Contents

1 Introduction 1References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Electoral Rules and Leader Selection: Experimental Evidence fromUgandan Community Groups 92.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2 Background: Karamoja and BRAC Uganda . . . . . . . . . . . . . . . . 162.3 Setup of saving groups and leader selection methods . . . . . . . . . . . 182.4 Conceptual framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.5 Data and Descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.7 Welfare effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.8 Discussion and Concluding remarks . . . . . . . . . . . . . . . . . . . . 41References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59Appendix 2: Variable construction details . . . . . . . . . . . . . . . . . . . . 67

3 Preparing for Genocide: Community Meetings in Rwanda 693.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853.6 Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893.7 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 91References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

ix

x CONTENTS

Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

4 Selection into Borrowing: Survey Evidence from Uganda 1154.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154.2 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . . 1204.3 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1224.4 Survey methodology and Data . . . . . . . . . . . . . . . . . . . . . . . 1244.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1354.6 Validation checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1434.7 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 146References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159Appendix 2: Loan contract variations . . . . . . . . . . . . . . . . . . . . . . 168

5 Credit Contract Structure and Firm Growth: Evidence from a Ran-domized Control Trial 1695.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1695.2 Conceptual framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 1735.3 Experimental design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1755.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1805.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1835.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1875.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Sammanfattning 209Referenser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216

Chapter 1

Introduction

This thesis consists of four self contained chapters that all revolve around the themes

of development policy and political economy aspects of the implementation of develop-

ment programs. Chapters 2, 4 and 5 analyze development initiatives related to financial

inclusion of the poor while chapter 3 deals with a state-led development project with

political undertones.

In developing countries, NGOs and other external agents often assume responsi-

bility for the provision of critical services that the state fails to provide, and for the

delivery of financial services inadequately supplied by banks and other formal institu-

tions (Baland et al., 2011; Casey et al., 2012; Grossman; 2014). In the past decades,

leading NGOs have increasingly favored development projects that involve the local

community in decision making (Mansuri and Rao, 2012). This is believed to increase

the legitimacy and long-term sustainability of projects. Evaluations of such projects

point to several advantages of direct local participation as compared to central de-

cision making. However, so far, we know very little about the relative effectiveness

of different types of direct participation. Moreover, with the exception of Grossman

(2014), previous studies are silent on the subject of leaders, despite that the mode

of governance in community-led development projects can be crucial for their service

delivery and outreach.

Chapter 2, Electoral Rules and Leader Selection: Experimental Evidence

from Ugandan Community Groups, studies self governance and leadership in

1

2 INTRODUCTION

NGO-initiated local savings groups for young women in an impoverished area of

Uganda. More specifically, it studies how the design of electoral rules determines lead-

ership selection in and performance of the savings groups. Despite a large body of work

documenting how electoral systems affect policy, less is know about their impact on

leader selection. Moreover, exogenous variation in electoral rules is notoriously hard to

find, in particular for real world contexts. We randomly assigned Ugandan community

saving groups to use one of two distinct methods when selecting leaders for the first

time: vote by secret ballot or open discussion with decision making by consensus. Ran-

dom assignment allows us to estimate the causal impact of the rules on leader types

and on measurable outcomes resulting from the leaders’ implemented policy: member

retention, savings and loans. We find that vote groups elect leaders more similar to the

average member while discussion group leaders are richer and more educated than the

average member. Further, dropout rates are significantly higher in discussion groups,

particularly for the initially poorer members. After 3.5 years, vote groups are larger in

size and their members save less and get smaller loans than discussion group members.

We thus find that the leader election rule affects leader types and community group

outcomes, with secret ballot vote creating more inclusive groups while open discussion

leads to lower financial inclusion of society’s poorest members. Our findings are con-

sistent with elite capture being higher in discussion groups, leading to outcomes less

representative of the preferences of the average member. This is in line with studies

of the introduction of the secret ballot (Baland and Robinson (2008) and Hinnerich

and Pettersson-Lidbom (2014)). Given the crucial role played by community groups

in delivery of many public and financial services in low-income contexts, our study has

policy implications for public service delivery in developing countries.

Community meetings and civic organizations are widely believed to foster social

capital by providing arenas for people to meet, exchange ideas, solve free-rider prob-

lems, and create public goods (Grootaert and van Bastelaer, 2002; Guiso et al., 2008;

Knack and Keefer, 1997; Putnam, 2000). This view partly motivates the increasing

focus of development agencies on "community driven" development projects, in which

3

deliberative forums and grass root participation play a central role (see Mansuri and

Rao (2012) for a recent overview). A recent literature shows that social forums and

civic organizations can also serve to enforce ties within social groups, and increase

tensions between them, rather than providing forums for bridging between members

from different social groups, thereby highlighting a more destructive potential of such

forums (Satyanath et al., 2015).

Chapter 3, Preparing for Genocide: Community Meetings in Rwanda, re-

lates to this work by studying a very different kind of development program, that due

to its political nature had devastating consequences. The practice of mandatory com-

munity work has been present in Rwanda since pre-colonial times and similar practices

existed during the early post-colonial period also in other East and Central African

countries (Guichaoua, 1991). During the period of 1973-1994, the mandatory com-

munity work became a state policy. Every Saturday, Rwandan villagers had to meet

to work on community infrastructure, a practice called Umuganda. The practice was

motivated by development arguments, but was also highly politicized and, according

to qualitative evidence from scholars such as Straus (2006) and Verwimp (2013), reg-

ularly used by the local political elites for spreading propaganda in the years before

the genocide. This paper presents the first quantitative evidence of this (ab)use of

the Umuganda community meetings. Identifying the causal effect of these meetings

on participation in genocide is difficult for two reasons. First, we lack data on the

number of people participating in Umuganda or the number of meetings taking place

in a given locality. Second, even if that data existed our estimates would likely suf-

fer from omitted variable bias. To establish the causal link between meeting intensity

and participation in genocidal violence we therefore exploit cross-sectional variation

in meeting intensity induced by exogenous weather fluctuations. The assumption that

we make is that when it rains heavily, the community meeting is either cancelled or

less intensive. Using daily rainfall data from the period 1984-1998 and sector-level

prosecution data from the Rwandan Genocide in 1994, we find that an additional

rainy Saturday resulted in a five percent lower civilian participation rate in genocide

4 INTRODUCTION

violence. We find no results of rainfall on other weekdays on genocide participation.

Moreover, this result is entirely driven by localities that were governed by the pro-

genocide Hutu parties. In the few places with the pro-Tutsi minority in power, the

effects are reversed. These results indicate that Umuganda meetings were indeed used

as an arena to mobilize and prepare civilians for the genocide. Despite the specific

geographical focus of this paper, we argue that examining the possibly negative effect

of these community meetings is of more general importance. While the attitude to

Umuganda and similar initiatives is generally positive, we show evidence of a "dark

side" to these community meetings where social capital does not bridge the societal,

ethnic divides but rather enforces bonding within the different ethnic groups. Under-

standing this process is even more important since, despite its history, Umuganda was

formally reintroduced in Rwanda in 2008, and similar practices have been installed in

Burundi and are have recently been proposed in Kenya.1

One of the most widely praised forms of development aid in the past decades

is microfinance. Microcredit and the broader concept of microfinance became well

known to the general public as Grameen bank and its founder Mohammad Yunus were

awarded the Nobel Peace Price in 2006. The idea behind microfinance is that small

loans can help poor people improve their livelihood through small-scale commercial

activity. As Amendariz de Aghion and Morduch (2005) write in their book about

microfinance: "Microfinance presents itself as a new market-based strategy for poverty

reduction, free of the heavy subsidies that brought down large statebanks. In a world

in search of easy answers, this win-win combination has been a true winner itself".

Despite widespread enthusiasm about microfinance as a tool for poverty-alleviation,

recent evaluations of microfinance initiatives have, however, found its long run impact

on firm growth and the welfare of borrower-households to be limited (Banerjee et al.,

2015). Chapters 4 and 5 of this thesis examine possible ways in which changes to the

standard microfinance contract could lead microfinance to better fulfill its promised

1For details about the Kenyan case, see Daily Nation (March 2016).

5

objective of business growth. Both studies focus on micro, small and medium sized

enterprises and concern individual loans.

Chapter 4, Selection into Borrowing: Survey Evidence from Uganda, re-

ports the results of a survey that elicits loan demand among a representative sample

of firm owners in urban Uganda. A literature in credit contract theory shows that

raising the price of credit (the interest rate) can lead to either advantageous selection

effects (Stiglitz and Weiss, 1981) or adverse selection effects (De Meza and Webb,

1987), in terms of the likelihood of project success, while increasing the collateral size

will lead to advantageous selection (Stiglitz and Weiss, 1981; Wette, 1983). Examining

the selection into microfinance is particularly relevant, as this market is characterized

by credit rationing partly due to asymmetric information. Existing studies of micro-

finance focus on individuals or firms that are already borrowers, and can therefore

provide only limited insight into how contractual changes would affect credit demand

and investment behavior through changes in the composition of the borrower pool. I

study loan attitudes among a representative sample of entrepreneurs, most with no

experience of borrowing, in core sectors within both retail and manufacturing. Hy-

pothetical loan demand questions are used to test whether firm owners respond to

changes in loans’ contractual terms and whether take-up varies by firms’ risk type and

firm owner characteristics. The results indicate that contracts with lower interest rates

or with less stringent collateral requirements are likely to attract less risky borrowers,

in terms of both stated risk behavior and the riskiness of their business environment.

This is true also when controlling for wealth. These results are more pronounced among

manufacturing firm owners, something which is likely to be explained by differences

in available investment options. Less wealthy firm owners are more likely to borrow

if collateral rates are lowered. The results are robust across different ways of defining

riskiness and suggest that there is scope for improvement of standard financial contract

terms. Chapter 5, Credit contract structure and firm growth: Evidence from

a randomized control trial, studies the effects of credit contract structure on firm

outcomes among small- and medium- sized firms in Uganda. We build on recent work

6 INTRODUCTION

which suggests that take-up and effectiveness of microfinance may improve if contrac-

tual terms are changed (Field et al., 2013; Karlan and Zinman, 2008). A randomized

control trial was carried out to distinguish between some of the key constraints to

efficient credit use connected to firms’ business environment and production function,

namely (i) backloaded returns; (ii) uncertain returns; and (iii) indivisible fixed costs.

Firms that participated in the experiment had been approved for borrowing from our

partnering NGO and, as part of our experiment, received rebates that subsidized the

equivalence of two out of 12 monthly repayments during their one year loan cycle. The

findings presented in Chapter 5 are preliminary results from the first 754 out of 2,340

firms to have completed their loan cycle. We find that firms that were given a 2-month

grace period at the beginning of the loan cycle increased their profits and household

income relative to firms that received a rebate later in the loan cycle, and to the control

group. They also increased the number of paid employees, while decreasing the number

of unpaid ones, but wage expenditures did not increase in accordance. Further, the

households of firm-owners in the early grace-period group started significantly more

new household-owned firms than the households of firm-owners that received a rebate

later in the loan cycle, and the control group. Firms that were offered a flexible grace

period scheme, in which they were free to skip repayments in any 2 months of their

choice, predominantly chose to use these rebates in the first months of the loan cycle.

These findings provide some support for backloadedness of returns being a more im-

portant constraint than the uncertainty of returns. Firms that received a cash subsidy

at the start of the loan cycle increased their number of employees relative to the control

group, and they also increased their wage costs. To the extent that this implies that

they hired higher quality workers, which can be seen as an indivisible investment, this

finding provides suggestive evidence for the importance of indivisible costs hampering

investments.

REFERENCES 7

References

Amendariz de Aghion, B. and Morduch, J. 2005. The Economics of microfinance:

Cambridge, MA: MIT Press.

Baland, J. M. and J. A. Robinson. 2008 Land and Power: Theory and Evidence

from Chile. The American Economic Review, 1737-1765.

Baland, J.M., R. Somanathan and L. Vandewalle. 2011 Socially disadvantaged

groups and Microfinance in India. University of Namur, Department of Economics

Working Paper no. 1117.

Banerjee, A., Karlan, D. and Zinman, J. 2015. Six randomized evaluations of

microcredit: introduction and further steps. American Economic Journal: Applied Eco-

nomics, 7(1), pp.1-21.

Casey, K., R. Glennerster and E. Miguel. 2012. Reshaping Institutions: Evidence

on aid impacts using a preanalysis plan. The Quarterly Journal of Economics, 1755,

1812.

Daily Nation (retrieved on March 5, 2016). Emuhaya MP Seeks to Introduce Manda-

tory Community Cleaning Day, http://allafrica.com/stories/201603030209.html.

DeMeza, D. andWebb, D.C. 1987. Too much investment: a problem of asymmetric

information. The quarterly journal of economics, pp.281-292.

Field, E., Pande, R., Papp, J. and Rigol, N. (2013. Does the classic micro-

finance model discourage entrepreneurship among the poor? Experimental evidence

from India. The American Economic Review, 103(6), pp.2196-2226.

Grossman, G. 2014. Do Selection Rules Affect Leader Responsiveness? Evidence

from Rural Uganda. Quarterly Journal of Political Science 9.1: 1-44.

Grootaert, C. and T. van Bastelaer. 2002. Understanding and Measuring Social

Capital: A Multi-Disciplinary Tool for Practitioners, Washington, World Bank.

Guichaoua, A. 1991. Les Travaux Communautaires en Afrique Centrale, Revue Tiers

Monde, t.XXXII, n. 127, July-September, pp. 551-573.

8 INTRODUCTION

Guiso, L., Sapienza P. and L. Zingales. 2008. Alfred Marshall Lecture: Social

Capital as Good Culture, Journal of the European Economic Association, 6(2-3), pp.

295-320.

Hinnerich, B. T., and P. Pettersson-Lidbom. 2014. Democracy, Redistribution,

and Political Participation: Evidence From Sweden 1919-1938. Econometrica, 82(3),

961-993.

Karlan, D.S. and Zinman, J. 2008. Credit elasticities in less-developed economies:

Implications for microfinance. The American Economic Review, pp.1040-1068.

Knack, S. and P. Keefer. 1997. Does Social Capital Have an Economic Payoff? A

Cross-Country Investigation, Quarterly Journal of Economics, 112(4), pp. 1251-1288.

Mansuri, G. and V. Rao, 2012. Localizing development: does participation work?.

World Bank Publications.

Putnam, R. D. 2000. Bowling Alone, Free Press, New York.

Satyanath, S., Voigtlaender, N. and H.J. Voth. 2015. Bowling for Fascism: Social

Capital and the Rise of the Nazi Party, NBER Working Paper no. 19201.

Straus, S. 2006. The Order of Genocide: Race, Power, And War in Rwanda, Cam-

bridge University Press, 1 edition.

Stiglitz, J.E. and Weiss, A. 1981. Credit rationing in markets with imperfect in-

formation. The American economic review, 71(3), pp.393-410.

Verwimp, P. 2013. Peasants in Power: The Political Economy of Development and

Genocide in Rwanda, Springer, Heidelberg.

Wette, H. C., 1983. Collateral in credit rationing in markets with imperfect infor-

mation: Note. Wette, H.C., 1983. The American Economic Review, 73(3), pp.442-445.

Chapter 2

Electoral Rules and Leader Selection:

Experimental Evidence from Ugandan

Community Groups∗

2.1 Introduction

Leader characteristics and representativeness is widely believed to be important for

the way political entities and organizations perform (Besley, 2005). Numerous studies

put forward links between electoral rules and policy outcomes (see Cox, 1997; Huber et

al., 1993; Myerson 1993; Persson and Tabellini, 2000; 2005; Hinnerich and Pettersson-

Lidbom, 2014). A less explored subject is the role of electoral rules for selecting leaders

with certain characteristics, and how these characteristics impact on the quality of

policy outcomes. As noted by Beath et al. (2014), exogenous variation in electoral

∗This paper is co-authored with Erika Deserranno and Munshi Sulaiman. We would like to ex-tend our gratitude to the staff of the BRAC Uganda Karamoja initiative, in particular to AlbertSsimbwa and the BRAC Uganda Research and Evaluation Unit. We also thank Emanuele Brancatifor excellent research assistance. This paper has benefited from discussions and helpful commentsfrom Ingvild Almås, Tessa Bold, Jonathan de Quidt, Selim Gulesci, Andreas Madestam, TorstenPersson, Johanna Rickne, Anna Sandberg, David Strömberg, Jakob Svensson and Anna Tompsett,as well as from seminar participants at the IIES, the OXDEV Workshop 2015, Hebrew University,Ben Gurion University, CERGE-EI, Universidad de los Andes and Warwick. Miri Stryjan is gratefulfor funding from Handelsbanken’s Research Foundation, the Mannerfelt Foundation and the NordicAfrica Institute.

9

10 ELECTORAL RULES AND LEADER SELECTION

rules is notoriously hard to find, in particular for real world contexts.

This paper provides causal evidence of how leader electoral rules affect the types of

leaders selected and, in turn, the quality of policy outcomes. In an experimental setting,

we are able to isolate the impact of these rules on the provision of a specific service: a

system for personal savings and loans. Leader selection procedures were randomized

across community groups whose objective it is to make savings and loans possible for

vulnerable members of the community. Over a 3.5 year period, the performance of

the groups was monitored both in terms of continued membership and the amount of

savings accumulated by, and loans given to, members.

Our experiment contributes to the understanding of service delivery in developing

countries. In these contexts, NGOs and community groups often assume responsibility

for the provision of critical service and public goods that the state fails to provide,

and the delivery of financial services inadequately supplied by banks and other formal

institutions.1 The governance of such groups is one factor that affects their service

delivery.

We study how electoral rules affect leader selection using 2 different rules for elect-

ing leaders: (a) secret ballot plurality voting and (b) open discussion with consensus

decision. These procedures are central to the practice of direct democracy, and may

strongly affect the selection of leaders. Leaders can vary in terms of their skill level

and their representativeness. The previous literature has highlighted a trade-off be-

tween these two aspects in terms of service delivery (see e.g. Beath et al., 2014), since

measures of skills, such as education, tend to be correlated with higher socio-economic

positions while a leader closer to the median is representative of a larger fraction of

the electorate. When decisions are made in an open discussion format, less influen-

tial members may feel intimidated and refrain from contributing to decisions. This

1For more information about the role of NGOs and community groups in public service delivery:see Bernard et al. (2008), Casey et al. (2012) and Grossman (2014). For financial services: 97 millionIndian households were covered by the self-help program NABARD in 2011 (Baland et al., 2011) andaccording to data collected by VSL Associates, a consultancy, Village Savings and Loan Associations,i.e. groups with a model similar to that of the savings and loans groups studied here, reach close to12 million people worldwide, 10 million of these are in Sub-Saharan Africa (Vsla.net, 2015).

2.1. INTRODUCTION 11

tilts selection decisions towards leaders with more visible skill indicators that are less

representative for low-income group members. The representativeness of leaders can,

in turn, have an impact on public goods delivery. If leaders are closer to the average

member, public goods delivery can be expected to shift toward lower income members

(Besley and Coate 1997; Osborne and Slivinsky 1996; Phillips 1995; Pitkin 1967).

The specific context of the study is a saving program in the most impoverished

region of Uganda: Karamoja. Members are organized into savings groups, each group

jointly keeping a savings fund from which members can take loans. These groups

were founded by the NGO BRAC Uganda and are a version of self-help groups that

have become popular in Sub-Saharan Africa in the last decade (see e.g. Ksoll et al.,

2013; Greaney et al., 2016). The groups aim at empowering and improving livelihood

options for young women and, as a step towards empowerment and sustainability, they

are self-governed through a managing committee, consisting of group members. Before

the committee-formation, 46 groups were randomly assigned to elect their committees

through secret ballot, plurality rule voting (vote groups). In the remaining 46 groups,

committee members were nominated and agreed upon in an open discussion with

consensus (discussion groups). Both procedures took place at a meeting attended by

all group members, and in both cases the election was preceded by a discussion. The

difference lies in the openneness of decision making: one procedure (vote) imposed a

secret vote at the end of the discussion while the other (discussion) did not.

We find that selection rules affect both the type of leader chosen and the subse-

quent performance of the savings group. First, vote groups elected leaders who were

more representative of the average group member in terms of economic status, while

groups that selected leaders via open discussion chose less representative leaders who

were richer and more educated than the average member. For example, leaders in dis-

cussion groups were 46 percent more likely to have employment connections and 25

percent more likely to have some education as compared to regular members. They

also had larger asset holdings and scored significantly higher on a wealth index. Mean-

while, none of the differences between leaders and regular members were statistically


or economically significant in the vote treatment. These findings are consistent with

theories of representation and elite capture where open participatory selection proce-

dures give power to those who are already powerful.2 However, they are also consistent

with open discussion improving coordination and generating leaders with higher skill

levels as compared to the secret vote.3

Second, service delivery on the extensive margin was lower in the discussion groups:

while dropout was high in both types of groups, it was significantly higher in discus-

sion groups, where the poorest members were more likely to drop out. Specifically,

in the first year, dropout was 32 percent higher (15 percentage points) in discussion

groups than in vote groups, and members in the lowest quartile of the group’s wealth

distribution were 12.3 percent more likely to drop out in the first year compared to

wealthier members, while members with no market income at baseline were 18.2 per-

cent more likely to leave the discussion groups than those with market income4 and

members who kept no savings at baseline were 33.4 percent more likely to drop out

from discussion groups than to those with initial savings. These findings are in line

with committee members elected under the vote treatment having policy preferences

closer to those of the average member. Third, 3.5 years after committee formation,

service delivery on the intensive margin was lower in vote groups: members in vote

groups were 15 percent less likely to be saving at endline and also reported getting

smaller and fewer loans than their counterparts in discussion groups. The dropout has

implications for welfare: members who dropped out of groups were significantly less

likely than stayers to have access to saving or loans at the time of the endline sur-

vey. Moreover, members who were initially poor are less likely to have access to such

services (from any service provider) at endline if randomly assigned to a discussion

group.

Taken together, the results suggest that the discussion treatment can lead to more

2see e.g. Baland and Robinson (2008) and Hinnerich and Pettersson-Lidbom (2014).3That open discussion, in particular if it takes the form of deliberation, may lead to better quality

outcomes is suggested by Humphreys et al., 2006 among others.4Members with no market income were either subsistence farmers/animal rearers or dependents.


efficient savings groups through selective dropout, but at the expense of more inequal-

ity in access to savings and credit. In terms of mechanism, we provide supportive

evidence that elite capture is taking place in the discussion group while we find no

support for discussion groups being more economically efficient than vote groups. The

effects thus seem to work through leader representativeness rather than leader skill

level. Hence, the appropriate method for leader selection ultimately depends on the

objective and target group of a program.

This paper adds to several strands of literature. First, we contribute to the lit-

erature on electoral systems and their role for policy outcomes, especially to new

knowledge about the selection of leaders. As pointed out by Besley (2005), political

selection is an overlooked area in the political science and economics literature. Just

like the literature on electoral rules, the literature on leader selection is primarily con-

cerned with understanding the selection of representatives at the national or county

level. As a consequence, the electoral rules studied to date are models of represen-

tative democracy, where variation in district magnitude or the number of political

parties can affect policy. A recent example is Beath et al. (2014) who compare at-large

voting to voting by districts in an experiment in Afghanistan. Closer in spirit to our

study, Grossman (2014) studies leaders of Ugandan farmer group councils and finds

direct vote by farmer group members to be superior to appointment by council rep-

resentatives. Our study unpacks this finding by comparing two different participatory

methods involving a comparable set of agents, and also has the advantage of exogenous

variation in the leader selection rule.

Our findings also add to the literature on participatory democracy and the role of

openness in decision making. The decision-making models we study are not feasible

for national politics but are typical within the field of direct democracy.5 Consensual

5"Direct democracy" typically refers to systems in which community members directly decideon policy outcomes and thus, where there is no intermediate step of "representation". However,although this paper deals with the selection of leaders, the methods employed are the most centralones in the direct democracy literature and practice. For example, Matsusaka (2005) writes that"Direct democracy is an umbrella term that covers a variety of political processes, all of which allowordinary citizens to vote directly on laws rather than candidates for office. The town meeting, in whichcitizens assemble at a particular place and time to make public decisions, is the earliest form of direct


discourse has been found to be a way for leaders to maintain and reaffirm a social order

(Humphreys et al., 2006 and Murphy, 1990). Moreover, under open decision making,

less powerful members can be coerced or intimidated into supporting certain proposals

(Hinnerich and Pettersson-Lidbom, 2014) while the opportunity to cast a secret vote

has been shown to increase the representation of economically disadvantaged groups

(Baland and Robinson, 2008).

We recognize that one must be careful in generalizing our findings to the national

level. However, in many developing countries, service delivery is primarily provided by

agents such as NGOs and community based groups. Studying the design of electoral

systems at the macro level is not sufficient to understand efficient service delivery in

these contexts. Our results offer insights into public good delivery that are complemen-

tary to those of studies of national and district elections. This also relates our study

to the growing literature on Community Driven Development (CDD) - development

projects involving the local community in public service delivery provided by exter-

nal agents (see Mansuri and Rao (2012) for a recent review). Previous studies that

evaluate such projects indicate that direct local participation has several advantages

as compared to central decision making by an NGO. However, so far, we know very

little about the relative effectiveness of different types of direct participation. More-

over, previous studies focus on systems where decisions are made directly about policy

outcomes, which requires an infrastructure provided by external agents for their func-

tioning. Our study addresses both these shortcomings. To our knowledge, this is the

first study to compare two participatory decision-making systems in a development

program setting. Moreover, holding fixed who had access to the decision-making forum

gives us an advantage over previous studies where typically discussions with consensus

decision making only take place among a limited group of people, such as a local elite

(Beath et al., 2012) or an elected body of representatives (Grossman, 2014). While

pure direct democracy setups such as those brought forward in previous evaluations

of CDD projects (Olken, 2010; Madajewicz et al., 2014; Beath et al., 2012) can be

democracy /.../ The most prominent form of direct democracy today is an election in which citizensvote yes or no on specific laws listed on the ballot...".


suitable for case-by-case decision making, they are less appropriate for services contin-

uously provided by community organizations. In such cases, the role of leaders becomes

important. Our findings can thus offer policy advice to external actors that want to

put in place sustainable and equitable/inclusive mechanisms for service provision.

Finally, this paper also increases the knowledge in the specific policy area of how

to increase saving among the poor. Using statistics from IMF, Aggarwal et al. (2012)

estimate that less than 19% of the population of Sub Saharan Africa had a bank

account in 2011.6 A number of recent studies investigate the take up of such formal

saving technologies, for example Dupas and Robinson (2013a, 2013b). For people living

below the poverty line, informal financial institutions for saving, such as self-help

groups or Village Savings and Loan Associations are more accessible than banks and

are increasing in importance in Sub-Saharan Africa and Asia.7 A handful of recent

papers study these groups (Burlando and Canidio, 2015; Greaney et al., 2016; Ksoll

et al.,2013). These studies measure the economic performance and vary the purely

economic incentives of administrators or members, but abstract from the organization

of group members. We complement this previous work with our focus on the governance

of groups.

In the next section, we present the context in which the study took place and

the partnering NGO Program. Section 3 explains the setup of the program and the

experiment. In section 4, we lay out the conceptual framework and section 5 presents

the data. Section 6 presents the results for leader types and group policy outcomes.

Section 7 discusses the welfare effects of these findings. Section 8 concludes the paper.

6This excludes South Africa, where a significantly higher share of the population is banked thanin the rest of the region.

7See Ksoll et al. (2013) and Greaney et al. (2016).


2.2 Background: Karamoja and BRAC Uganda

2.2.1 Karamoja region

This study took place in Karamoja which is located in the North-Eastern corner of

Uganda, bordering Kenya and South Sudan.8 With its dry climate, it is the poorest

region in Uganda. According to the 2014 Uganda Poverty Status Report, 74% of its

population lived below the local poverty line (1 USD per day) compared to 19.7% in

the country as a whole.9 10 The inhabitants of the region traditionally relied on agro-

pastoralism and pastoralism for their livelihood, but these livelihood options have

become compromised in the last few decades, due to conflict and insecurity combined

with harsher climate conditions. This has resulted in Karamoja having the largest

number of food insecure people in Uganda. Other development indicators also lag

behind those of the rest of the country. Figures taken from an 2004 Uganda Bureau

of Statistics survey by Irish Aid show that the literacy rates in the region were 21%

as compared to a national average of 68%, and that 60.3% of the 6 - 25 year olds had

never been to school as compared to only 13.8% nationally.

2.2.2 BRAC Uganda and the Karamoja Initiative

The authors collaborated with the NGO BRAC Uganda. BRAC is a large non-profit

organization founded in Bangladesh in 1974, currently active in 12 developing countries

in Asia, Sub-Saharan Africa and the Caribbean. BRAC was launched in Uganda in

2006 and had by 2013, it had become one of the largest development organizations

and micro finance institutions in Uganda. Its core activity is microfinance, and other

programs include a system of health workers, agricultural extension, and self-help8For more information on the socioeconomic characteristics of Karamoja and BRAC’s activities

in the region, see Czuba; 2011, 2012a, 2012b.9The National poverty line of Uganda is 1 USD per day, but the country employs 8 additional

local poverty lines, that are adjusted to the consumption baskets in urban and rural areas in each ofthe four main regions, for more details, see Appleton, 2003.

10According to the UNDP Millennium Development Goals Report, the fraction of people living onless than 1.25 USD per day in 2015 is 14% in the World’s Developing regions and 41% in Sub-SaharanAfrica.

2.2. BACKGROUND: KARAMOJA AND BRAC UGANDA 17

groups for young women.

Our experiment takes place within BRAC Uganda’s "Karamoja Initiative" which

started in 2011 in five of the seven districts of the Karamoja region. As traditional

livelihoods have become more difficult and solely relying on agriculture is not a viable

option in Karamoja, small-scale market activities have gained increased importance

in recent years. This provides the motivation for BRAC’s activities for adolescents

in the region. Partly modeled on BRAC’s self-help groups for young women,11, the

Karamoja Initiative targets children and young women with the objective of improv-

ing education take-up among children and, in parallel, to provide ways of promoting

income generating activities (IGA) for young women. BRAC’s activity in the 114 lo-

cal Youth Development Centers (YDC’s) in Karamoja is structured around 9 branch

offices, each with a defined catchment area and employing 2-4 members of staff who

monitor the activities of their centers.12 These centers are typically located in small

houses or huts rented by BRAC in the targeted villages and communities. Each center

employs two women: One caretaker and one mentor, usually recruited from within the

community. The centers are open every weekday. In the morning hours, the caretaker

receives pre-school aged children and in the afternoons the mentor keeps the center

open for adolescent girls and young women. During its opening hours, members of the

center can engage in leisure activities such as board games combined with "life skills"

training sessions led by the mentor, on topics such as reproductive health, relationships

and water and sanitation. The center members can also join the Income Generating

Fund (IGF) savings group which is our focus in this paper. This system for savings was

introduced to the YDC members in mid-2011. Because monetization of the economy is

relatively recent in Karamoja, most YDC members had little capital at their disposal

and limited experience in managing financial flows. The objective of the IGF was to

encourage savings and facilitate the starting of business activities. Group members’

11BRAC’s self-help groups in Uganda are called Empowerment and Livelihood of Adolescents (ELA)clubs. They are present in 80 districts of Uganda (excluding Karamoja) and have been studied byBandiera et al. (2014).

12Three in the Napak district, two in the Nakapiripirit district, two in the Moroto district and onein the Kotido and the Amudat districts, respectively.


savings are collected at weekly meetings and placed in a box, initially safeguarded by

BRAC staff but eventually to be handed over, along with the responsibility for the

other IGF activities, to the saving members. After an initial 12 months of saving, the

groups started providing loans to their saving members. Occasionally, more structured

courses in income generating activities are offered to members of the center. These are

courses in agriculture, poultry rearing, hairdressing and baking with the aim of pro-

viding skills for starting small scale business activity, but by 2015, the saving meetings

are the only structured activity for adolescents hosted regularly at the Youth centers.

2.3 Setup of saving groups and leader selection meth-

ods

In mid-2011, BRAC started the YDC in Karamoja. After a few months, the IGF

groups were started as one component of the activities of the centers. Young women

in the age range of 13-21 were invited to join the groups and start saving. In January

2012, local BRAC staff members instructed the groups to form committees. This was

an important step in handing over ownership and governance of the groups to the

members themselves.

Before committees were chosen, the committee selection method was randomly

assigned to the groups. The randomization was carried out by the BRAC Research

Unit under the supervision of one of the authors. Randomization was stratified at the

branch level to ensure variation in the appointment method between the groups within

each branch, and the smallest and largest groups were excluded before randomization,

leaving 92 groups in our study.13 In all groups, the mentor and a staff member from

the corresponding branch office informed the members of the IGF savings group that

the group was to select a committee. Information was given about the role of the

committee and about each committee position. The members were told that they13Groups with less than 10 or more than 30 members were excluded to leave a set of more com-

parable groups in the experiment. The groups that were not included in the experiment selectedcommittees according to the discussion setup.

2.3. SETUP OF SAVING GROUPS AND LEADER SELECTION METHODS 19

would meet again approximately one week later to select the committee and were also

told that desirable characteristics of a committee member were that she should be

trustworthy with money and accepted in the community. At the second meeting, the

members were reminded by the local BRAC staff about which committee positions

were to be chosen, and the specific tasks associated with each position. The positions

were: chairperson, treasurer, secretary, keyholder 1 and keyholder 2. The role of the

two latter members is to store keys to the saving box, which was to be kept at the

house of the treasurer. The mentor was usually suggested to become the secretary or

the treasurer. The selection of committees then happened in two distinct ways.

Open Discussion: The group members openly discussed each position and anyone

who was a saving member could nominate candidates for the position. Other members

could then second or oppose the nomination openly until the group agreed on a name.

Then, they proceeded to discuss and fill the next position.

Secret Ballot (Vote): The group members openly nominated candidates for each

position. For each position, at least two candidates were required. Members would

then vote by writing the name of the person they preferred for each candidate and

drop it into one of 4 boxes or baskets, one for each voteable position.14. BRAC staff

assisted with writing (for those unable to write themselves) and with compiling the

votes.

In both appointment systems used, all members were invited to attend the meeting,

and potential committee members were discussed. In the discussion treatment, this

open discussion was the means to make the final selection for each committee position.

In the vote treatment, the discussion only served to nominate at least two people for

every position, after which the final decision was made by secret ballot. The role of the

committee involves tasks that can be divided into four main areas of responsibility: The

first is to encourage members to attend weekly saving meetings and to save regularly.

The second is concerned with the safeguarding of the money in a saving box, and

the third is allocating loans to members and ensuring that these are being repaid.14The mentor was automatically given a position either as treasurer or as secretary so only four

positions remained to be voted for.


Finally, the committee is responsible for keeping books at the group level in order to

keep track of the savings, loans and interest rate amounts. Newly elected committee

members received a few days of training by BRAC, organized at the branch level.15

The committees officially had a term limit of one year. After this, every group was free

to decide how to appoint their next committee. In the summer of 2012, approximately

6 months after the committee appointment, most groups started giving out loans

from their pooled savings. The decision about when to start lending was taken at

the branch level. As the saving groups became free to adjust their bylaws, most of

the groups started following a model used by Village Savings and Loans Associations

(VSLAs) that became increasingly popular in the Karamoja region. This is a model

with saving cycles, typically one year in length, at the end of which a "share out" takes

place. In this meeting, all funds in the saving box including the interest rate generated

by lending, is shared among the saving members according to their level of savings.

After this, a new saving group is formed, and a new committee is appointed.16

Figure 2.1 shows the timeline of the program implementation and data collection

activities from 2011 until 2015.

2.4 Conceptual framework

In this section, we discuss the characteristics potentially relevant for leaders that have

been highlighted in the previous literature, and relate them directly to the role and

tasks of leaders in our context. Then, we outline the hypotheses about how the electoral

15The training covered basic concepts of financial literacy. Committees from all centers within onebranch attended the training together. This ensures that committee members in both treatmentsattended identical training sessions, independent of method used when electing them. Spillover ef-fects between treatments from this joint training are not likely since the mechanisms through whichtreatment would affect the groups played out either before the training took place, in the meetingitself, through the appointment methods used producing different types of committee members, orthrough the feeling of legitimacy of regular members who did not attend this training.

16In practice, local BRAC staff from the branch office associated with a given youth center, whovisit each group at least one time per month, still function as a type of mentors for group membersand can be influential in their governance. Since we are not able to observe exactly how this affectsthe group performance, we include branch fixed effects in the analysis to account for this unobservablecharacteristics that are common for the groups within one branch and fixed over time.

2.4. CONCEPTUAL FRAMEWORK 21

rule can be expected to affect these characteristics and how this, in turn, may affect

the policy outcomes we study: member retention, savings and loans.

2.4.1 Leader representativeness and skill level

The literature on political selection highlights a tradeoff between leader representa-

tiveness and leader quality in terms of skills.

Representativeness of leaders would imply leaders with similar preferences to

those of the median group member, in terms of rules for savings and loans and the

ensuing stringency. It can be measured in terms of similarity in observed socio eco-

nomic characteristics. Grossman (2014) shows that personal ties substitute for rule

enforcement in Ugandan farmer groups. In a similar way, friendship or kinship ties

with leaders can proxy for trust and convey information about commonly held prefer-

ences. Being a member of the same tribe can work in a similar way (Alesina and La

Ferrara, 2002).

Representativeness can affect outright favoritism, for example by leaders offering

loans to members of their own group, or being more lenient with the repayment from

such members. It may also lead to policy more targeted to the preferences of the own

group: by setting the level of minimal savings high, a rich committee member can push

out poorer members or, conversely, by choosing a low interest rate and offering few

loans, which makes the aggregate money grow at a slower pace, a committee member

can push economically successful members out of the group, as they may have more

profitable outside options.

Skills are features that affect every group member in the same direction, what

Besley et al. (2005) refer to as a valence issue. A leader with higher skills makes

everyone in the group better off. Examples would be better accounting skills, general

honesty and reliability with money, and the ability to make members who took loans

repay them to the group. Since the objective of the groups we study is to encourage

savings and give loans for profitable market activities, skill indicators in our setting are

variables such as education, economic performance at the baseline, market experience


and labor market connections.

2.4.2 Treatments’ relative effects on representativeness and skills

of leaders

Recent work predicts that elite capture and intimidation may lead discussion treat-

ment groups to select less representative leaders (Hinnerich and Pettersson-Lidbom,

2014). Introducing a secret poll has been shown to generate leaders that are more rep-

resentative of the preferences of the electorate (Baland and Robinson, 2008). Based on

these findings, we would expect the leaders emerging in the vote treatment to be closer

in characteristics and preferences to the average group member than leaders selected

in discussion. In addition, the Discussion framework may provide better conditions

for coordination which could lead to more informed decisions regarding leader skills

(Humphreys et al., 2006).

If the coordination channel is at work, it would lead to positively selected committee

members, in terms of economic characteristics, in particular education and wealth. If

we observe that committees in discussion groups are richer and more educated than

the non-committee members of their group, this can thus be a result of either the

discussion treatment facilitating coordination and leading to more skilled leaders, or

of elite capture taking place in the discussion treatment. To disentangle the two,

we need to follow the group performance over time and construct measures of group

economic effectiveness.17

17One additional mechanism that can differ between open discussion and secret vote is legitimacy.The literature shows that people like to participate in decision making and a more inclusive appoint-ment procedure may lead to greater feeling of legitimacy among a higher share of members. Legitimacyeffects have been observed in both Olken (2010) and Beath et al. (2012). As vote treatment is themore inclusive process in our setting, it is also expected to entail greater feeling of legitimacy here.Dal Bó et al. (2010) use a lab experiment to disentangle satisfaction with the decision-making pro-cess from satisfaction with the outcome, and show that being "heard" increased subjects’ feeling oflegitimacy, regardless of whether the rule they had voted for was chosen.

2.5. DATA AND DESCRIPTIVES 23

2.4.3 Hypotheses

Group performance in public service delivery is affected by leader characteristics. A

member’s utility from the group is affected by the skill level of the leader and by the

closeness of the leader’s group affiliation to the member’s own group. A higher skill of

a committee member has a positive effect on each member, regardless of her own group

affiliation. The skill level is thus expected to increase the overall level of savings and

loans in a group. The distance between a member’s group affiliation and the leader’s

group affiliation implies a cost for the member. The leader’s group affiliation thus

affect the types of members asymmetrically and the larger is the difference, the higher

is the cost imposed on the member. If the distance between a member’s own type and

the leader type is too large, the member will be better off by leaving the group.

2.5 Data and Descriptives

2.5.1 Baseline census and survey data

A baseline survey was administered to the saving members before the treatment, in

September-November 2011. The baseline survey was preceded by a census of the groups

in August 2011, listing the name and age of all members. 85% of the listed members

were interviewed at the baseline. Reassuringly, there is no difference between treat-

ments in the share of members interviewed. Baseline summary statistics and random-

ization checks of the baseline variables are presented in Table 2.1.18

Table 2.1 shows the means in the population and by treatment of a number of

variables that measure group characteristics (such as group size), the economic per-

formance of members at baseline, or the socio-economic characteristic of the members

and their household. The last two columns of table 2.1 report treatment differences

and p-values from weighted least squares regressions of the following form:

18Additional explanations for construction of baseline variables are provided in Appendix 2.


Ygb = α +βVoteg +θb + εgb, (2.1)

where Ygb is the mean of the variable in group g, branch b, Voteg is an indicator

variable equal to 1 if the group received the vote treatment, θb is a branch indicator,

controlling for unobserved characteristics at the branch level such as the local BRAC

staff serving the centers within a branch, or local market or agriculture conditions.

The regressions are weighted by the number of individuals interviewed in the group at

baseline. The sample is balanced across treatments: Table 2.1 shows that groups in the

vote treatment and the discussion treatment are, on average, comparable on baseline

characteristics. The only variable that is significantly different at the 1% level is the

share of members that had received a training in income generating activities (IGA)

offered by BRAC. These are practical trainings focusing on skills such as baking or

animal rearing, occasionally offered at the branch level to a few members from each

group within the branch. To ensure that this characteristic does not drive observed

differences between treatments, we control for the baseline share of members with IGA

training in our regressions that focus on baseline variables.

2.5.2 Committee member data

The next set of data that we use is the complete list of the original committee mem-

bers, i.e. the committee members that were selected during the treatment which took

place in January 2012. Merging this list with the baseline data enables us to analyze

differences in predetermined characteristics between the committee members elected

in vote (secret ballot) treatment compared to those selected in discussion treatment.

Out of the 462 people listed as original committee members of the groups, we have

baseline data for 323. Out of the remaining 139 original committee members, 37 were

listed as members in the census made before the baseline, but were not interviewed at

the baseline due to non-response, and the other 102 were not listed as members in the

census made before the baseline; these are hence members who joined the groups in

2.5. DATA AND DESCRIPTIVES 25

the time period between the baseline (September-December 2011) and the committee

selection (January 2012). Importantly, there is no difference between the treatments in

the number or the share of committee members interviewed in the baseline survey.19

2.5.3 Follow-up in 2013

One year after the treatment, we collected individual level data for all baseline members

on their current membership status and group level data on savings and loans in the

group during the first year of self governance. We use this data to examine dropout

patterns from the groups between 2011 and 2013.

2.5.4 Census data collected in 2015

A census in March 2015 collected information on all current members in the groups, and

followed up on all baseline members. This census provides individual-level information

on the membership status of each of the baseline members who we classified into

stayers or dropouts. In addition, we have information on all new members, i.e. those

who joined the group after the treatment. We use the census data to examine dropout

patterns from the groups between 2011 and 2015, and it also provides the sampling

frame for the endline survey.

2.5.5 Endline survey

An endline survey was collected in May-July 2015. This survey was conducted within

a random sample of each of the three sub-groups defined in the 2015 census: Stayers,

Dropouts and New Members. This survey collected socio-economic data similar to the

baseline data, and also data on network variables for the original leaders and person-

ality features and satisfaction measures of group members. All members in the census

19The committee size deviated from 5 in four cases: One group had 4 committee members, 2groups had 6 committee members and one group had 7 committee members. These differences arenot significantly correlated with group treatment status.


could not be interviewed for cost and time reasons. Since our main focus is the treat-

ment effect on members who were present when the leaders were selected, we sampled

all stayers and only a random sample of new members (42 %). We also sampled 41%

of the dropouts. The over-sampling of stayers in relation to dropouts was motivated

by cost reasons as stayers were easier to locate and interview than dropouts. Table A.7

shows sampling and attrition rates. All regressions using endline data include sample

probability weights to account for oversampling of stayers in relation to dropouts and

new members. In the endline survey, we also collected group-level policy variables in

an additional module. This module collected information from one committee member

of each group about the interest rate imposed on loans, whether the group employs a

VSLA model with annual share-outs, questions about the saving cycle, and how the

committee deals with loan defaults and with members who are not saving on a regular

basis. Wealth score, assets and income variables are constructed in a way similar to the

one described for the baseline survey.20 Social network variables were also collected in

the endline survey. Here, the focus was on the original members and the committee

members. To all respondents, old and new, we posed a series of questions about the

five original committee members. This information serves to construct basic network

measures and to investigate how members evaluated the performance of original com-

mittee members. To all old respondents (stayers and dropouts) we also asked, for each

of the original group members, if they knew who the person is and if they were a close

friend or relative known before joining the savings group.

2.6 Results

In this section, we present the results of the electoral system on committee member

selection, member retention, and savings and loan services at the intensive margin. The

estimating equations are presented along with the corresponding results. Throughout

20For assets, we only use the respondent’s self reported value for the assets houses and land holdings.For the other assets included in the roster, electronics and various livestock and poultry, we imputetypical asset values for the region collected by BRAC Research and Evaluation Unit.

2.6. RESULTS 27

the results section, the base category will be the discussion treatment, while we regress

outcomes on an indicator for the group being exposed to the vote treatment. All

individual level regressions standard errors are clustered at the group level (the level

of randomization).

2.6.1 Leader selection

The first question we address is the effect of the appointment method on the type of

leaders selected. First, we restrict our sample to committee members, and estimate

the following linear model:

Yigb = α +βVoteg +θb + εigb, (2.2)

where Yigb: an economic or social characteristic of committee member i in group

g, branch b, θb is a branch indicator, controlling for unobserved characteristics at the

branch level such as the local BRAC staff serving the centers within a branch, or local

market or agriculture conditions. β is the coefficient of interest, measuring the effect

of the vote treatment on the likelihood that the committee members appointed have

the predetermined characteristic Y .

Columns 1-4 of Table 2.2 show the result of estimating equation 4.1 using proxies for

the economic performance as the dependent variable. On average, leaders (committee

members) in the vote treatment are poorer than leaders in the discussion treatment,

as measured by a lower value on the Wealth index21. The effect of vote treatment on

the likelihood that leaders have received employment advice from anyone outside the

household is negative, but just below conventional confidence levels. For robustness, we

also estimate equation 4.1 using other measures of wealth such as income, assets or the

raw wealth score, rather than a member’s position in her group’s wealth score distribu-

21The Wealth index is composed using the Uganda Progress out of Poverty index, compiled by theGrameen foundation (2011). This index combines information on poverty indicators such as materialof roof, walls and floor of a household’s main house, its ownership of shoes and clothes, access towater, power sources and sanitation, and education level in the household.


tion, as a measure of relative wealth (see Table A.1 in the Appendix).22 Columns 5-7

of Table 2.2 show the result of estimating equation 4.1 using socio economic character-

istics as dependent variables. Leaders in vote treatment score lower on socioeconomic

proxies: they are 27 percent more likely to have children and were, on average. 0.9 years

younger at the time of birth of their first child.23 They are also 34 percent less likely to

have migration experience than leaders in the discussion treatment. Migration among

young women in Karamoja is normally short-term migration for cultivation or casual

work in farming on more fertile land in neighboring districts, or sometimes migration

to urban areas for studies or for unskilled work. Migration experience is positively

correlated with wealth in our sample. The last three columns of Table 2.2 show the

results for social characteristics, such as tribe and the number of connections to other

group members. Leaders in vote groups are more likely to belong to the majority tribe

of their group and also somewhat more likely to have reported fellow leaders to be

among their four best friends at baseline - these differences are not statistically sig-

nificant at conventional confidence levels, however. Overall, the variables measuring

wealth or economic situation of the leaders are consistently higher in the discussion

treatment, albeit the difference between treatments is not always statistically signif-

icant. This suggests a positive selection of leaders on economic characteristics in the

discussion treatment.

The next question that we analyze is whether leaders in vote treatment groups

select leaders that are more similar to the average group member measured in socio-

economic variables, or more likely to be their friends, compared to the leaders of

discussion treatment groups. The upper panel of figure 2.2 plots the position in the

group’s wealth score distribution for leaders (committee members) and non leaders

separately, by treatment. The lower panel of the figure does the same for the initial

assets held at the time of baseline. In discussion groups, the distribution for leaders

22For log income and for the raw wealth score, the results confirm the finding that leaders invote groups are poorer. For asset score, the point estimate is negative but no longer statisticallysignificant when controlling for the baseline share of group members that had received training inincome generating activities (see Table A.1 in the Appendix).

23This variable is only available for members with children, which constitute 67% of the sample.

2.6. RESULTS 29

has more mass at higher levels of the wealth score and asset distribution in discussion

than that for non leaders. This indicates that leaders in discussion treatment groups

are positively selected on wealth and assets, while the distribution among leaders in

vote groups more closely follows that of non leaders.24 To examine differences between

leaders and non leaders in a more systematic way, we estimate the following OLS

regression:

Yigb = α +βVoteg + γLeaderi +σ [Voteg×Leaderi]+ηg + εigb, (2.3)

Yigb is a characteristic that proxies either for economic performance or the social

connectedness of a group member, CM is an indicator for whether an individual was

selected to become a leader (committee member), Voteg indicates the vote treatment

and ηg is a group indicator. γ [γ+σ ] estimates the correlation between the variable

Y and being a leader in the discussion [vote] treatment. The coefficient of interest is

σ , indicating the difference in the correlation between characteristic Y and becoming

a leader between the two treatments. The sum of γ and σ , shows how leaders differ

from non leaders if the treatment is vote.

Table 2.3 shows the result from estimating equation 4.2 with baseline characteristics

proxying for economic performance (columns 1-4), socio economic variables (columns

5-7) or social proxies (columns 8-10) of members as the dependent variable. For three

out of the four economic outcome variables, it applies that for the discussion treatment

groups, the mean difference between leaders and other group members is statistically

significant at least at the 95 percent confidence level. For example, leaders in discussion

groups are 46% (12.9 ppt) more likely to have employment connections and 25% (12.2

ppt) more likely to have some education than regular members. The differences for a

wealth index and for asset holdings point in the same direction. For the vote treatment,

none of the differences in economic variables between leaders and regular members are

statistically or economically significant. The interaction terms with the vote treatment24Wealth score indicates the position (decile) of an individual in the wealth score distribution of her

group. The score is constructed using the Uganda PPI index compiled by the Grameen foundation(2011). Higher value=less poor. The wealth score does not include the value of household assets.


attenuates the difference between leaders and non leaders, meaning that in vote groups,

leaders resemble non leaders more than in discussion groups. Looking at the sum of σ

and γ , we cannot reject the hypothesis that the difference between leaders and regular

members in vote groups equals zero for any of the four variables.25

Thus, in general, on economic performance variables, the vote treatment leaders are

more similar to the average of regular members in their group than in the discussion

treatment where leaders are positively selected in terms of wealth, education and

market connections.

Columns 5-7 of Table 2.3 show the result from estimating of 4.2, using baseline

socio-economic proxies as dependent variables, and columns 8-10 of Table 2.3 show the

result from estimating of equation 4.2 with social connectedness-measures of members

as the dependent variables. For the socio-economic variables children, age at first birth,

and migration experience, the difference between leaders and regular members in vote

groups is also significant at least at the 95 percent confidence level and indicates lower

socio economic scores among leaders in vote groups. The hypothesis that the value of

this characteristic is the same for leaders as for regular members within a given group

if the treatment is vote can be rejected at the 99 percent confidence level for all social

connectedness variables. In vote groups, leaders are more likely than other members

to have strong links both to other group members and to other leaders (committee

members), as measured in the baseline survey. Leaders in vote are more likely to come

from the majority tribe of the group than non leaders. Overall, the results for socio-

economic characteristics indicate a negative selection for leaders in vote treatment.

Taken together, the results in this section suggest that in the vote treatment,

members do, to a higher extent, elect people who are similar to themselves, and for

individuals more socially connected within the group: they are more likely to belong to

the majority tribe and more likely to be well connected to other members who ended

up becoming leaders. In the discussion treatment, leaders are positively selected in

25The results are robust to adding the individual baseline wealth score or assets as a control,to ensure that the results are not driven by any correlation with leader wealth (Table A.2 in theAppendix shows results controlling for the wealth score).

2.6. RESULTS 31

terms of wealth, education and market connections.

2.6.2 Results on member retention and dropout

2.6.2.1 Dropout rates in each treatment

From January 2012, the time of the committee appointment, until March 2013, 45%

of the original members left the vote treatment groups and 60% left the discussion

treatment groups. The dropout continued between 2013 and the endline but at a

slower rate, and the difference between treatments became less stark over time. In

March 2015, 33.5% of the original members remained in the vote groups compared to

27.3% in the discussion groups

Figure 2.3 shows the share of original members that was still saving in the group

in March 2013 and in March 2015, respectively. Regressing dropout on treatment

shows that this large difference between treatments after one year is highly statistically

significant. Table 2.4 shows the result of this regression, which is an estimation of the

model:

Dropouttigb = α +βVoteg +θb + εigb, (2.4)

t ∈ {2013,2015}, Dropout is a binary outcome variable for individual i, group g

branch b, equal to 1 if the individual had dropped out of the group at time t, which is

either in 2013 or 2015, and 0 otherwise. θb is a branch indicator.

As can be seen from both Figure 2.3 and Table 2.4, the difference between treat-

ments in dropout by 2015 is weaker, when clustering standard errors at the level of

randomization and controlling for branch indicators, the p-value of the difference be-

tween the treatment is 0.11. The sizable dropout from both types of treatment groups

means that composition effects need to be taken into account in order to understand

differences between the groups at endline. In the next subsection, we examine who

drops out in more detail.


2.6.2.2 Results for selective dropout

To understand how dropout affects the efficiency of the group and to understand the

welfare implications of the dropout, we need to know what characteristics determine

dropout and if the dropout pattern differs across treatments in terms of what type of

members leave the group. To examine this, we estimate the following OLS regression:

Dropouttigb = α +βXigb +σ [Voteg×Xi]+ δVoteg +θb + εigb, (2.5)

where t ∈ {2013,2015} and X is an economic characteristic, proxying for poverty,

of member i measured at baseline, such as a dummy for being in the bottom 25%

of the wealth or asset distribution of her group or a dummy for having no market

income. θb is a branch indicator. β and σ are the coefficient of interest. β tells us the

predictive power of characteristic X on dropout if Vote=0, that is, if the treatment is

discussion. σ is the treatment difference in the effect of characteristic X on dropout

for a member in the vote treatment. The sum of the coefficients β and σ tell us to

what extent characteristic X affects dropout if the treatment is vote.

Table 2.5 shows the result from the regression of dropout by 2013 on five eco-

nomic characteristics (poverty proxies) measured at baseline. For all five variables, the

coefficient measuring the predictive value of the variable on dropout is positive and

significantly so for all variables except assets. This tells us that dropout in discussion

treatment is negatively correlated with wealth and with having savings, loan experi-

ence and market income at baseline. Specifically, members in the lowest quartile of

the group’s wealth distribution were 12.3 percent (7.1 percentage points) more likely

to drop out in the first year compared to wealthier members, while members with no

market income at baseline were 18.2 percent (10.3 ppt) more likely to leave the discus-

sion groups than those with market income. Members who kept no savings at baseline

were 33.4 percent (19.4 percentage points) more likely to drop out from discussion

groups than those with initial savings, and members with no loan experience were

53.5 percent (24.5 ppt) more likely to leave discussion groups than those with loans at

2.6. RESULTS 33

baseline. The interaction term between each of the five economic variables and the vote

treatment goes in the opposite direction, implying a much smaller effect of economic

background variables on dropout patterns in vote groups. Tests of whether dropout

rates are the same for poor people across the two treatments are rejected at the 95%

confidence level for all five economic variables, while tests of whether dropouts and

stayers in the vote treatment are similar can not be rejected for any of the variables

(p-values reported below table).

Table A.3 in the Appendix shows the same regressions but now the dependent

variable is dropout by the year 2015. The results point in the same direction but are

not as conclusive and significance levels are lower. This is likely to be explained by the

high initial dropout rates implying that the groups are, by 2015, less affected by the

original treatment.

To sum up, members who had lower economic indicators at baseline leave the dis-

cussion groups at a significantly higher rate than richer members while vote groups

retain a member group which is more diverse in terms of their initial economic char-

acteristics and thus more inclusive of initially poor members.

2.6.2.3 Results for group size in 2015

To understand if the dropout leads to smaller groups over time, or if it is compensated

by new members joining the groups, Figure 2.4 shows the distribution of group sizes

in 2015, separately for vote treatment groups and for discussion treatment groups.26

Data on group size comes from the 2015 census and is the number of old stayers plus

the new members. As shown in the figure, discussion groups are smaller than vote

groups. Regressing group size on the treatment also yields a positive point estimate,

but the relation is not statistically significant. However, the number of observations

are fewer: only 84 out of the 92 initial groups are active in 2015, and some of those

were very small at the time of the 2015 census. Table A.8 in the Appendix presents the

26For 2013, we only have dropout data but do not have access to data for new members and cantherefore not perform the same exercise.


regression of group size in 2015 on the treatment. For the full sample of groups, the

difference between treatments is not significant. When restricting the sample to the

84 groups that are active at the time of the endline survey (column 3), we approach

conventional significance levels and when restricting the sample to groups with above

5 members at endline (column 4), the p-value is 0.11. To sum up, vote groups are

slightly larger at endline but the overall group level activity is low in terms of number

of active members.

2.6.3 Group performance in savings and loans delivery

Next, we turn to the question of how the appointment method affects the perfor-

mance of the groups in terms of savings and loans activity measured at the endline.

To examine the effect of treatment on savings in 2015, the following OLS model is

estimated:

Y 2015igb = α +βVoteg +θb + εigb, (2.6)

where Y is a dummy for having savings in the group, a dummy for having savings

anywhere, or log savings (intensive margin) and θb is a branch indicator. Regressions

include sample weights to account for the over-sampling of stayers as compared to

dropouts in the endline survey.

Table 2.6 shows the result from estimating equation 2.6 on all old members, as well

as on old members broken down by whether they are stayers or dropouts in 2015. The

dependent variable in columns 1, 3 and 7 of Table 2.6 is a dummy for having savings

in 2015 (regardless of where they are kept), and in column 5 the dependent variable

is a dummy taking the value of one if the individual is saving in the BRAC group

(available for stayers only).27

For the union of stayers and dropouts, the point estimate for the effect of vote

treatment on savings in 2015 is negative across specifications, but not statistically27Here, it is important to note that while stayers can have savings both in the BRAC group and

elsewhere, dropouts, by definition, have no saving in the BRAC group.

2.6. RESULTS 35

significant. Looking separately at stayers and dropouts; stayers in vote groups are 11

% (9 ppt) less likely than stayers in discussion groups to be saving at endline. This

result is statistically significant at the 90 percent confidence level. Focusing on saving

in the BRAC group as the dependent variable, the results point in the same direction:

stayers in vote groups are 15 % less likely than stayers in discussion groups to be saving

in the group at endline. We see no difference in the saving behavior of dropouts across

the two treatments. Columns 2, 4 and 8 of Table 2.6 show the result from estimating

equation 2.6 with log savings in 2015 as the dependent variable. Although the point

estimates are not statistically significant, the results point in the same direction: Vote

groups members save less also at the intensive margin.

To examine the effect of treatment on loans in 2015, we estimate the following OLS

regression:

Yigb = α +βVoteg + δXigb +θb + εigb, (2.7)

where Y is a dummy for having ever taken a loan (extensive margin), the number

of loans taken, or log loans (intensive margin), and X is a control for assets at baseline.

Assets at baseline are chosen as a control because it is an important predictor of who

is given a loan. As before, θb is a branch indicator. Regressions include sample weights

to account for the over-sampling of stayers as compared to dropouts in the endline

survey.

To measure the loan activity within BRAC savings groups, we use 3 different

variables: (i) extensive margin of loans, (ii) intensive margin of loans for those with

nonzero loan values and (iii) number of loans ever given by the group (values of zero

are also included). In order to understand the welfare impact of any effects, we also

use loans taken anywhere (where we have information for the extensive margin only)

as outcome variables .

Table 2.7 shows the result for loans within the group, for all old members as well

as on old members broken down by whether they are stayers or dropouts in 2015.

There are no statistically significant effects on the likelihood of having taken a loan


(extensive margin), although when reducing the sample to stayers, the coefficients are

negative for the vote treatment and closer to conventional confidence levels than in

the whole sample or in the sample of dropouts. On the intensive margin, as shown

in columns 2, 5 and 8, of Table 2.7, stayers who took loans in vote treatment groups

took smaller loan amounts: the coefficient of the vote treatment is negative across

ways of measuring loan size, but the results are not significant at conventional levels.

Any difference between treatments in the number of loans and loan size appears to

be driven by stayers; the coefficient for dropouts is less negative than for stayers and

for the intensive margin, dropouts from vote groups appear to borrow larger amounts

than dropouts in discussion groups. Finally, columns 3, 6 and 9 of Table 2.7 show the

results for the number of loans taken in the group. Results are similar.28

Finally, we estimate the effect of treatment on new members, i.e. members that

joined the groups after the committee selection in January 2012, and are only in-

terviewed at the endline. Since we do not have any baseline data for this group, we

estimate the following equation:

Yigb = α +βVoteg + δXg +θb + εigb, (2.8)

where Y is a dummy for ever having taken a loan (extensive margin), the number of

loans taken, or log loans (intensive margin), and X is a group level control. As before, θb

is a branch indicator. Columns 1-4 of Table 2.8 show the result of estimating equation

2.8 using savings at the endline as the outcome variable. At the extensive margin, we

see that also among new members, the likelihood of being actively saving is smaller in

vote groups. This difference appears to be driven by savings kept outside of BRAC.

On the intensive margin, point estimates for the vote treatment are positive but not

statistically significant. Columns 5-8 of Table 2.8 show results from the estimation of

equation 2.8 using loans at endline as the outcome variable. Among new members,

the number of loans accessed through BRAC is significantly lower for members of the28Table A.4 in the Appendix shows results for ever having taken loans anywhere. As for the like-

lihood of having taken a loans within the group, the coefficient is negative for vote groups but theestimate is not significant at conventional levels.

2.6. RESULTS 37

vote treatment group. There appears to be no difference between treatments in having

received a loan, nor in the loan size conditional on having borrowed - the coefficients

are negative for the vote treatment but far from significant.

To sum up, a lower fraction of members are actively saving in vote treatment

groups, than in discussion groups in 2015. The number of loans taken in vote groups

is also lower, both for old and new members. The amount of loans taken is smaller

in vote groups. The results are statistically weak, and the result for loans are only

significant in the smaller subsamples of respondents for whom we have baseline data

on assets. However, the point estimates are negative for vote groups across the different

models and robust to alternative ways of measuring savings and loans.

2.6.4 Potential mechanisms

The results found for leader types and subsequent dropout patterns indicate that elite

capture may be taking place in the discussion groups, with powerful members electing

leaders of their own type, favoring members that are similar to themselves. The results

may however also be explained by coordination taking place in the open discussion,

resulting in more skilled leaders. In this subsection we discuss and present suggestive

evidence that can help us distinguish between these two mechanisms.

Elite capture: To obtain a measure of whether leaders were favoring a specific type of

members, we examine loan allocation data. The committee decides which loan appli-

cations to approve and thereby have the opportunity to favor certain members. Since

dropout was already occurring during the time when groups started giving out loans,

it is difficult to isolate the causal impact of loan allocation on dropout. For this reason,

we focus on stayers since for these members the problem of selection out of groups is

less severe. Table 2.9 shows the loan allocation for members by their baseline poverty

status. Members who were poor at baseline or who had no loan access at baseline are

significantly less likely to be granted a loan (columns 1 and 3) and get fewer loans

(columns 4 and 6) in discussion groups than non poor members. The interaction be-


tween poverty indicators and the vote treatment cancels out this negative effect of

poverty on loan access within the groups. The results for aggregated loan amount

point in the same direction but are less statistically significant.29

Efficiency: To measure whether discussion groups are more efficient than vote groups,

and if wealth is correlated with efficient loan behavior of members, we examine de-

fault data. Table 2.10 shows the propensity to default on a loan as a function of the

electoral rule and poverty, proxied as before by the member’s position inverse wealth

distribution. There are no significant differences between vote and discussion group

in the likelihood to default on a loan, and being poor does not appear to affect the

likelihood of default, conditional on having been offered a loan.30

Our findings suggest that, in our setting, the open discussion rule entails elite

capture rather than efficiency. Leaders in discussion groups favor richer members when

approving loans, while such members are not better borrowers, as measured by default

rates.31

2.7 Welfare effects

We have showed that the electoral rule for appointing leaders leads to different pol-

icy outcomes. Groups that elected their leaders through open discussion have higher29Table A.5 shows the loan allocation for the sample of all old members (stayers and leavers).

Results are stronger than for the subsample of stayers only due to the selection out of groups, withpoor people in discussion groups more likely to be among the leavers and thus without access to thegroup’s loans.

30The measure for default is taken from the endline survey and is self-reported. The findings areconsistent with default information from group-level administrative data from BRAC available uponrequest.

31A third potential mechanism behind why more members, and poorer members, stay in the votegroups than in the discussion groups is legitimacy. If members of vote groups perceived the leaderelection mechanism as more legitimate this may explain their staying in the group. Although difficultto disentangle from elite capture, we attempted to measure the perception of fairness and voice byquestions asked in the endline survey. Among original members (stayers and leavers), we find nodifference between the treatments in the perceived fairness of the way the initial meeting was beingconducted. Nor do we find that members perceive themselves to have more "voice" in any of thetreatments. Results are available upon request.

2.7. WELFARE EFFECTS 39

dropout, in particular by economically weak members, as compared to groups who

elected leaders by secret vote. In this section, we discuss the interpretation of our find-

ings in terms of welfare effects on the entire target population of the savings and loans

groups, both for those who left the groups and those who stayed.

The objective of the groups studied here is to make saving and borrowing possi-

ble for the most vulnerable members of society; often those not allowed to join other

savings groups. In order to understand the long-run welfare implications of the differ-

ences in governance induced by our treatment, we do not only care about the mem-

bers remaining in the groups but also how the dropouts fare. The results presented in

the previous section suggest that for these members who dropped out of the groups,

there are no significant differences in savings and loans across treatments. To examine

whether stayers differ from dropouts in their access to savings and loans, we estimate

the following equations:

Yigb = α +βDropouti +θb + εigb (2.9)

Yigb = α +βDropouti +σ [Voteg×Dropouti]+θb + εigb, (2.10)

Yigb is a dummy for having savings (anywhere), log savings kept (anywhere) or a

dummy for having ever taken a loan from any lender. The coefficient of interest is β

which indicates if dropouts have a different value of the outcome variable than staying

members. The sample is restricted to old members.

The first two columns of Table 2.11 show the results from the estimation of equa-

tions 2.9 and 2.10 with loan dummy as the dependent variable, while the next two

columns use the saving dummy as the dependent variable. Dropouts are less likely to

save or to have taken a loan than stayers. This finding is highly significant and the

difference is large: Dropouts are about 56% (or 45 percentage points) less likely to be

saving at endline than staying members and are 31% (or 20 percentage points) less

likely to have taken a loan. The last two columns of Table 2.11 are estimated only for

those who save at endline and use log savings as the dependent variable. The dropouts


who are saving save larger amounts than the stayers who are saving, but this is clearly

a highly selected group given the low savings rate among dropouts. Thus, we can rule

out that members drop out to join other savings groups. Instead, dropouts display

substantially lower levels of savings and loans at endline. Columns 2, 4 and 6 in Table

2.11 show results from the estimation of 2.10, for each of the three dependent variables,

and tell us whether the correlation between dropout and access to savings and loans

differs between the two types of treatments. The interaction term suggests that this

is not the case.32

Finally, we examine if initially poor members are less likely to have access to saving

or loans in general, not only from the BRAC group. Table A.6 in the Appendix shows

that members who were initially poor are less likely to have access to such services at

endline if they were randomly assigned to a discussion treatment group. This holds

true both if poverty is proxied by market income, position in the wealth distribution,

or baseline access to loans.

Since a larger fraction of members left the groups in the discussion treatment, the

findings in this subsection imply that more members from the discussion treatment

groups end up without access to savings or loans. Moreover, these are disproportion-

ately the members who were initially poor or economically vulnerable. Given that

access to such financial services is instrumental for poverty reduction and that the ob-

jective of the groups was to offer access to such services, the overall welfare of original

discussion group members is lower at endline than that of vote group members.

32We asked all respondents at endline if they were currently saving (financially) somewhere elsethan in the BRAC savings and loans group. 29 % of members report saving elsewhere, compared to33% of non-members. Out of the non-members who save, 84% keep savings with another institution,while the remaining 16% report that they only keep saving with someone they know or at home.In comparison, among stayers, over 95% of those who save elsewhere do so with another institutionand only 4.65% keep their savings at home. We also asked all respondents about loans from othersources than the BRAC group. The share of respondent that have had access to such loans is similaracross dropouts and stayers, with a slightly higher share among dropouts (26% compared to 22%among stayers). The majority of those loans have been provided by other village savings and loansassociations. However, compared to dropouts, stayers clearly have had higher access to loans fromthe BRAC group.

2.8. DISCUSSION AND CONCLUDING REMARKS 41

2.8 Discussion and Concluding remarks

In this paper, we estimate the causal effect of the electoral rule used for leader selection

on the types of leaders and the performance of community savings and loans groups.

We specifically examine leader characteristics and group service delivery in terms of

continued membership of the vulnerable, and the savings and loans of individual mem-

bers over a period of 3.5 years.

We find that groups that chose their leaders in a secret ballot vote selected leaders

that were more representative of the average group member, as compared to leaders

appointed in an open discussion where leaders were positively selected in terms of

socio-economic characteristics. The dropout was high in both types of treatments but

substantially higher in discussion groups, which lost 60% of their original members

in their first year after committee selection, as compared to a 45% dropout rate in

vote groups. Moreover, in discussion groups, the dropout was substantially higher

for members with lower values on economic performance indicators at the baseline,

if measured in terms of wealth, income, loans or savings. Finally, at the endline, 3.5

years after selecting their first committee, vote groups are larger in size than discussion

groups. Also, their members are less likely to be active savers, while those who borrow

take smaller and fewer loans. We conclude that the secret poll voting creates more

inclusive groups while open discussion generates groups that are more exclusive and

selective in favor of more economically successful members.

This paper contributes to the understanding of public goods delivery in developing

countries, where community groups are assuming the responsibility for a large fraction

of the delivery of social and financial services to the poor. In particular, it is of relevance

for the growing literature on Community Driven Development (CDD). Knowledge of

how to set up inclusive mechanisms for local governance in such programs is important

for their long-term sustainability.

When interpreting the results in terms of external validity, it is important to no-

tice that we study savings groups where all members are female, young, have a low


education and are relatively poor, and where members of a group are recruited from

within the same community. Compared to other contexts in which decision making

and different democratic systems have been studied, ours is a very homogenous set-

ting. The results we find are likely to be a lower bound on the potential differences

that can arise from open discussion as compared to secret voting. An interesting exten-

sion would be to analyze similar treatments in more diverse groups where mechanisms

such as intimidation can be more pronounced and treatment effects are likely to be

stronger.

This study offers complementary explanations to the literature on efficiency and

inclusion in community groups in Sub-Saharan Africa. Our explicit focus on governance

and the role of group leaders enables us to reveal one potential mechanism behind the

exclusion of less economically able members and reconcile findings of previous studies.

Our findings suggest that the discussion treatment can lead to more efficient savings

groups through selection, but that this happens at the expense of more inequality in

access to savings and credit. In terms of mechanism, we provide supportive evidence

for the effect working through leader representativeness rather than leader skill level.

The appropriate method for leader selection ultimately depends on the objective and

target group of a program.

REFERENCES 43

References

Alesina, A., and E. La Ferrara. 2002. Who trusts others? Journal of public eco-

nomics, 85(2), 207-234.

Aggarwal, S., L. F. Klapper and D. Singer. 2012. Financing businesses in Africa:

The role of microfinance. World Bank Policy Research Working Paper, (5975).

Appleton, S. 2003. Regional or national poverty lines? The case of Uganda in the

1990s. Journal of African Economies, 12(4), 598-624.

Baland, J. M. and J. A. Robinson. 2008. Land and Power: Theory and Evidence


Baland, J. M., R. Somanathan and L. Vandewalle. 2011. Socially disadvantaged



Bandiera, O. N. Buehren, R. Burgess, M. Goldstein, S. Gulesci, I. Rasul

and M. Sulaiman. 2014. Women’s Empowerment in Action: Evidence from a Ran-

domized Control Trial in Africa (No. 50). Suntory and Toyota International Centres

for Economics and Related Disciplines, LSE.

Beath, A., F. Christia and R. Enikolopov. 2012. Direct democracy and resource

allocation: experimental evidence from Afghanistan. World Bank Policy research work-

ing paper WPS6133.

Beath, A., F. Christia, G. Egorov and R. Enikolopov. 2014. Electoral rules and

the quality of politicians: theory and evidence from a field experiment in Afghanistan

(No. w20082). National Bureau of Economic Research.

Bernard, T., M-H. Collion, A. de Janvry, P. Rondot, and E. Sadoulet. 2008.

Do Village Organizations Make a Difference in African Rural Development? A Study

for Senegal and Burkina Faso. World Development 36(11): 2188–2204.

Besley, T., and S. Coate. 1997. An economic model of representative democracy.

The Quarterly Journal of Economics, 85-114.


Besley, T. 2005. Political selection. The Journal of Economic Perspectives, 19(3),

43-60.

Besley, T., R. Pande and V. Rao. 2005. Participatory democracy in action: Survey

evidence from South India. Journal of the European Economic Association, 3(2?3),

648-657.

Burlando A., and A. Canidio. 2015. Financial inclusion of vulnerable households

through Savings and Borrowing Groups: Theory and Evidence from Uganda.



1812.

Cox, G. W. 1997. Making votes count: strategic coordination in the world’s electoral

systems (Vol. 7). Cambridge: Cambridge University Press.

Czuba, K. 2011. "Even if I have land, Can I eat it?" Perceptions of poverty in

Karamoja, BRAC Research and Evaluation Unit report.

Czuba, K. 2012a. Income Generating Activities and Savings Behaviour of Adolescent

Girls and Young Women in Karamoja, BRAC Research and Evaluation Unit report.

Czuba, K. 2012b. The causes and mechanisms of Karimojong child migration, IOM

Report.

Dal Bó , P., A. Foster, and L. Putterman. 2010. Institutions and Behavior: Ex-

perimental Evidence on the Effects of Democracy. American Economic Review 100(5):

2205-29.

Dupas, P., and J. Robinson. 2013a. Why Don’t the Poor Save More? Evidence

from Health Savings Experiments. American Economic Review, 103(4): 1138-71.

Dupas, P., and J. Robinson. 2013b. Savings Constraints and Microenterprise De-

velopment: Evidence from a Field Experiment in Kenya. American Economic Journal:

Applied Economics, 5(1), 163-192.

REFERENCES 45





961-993.

Huber E., C. Ragin and J.D. Stephens. 1993. Social democracy, Christian democ-

racy, constitutional structure, and the welfare state. American journal of Sociology,

711-749.

Humphreys, M., Masters, W., and M.E. Sandbu. 2006. The Role of Leaders in

Democratic Deliberations: Results from a Field Experiment in São Tomé and Príncipe.

World Politics, 58 (4), 583-622.

Grameen foundation. 2011. Progress out of poverty Index for Uganda, 2009/2010.

Grameen foundation. 2015. Progress out of poverty Index for Uganda, 2012/2013,

retrieved on July 1, 2015, http://www.progressoutofpoverty.org/country/uganda.

Greaney, B., J.P. Kaboski and E. Van Leemput. 2016. Can Self-Help Groups

Really Be "Self-Help"? FRB International Finance Discussion Paper No. 1155.

Ksoll, C., Lilleør, H., Lønborg, J. H. and O.D. Rasmussen. 2013. Impact of

village savings and loans associations: Evidence from a cluster randomized trial (No.

56). Rockwool Foundation Research Unit.

Madajewicz M., A. Tompsett and A. Habib. 2014. Community Participation in

Decision-Making: Evidence from a Field Experiment in Bangladesh.

Mansuri, G. and V. Rao. 2012. Localizing development: does participation work?.


Matsusaka, J.G. 2005. Direct democracy works. Journal of Economic Perspectives,

185-206.

Murphy, W. P. 1990. Creating the appearance of consensus in Mende political dis-

course. American anthropologist, 24-41.


Olken, B. A. 2010. Direct democracy and local public goods: Evidence from a field

experiment in Indonesia, American Political Science Review 104.02: 243-267.

Osborne, M. J., and A. Slivinski. 1996. A model of political competition with

citizen-candidates. The Quarterly Journal of Economics, 65-96.

Persson, T. and G. E. Tabellini. 2000. Political economics: Explaining economic

policy. Cambridge, MA: MIT Press.

Persson, T., and G. E. Tabellini. 2005. The economic effects of constitutions. MIT

press.

Pitkin, H. F. 1967. The Concept of Representation. Berkeley, CA: University of

California Press.

Uganda Poverty Status Report. 2014. Structural change and poverty reduction

in Uganda. Ministry of Finance, Planning and Economic Development, Republic of

Uganda.

Vsla’s website. 2015. INGO-SG global Outreach 2015-11, Retrieved November 10,

2015, vsla.net.

FIGURES AND TABLES 47

Figures and Tables

Figure 2.1: Timeline

November 2012 November 2014 May 2015

Data collection

Census collected

March 2015

Endline survey collected May-‐

July 2015

Program implementation

BRAC Karamoja saving groups: Timeline for program implementation and data collection activities

Follow up data collected March 2013

May 2013

Groups start lending June-‐Aug 2012

BRAC Saving groups started in Karamoja May-‐June 2011

Baseline survey collected Oct-‐Dec

2011

May 2012May 2011

Randomization + committee

choice Jan 2012

November 2011


Figure 2.2: Position in group wealth and asset distribution, leaders vs non leaders.0

4.0

5.0

6.0

7.0

8.0

9.1

.11

Kern

el D

ensi

ty

0 2 4 6 8 10Discussion groups: Wealth score

Non CMs CMs (Leaders).0

4.0

5.0

6.0

7.0

8.0

9.1

.11

Kern

el D

ensi

ty

0 2 4 6 8 10Vote groups: Wealth score

Non CMs CMs (Leaders)

.04

.06

.08

.1.1

2K

erne

l Den

sity

0 2 4 6 8 10Discussion groups: Log asset distr.


.04

.06

.08

.1.1

2K

erne

l Den

sity

0 2 4 6 8 10Vote groups: Log asset distr.


Note: The top panel shows the wealth score distribution at baseline separately for regular members (Non CMs) andcommittee members (CMs (Leaders)), by treatment. The wealth score is constructed using the Uganda PPI indexcompiled by the Grameen foundation (2011), and the value distribution indicates the position (decile) of anindividual in the wealth score distribution of her savings group. Higher value=less poor. Kernel density plot;Epanechikov Kernel, bandwidth: 2. A Kolmogorov-Smirnov test rejects equality of distributions in discussion (p-value0.01) and cannot reject equality of distributions in vote (p-value 0.8). The bottom panel shows the position of amember in the log asset distribution of her group at baseline separately for regular members (Non CMs) andcommittee members (CMs (Leaders)), by treatment. Kernel density plot; Epanechikov Kernel, bandwidth: 1.7. AKolmogorov-Smirnov test rejects equality of distributions in discussion (p-value 0.015) and cannot reject equality ofdistributions in vote (p-value 0.9).


Figure 2.3: Share of initial members staying in group

0.2

.4.6

Mem

bers

hare

sta

ying

2013 2015

VoteDiscussion

Note: Figure 2.3 above shows the share of original members that were stayers in March 2013 and in March 2015,respectively, by treatment. The bars indicate a 95% confidence interval around the mean. In 2013, the differencebetween the staying rate in vote and discussion groups is significant at the 99 percent confidence level (p-value 0.006,regression estimates displayed in Table 2.4). For 2015, the difference between the staying rate in vote and discussiongroups is just below the conventional confidence level with a p-value of 0.11 (Table 2.4).

Note: Figure 2.4 below shows the distribution of group size in 2015 by treatment. Kernel density plot;Epanechikov Kernel, bandwidth 3.5.

Figure 2.4: Group size 2015

0.0

2.0

4.0

6Ke

rnel

Den

sity

0 10 20 30Groupsize 2015

Vote Discussion

50 ELECTORAL RULES AND LEADER SELECTIONTable

2.1:Baseline

variables,treatment

balancecheck

Full

sample

mean

st.dev

Discussion

treatmVote

treatmAdjusted

p-valuemean

(N=46)

mean

(N=46)

Difference

Grou

pC

haracteristics

Average

initialIGFgroup

size(2011)

205.581

20.0219.98

0.0650.954

Initialcommittee

size5.033

0.2755.07

5.000.056

0.375Share

ofCMs2011

thatare

inBaseline

data0.705

0.2260.69

0.72-0.049

0.267Share

ofall2011

mem

bersin

Baseline

data0.843

0.1350.85

0.840.001

0.954G

roup

perform

ance

Characteristics

Shareof

mem

berswho

keptsavings

atbaseline

0.930.164

0.940.92

0.0160.617

Mean

baselinesavings

keptby

groupmem

bers,1000sUGX

12.7517.278

12.7712.73

0.0800.868

Shareof

mem

bersthat

hasany

debt(not

with

BRAC)at

baseline0.101

0.1610.11

0.090.022

0.545Mean

amount

ofloans

outstandingat

baseline,1000sUGX

0.8291.554

1.010.64

1.0280.269

Mem

ber

hou

sehold

characteristics

Mean

householdsize

5.9841.115

5.876.10

-0.2670.259

Mean

wealth

scoreof

household19.865

9.17321.03

18.702.154

0.121Mean

totalvalueof

assetsheld

atbaseline,1000s

UGX

1831.0121194.318

1908.951753.07

500.3030.438

Shareof

householdswho

ownland

0.7230.239

0.720.72

0.0100.832

Shareof

householdswho

ownhouse

0.9020.141

0.890.91

-0.0170.355

Shareof

householdmem

bersyounger

than18

0.5330.083

0.530.54

-0.010.509

Shareof

householdmem

bers(above

9y)who

areworking

0.360.23

0.340.38

-0.0260.317

Shareof

householdmem

bersaged

6-18with

someschool

0.6420.216

0.620.67

-0.0490.170

Shareof

householdmem

bersaged

6-14with

primary

school0.038

0.0450.03

0.04-0.007

0.436M

ember

characteristics

Mean

mem

berage

201118.534

1.63618.63

18.440.056

0.826Share

ofmem

berswho

arethe

main

earnerin

herhousehold

0.2710.203

0.260.28

-0.0250.468

Sharewith

noschooling

0.4260.257

0.460.39

0.0940.049

Sharewith

primary

school0.171

0.1410.16

0.18-0.028

0.332Share

ofmem

berswho

evergot

trainingin

IGA

0.2270.248

0.280.18

0.1030.007

Mean

totalearningsduring

last12

months,1000s

UGX

448.369330.126

477.49419.24

62.7390.320

Shareof

mem

bersever

married

orhad

partner0.836

0.2120.82

0.85-0.03

0.479Mean

number

ofchildren

1.8191.014

1.701.94

-0.3050.090

Mean

ageat

birthof

firstchild

18.2291.448

18.4018.06

0.4560.102

Sharefriends

ingroup

0.1650.058

0.170.16

0.0060.523

Shareof

mem

berswith

migration

experience0.303

0.2330.29

0.32-0.023

0.592

Notes:

Com

mittee

sizedeviated

from5in

fourcases:

One

grouphad

4com

mittee

mem

bers,2groups

had6com

mittee

mem

bersand

onegroup

had7com

mittee

mem

bers.Not

alloriginalgroupmem

bersor

originalcommittee

mem

bersare

representedin

thebaseline

data.Asthe

firstsection

ofthetable

shows,reassuringly,the

sharesinterview

edare

comparable

acrosstreatm

ents.Wealth

scoreis

thescore

obtainedin

anindex

constructedby

theGram

eenFoundation

(2011)measuring

povertyindicators

inUganda.

Higher

value=

lesspoor.

Sharefriends

ingroup:

anorm

alizeddegree

centralitymeasure

basedon

howmany

othermem

bersnam

edher

among

their2best

friendsin

thegroup

atbaseline,

normalized

bythe

number

ofother

mem

bersin

thegroup

atbaseline.C

olumns1-2

reportthe

mean

andstandard

deviationfor

thefull

baselinesam

ple.Colu

mns3-4

reportthe

means

indiscussion

treatment

andvote

treatment,respectively.D

ifferen

ceand

P-valu

esare

fromW

LSregressions

ofeach

characteristic’sgroup

mean

ontreatm

ent,controllingfor

branchfixed

effects,andweighting

bythe

fractionof

initialmem

bersin

thegroup

interviewed

atbaseline.*

p<0.1,**

p<0.05,***

p<0.01.


Tab

le2.2:

Differencesbe

tweenlead

ers(com

mitteemem

bers)across

treatm

ents,S

ocio-econo

mic

characteristics(sam

plerestricted

tolead

erson

ly)

Econo

mic

variab

les

Socioecon

omic

Social

conn

ection

prox

ies

Wealthscore

Logassets

Has

some

Employm

Migration

Has

Age

atMajority

Sharefriend

s#

Link

sto

distribu

tion

2011

educ

ation

netw

ork

expe

rience

child

ren

1stbirth

tribe

ingrou

potherCMs

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Votetreatm

-0.684∗∗

-0.218

-0.017

-0.063

-0.122∗

0.180∗∗∗

-0.878∗∗∗

0.065

-0.004

0.115

[0.334]

[0.239]

[0.065]

[0.052]

[0.063]

[0.062]

[0.330]

[0.040]

[0.018]

[0.132]

Discussionmean

5.766

13.656

0.660

0.333

0.357

0.671

18.710

0.884

0.200

0.553

Fixed

Effe

cts

bran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chIG

ACon

trol

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

Observation

s312

292

304

323

307

310

225

184

279

323

Adjusted

R2

0.034

0.170

0.087

0.169

0.088

0.109

0.125

0.126

0.094

0.094

Note:

Wealthscoredistribution

indicatesthepo

sition

(decile

)of

anindividu

alin

thewealthscoredistribu

tion

ofhergrou

p.The

scoreis

constructedusingtheUgand

aPPI

indexcompiledby

theGrameenfoun

dation

(2011).Highervalue=

less

poor.Lo

gAssets20

11:The

logged

valueof

assets

atba

selin

e,cond

itiona

lon

having

nonzeroassets.Has

noeducationdu

mmy=

1iftheindividu

alha

sno

scho

oling.

Employmentnetwork:

dummy=

1iftheindividu

alha

din

thepa

styear

received

help

from

anyone

outsidetheclose

family

formatters

ofem

ployment.

Thisis

aproxyforha

ving

labo

rmarketconn

ection

s.Migration

experience:du

mmy=

1iftheindividu

alha

sever

lived

outsidethevilla

ge.

Migration

isdo

neforshorttimeworkor

stud

iesan

dis

positively

correlated

withassets

andwealth.

Has

child

ren:du

mmy=

1iftherespon

dent

hasan

ychild

ren.

Age

at1st

birth:Indicatestheindividu

al’s

agewhenshefirst

gave

birth.

Low

erageindicatesaworse

socio-econ

omic

situation.

Onlyavailableforthosewho

have

child

renMajoritytribe:

dummy=

1iftheindividu

albe

long

sto

themostcommon

tribein

hersaving

sgrou

pSh

arefriend

sin

grou

p:ano

rmalized

degree

centralitymeasure

basedon

how

man

yother

mem

bers

named

heram

ongtheir2be

stfriend

sin

thegrou

pat

baselin

e,no

rmalized

bythenu

mbe

rof

othermem

bers

inthegrou

pat

baselin

e#

Links

toCMs:

thenu

mbe

rof

individu

alsna

med

amon

gher4be

stfriend

sat

baselin

ewho

laterbe

camecommitteemem

bers

(equ

ivalentto

thenu

mbe

rof

CMsin

this

sampleon

lyrestricted

toCMs)6=

therespon

dent.D

iscussionmean:The

meanvalueof

each

outcom

evariab

lein

thediscussion

treatm

ent.IG

ACon

trol:thesharethat

hadreceived

incomegenerating

activity

training

from

BRAC,is

includ

edas

acontrolin

allregression

sto

accoun

tforim

balancein

this

variab

leacross

treatm

ents

atba

selin

e.Rob

uststan

dard

errors

inbrackets,

clusteredat

thegrou

plevel(level

ofrand

omization).Regressions

areweigh

tedforthefraction

ofinitialcommitteemem

bers

interviewed

atba

selin

e,an

dinclud

ebran

chfix

edeff

ects.*p<

0.1,

**p<

0.05,***p<

0.01.


Table

2.3:Baseline

characteristicsof

leaders(com

mittee

mem

bers)com

paredto

othergroup

mem

bers(sam

ple:alloldmem

bers)

Econom

icvariables

Socioeconom

icSocialconnection

proxies

Wealth

LogAssets

Has

some

Empl

Migration

Has

Age

atMajority

Sharestrong

ShareCMs

score2011

educationNetw

orkexperience

children1st

birthtribe

linkslinked

to(1)

(2)(3)

(4)(5)

(6)(7)

(8)(9)

(10)

Leader3.940 ∗∗∗

0.1580.122 ∗∗∗

0.129 ∗∗∗0.057 ∗

0.132 ∗∗∗-0.012

-0.0350.044 ∗∗∗

0.030[1.244]

[0.141][0.040]

[0.031][0.034]

[0.043][0.408]

[0.024][0.015]

[0.026]Vote*Leader

-2.946 ∗-0.195

-0.080-0.123 ∗∗∗

-0.179 ∗∗∗0.063

-0.5480.097 ∗∗

-0.0040.066

[1.579][0.192]

[0.064][0.045]

[0.052][0.060]

[0.482][0.038]

[0.018][0.040]

Leader+Vote*Leader

0.994-0.037

0.0430.005

-0.122 ∗∗∗0.195 ∗∗∗

-0.561 ∗∗0.062 ∗

0.040 ∗∗∗0.096 ∗∗∗

[0.973][0.131]

[0.049][0.032]

[0.039][0.042]

[0.256][0.030]

[0.010][0.031]

Discussion

mean

19.50913.485

0.4870.279

0.2830.531

18.5570.542

0.1350.109

Fixed

effectsgroup

groupgroup

groupgroup

groupgroup

groupgroup

groupObservations

14491351

13941483

13701408

8401483

11071483

Adjusted

R2

0.3630.269

0.2100.330

0.2220.225

0.0780.690

0.1950.211

Note:

Wealth

scoreindicates

theposition

(decile)of

anindividualin

thewealth

scoredistribution

ofher

group.The

scoreisconstructed

usingthe

Uganda

PPIindex

compiled

bythe

Gram

eenfoundation

(2011).Higher

value=less

poor.Log

Assets

2011The

loggedvalue

ofassets

atbaseline,

conditionalon

havingnonzero

assets.Has

noeduc.:

dummy=

1ifthe

individualhas

noschooling.

Employm

entnetw

ork:dum

my=

1ifthe

individualhad

inthe

pastyear

receivedhelp

fromanyone

outsidethe

closefam

ilyfor

matters

ofem

ployment.

This

isaproxy

forhaving

labormarket

connections.Migration

experience:dum

my=

1ifthe

individualhas

everlived

outsidethe

village.Migration

isdone

forshort

timework

orstudies

andispositively

correlatedwith

assetsand

wealth.

Has

children:dum

my=

1ifthe

respondenthas

anychildren.

Age

at1st

birth:Indicates

individual’sage

when

shefirst

gavebirth.

Alow

erage

indicatesaworse

socio-economic

situation.Only

availablefor

thosewho

havechildren.

Majority

tribe:dum

my=

1if

theindividual

belongsto

themost

common

tribein

hersavings

group,Share

friendsin

group:anorm

alizeddegree

centralitymeasure

basedon

howmany

othermem

bersnam

edher

among

their2best

friendsin

thegroup

atbaseline,

normalized

bynum

berof

othermem

bersin

groupat

baselineShare

CMslinked

to:the

number

ofindividuals

named

among

her4best

friendsat

baselinewho

laterbecam

ecom

mittee

mem

bers,divided

bythe

number

ofcom

mittee

mem

bers6=

therespondent.

Discussion

mean

:Mean

valueam

ongnon-leaders

inthe

discussiontreatm

ent.Robust

standarderrors

inbrackets,

clusteredat

thegroup

level(level

ofrandom

ization).Regressions

areweighted

forthe

fractionof

initialcom

mittee

mem

bersinterview

edin

baseline,and

includebranch

fixedeffects.

*p<0.1,

**p<0.05,

***p<0.01.


Table 2.4: Dropout over time, all old members

Dep. var Dropout 2013 Dropout 2015(1) (2) (3) (4)

Vote treatm -0.149∗∗∗ -0.143∗∗∗ -0.062∗∗∗ -0.064[0.025] [0.051] [0.022] [0.040]

Discussion mean 0.607 0.607 0.727 0.727Fixed effects none branch none branchObservations 1624 1624 1811 1811Adjusted R2 0.022 0.232 0.004 0.190

Note: The dependent variable is an indicator of having dropped out by2013 (columns 1-2) or 2015 (columns 3-4). Columns (1) and (3): OLS re-gression with robust standard errors. Columns (2) and (4): OLS regressionwith robust standard errors clustered at group level, and controlling forbranch fixed effects. p-value for dropout in 2015 is 0.11. The fewer ob-servations in 2013 compared to 2015 are due to the followup data beingcollected using an incomplete list of original members. 2013. Results arerobust to imputing 2015-dropout-status for the 187 members missing in2013. * p<0.1, ** p<0.05, *** p<0.01


Table

2.5:Dropout

by2013

relatedto

economic

variablesat

baseline,alloldmem

bers,extensivemargin

Poverty

proxy:Poor

(bottom25%

Poor

(bottom25%

Has

nomarket

Has

nosavings

Has

noloans

ofwealth

distr.)of

assetdistr.)

incomein

2011in

2011in

2011(1)

(2)(3)

(4)(5)

Characteristic

0.071 ∗∗0.046

0.103 ∗∗0.194 ∗∗

0.245 ∗∗[0.034]

[0.028][0.047]

[0.054][0.086]

Vote*characteristic

-0.065-0.069 ∗

-0.153 ∗∗-0.187 ∗

-0.133[0.047]

[0.040][0.071]

[0.098][0.123]

Vote

-0.103 ∗-0.113 ∗∗

-0.093-0.115 ∗

-0.019[0.059]

[0.056][0.061]

[0.059][0.129]

Discussion

mean

0.5770.582

0.5650.581

0.458P-value

Disc.=

Vote

forChar=

10.011

0.0040.002

0.0020.008

P-value

Char0

=Char1

inVote

0.8630.437

0.3420.940

0.170Fixed

Effects

branchbranch

branchbranch

branchIG

AControl

yesyes

yesyes

yesObservations

1,4061,312

1,3671,356

1,273Adjusted

R2

0.2220.247

0.2330.233

0.237

Note:

The

dependentvariable

inall

regressionsis

anindicator

ofhaving

droppedout

by2013.

The

independentvariables

aredum

mies

proxyingfor

povertyin

differentways.

The

2first

columns

indicatebeing

inthe

bottom25%

ofthe

wealth

distributionor

assetdistribution,

respectively,of

one’sgroup.

Has

nomarket

incomeindicates

beingeither

asubsistence

farmer

oradependent

atbaseline.

Has

nosavings

indicatesnot

savingat

thetim

eof

thebaseline

surveyand

Has

noloans

indicatesnot

havingtaken

loansfrom

otherlenders

thanBRAC

(atthe

timeof

baseline,theBRAC

groupshad

notstarted

givingout

loans).The

resultsare

robustto

reducingthe

sample

onlyto

individualswith

non-missing

valuesfor

theasset,

wealth

score,incom

e,savings

andloans

indicatorvariables.D

iscussionmean

:Shareofdropouts

inthe

discussiontreatm

entfor

mem

berswith

independentvariable

=0.p-value

Discussion

=Vote

forChar=

1:pvalue

fromatest

ofwhether

dropoutrate

isthe

sameacross

treatments

forindividuals

with

povertycharacteristic=

1.p-value

Char0

=Char1

inVote

isfrom

atest

ofwhether

dropoutsare

similar

(interm

sofeach

povertycharacteristic)

tostayers

inthe

votetreatm

ent.IGA

Control:the

sharethat

hadreceived

incomegenerating

activitytraining

fromBRAC,is

includedas

acontrolin

allregressionsto

accountfor

imbalance

inthis

variableacross

treatments

atbaseline.

Robust

standarderrors

inbrackets,

clusteredat

thegroup

level(level

ofrandom


includebaseline

sample

weights,

andbranch

fixedeffects.

*p<0.1,

**p<0.05,

***p<0.01.


Table 2.6: Savings in 2015, extensive and intensive margin, all old members

Sample: All Stayers Dropouts

Dep.var.: Extensive Intensive Extensive Intensive Extensive Intensive Extensive IntensiveAnywhere Anywhere In BRAC group Anywhere

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

Vote treatm -0.035 -0.127 -0.090∗ -0.206 -0.110∗ -0.114 -0.012 -0.088[0.050] [0.153] [0.050] [0.202] [0.061] [0.188] [0.059] [0.244]

Discussion mean 0.593 10.962 0.831 10.794 0.757 10.398 0.360 11.357Fixed effects branch branch branch branch branch branch branch branchObservations 878 421 436 302 435 267 442 119Adjusted R2 0.034 0.106 0.129 0.134 0.173 0.144 0.020 0.072

Note: The dependent variable in columns 1 and 3 and 7 is an indicator for having Savings, anywhere, in 2015. The depen-dent variable in column 5 is an indicator for having Savings in the BRAC group in 2015, and is only available for stayers. Thedependent variable in columns 2 and 4 and 8 is log Savings in 2015 kept anywhere, conditional on having nonzero savings. Thedependent variable in column 6 is log Savings in 2015 in the BRAC group, conditional on having nonzero savings, and is onlyavailable for stayers. The mean value of the dependent variable for the discussion treatment. is displayed below the table.Standard errors in brackets are clustered at the group level (level of randomization). Regressions include endline sample weights,and branch fixed effects. * p<0.1, ** p<0.05, *** p<0.01.

Table 2.7: Loans in BRAC group by 2015, All old members; stayers and dropouts

Sample: All Stayers Dropouts

Dep.var.: Extensive Intensive # Loans Extensive Intensive # Loans Extensive Intensive # Loans(1) (2) (3) (4) (5) (6) (7) (8) (9)

Vote group -0.047 -0.035 -0.066 -0.083 -0.210 -0.202 -0.039 0.175 -0.021[0.048] [0.138] [0.098] [0.062] [0.182] [0.161] [0.052] [0.180] [0.095]

Discuss. mean 0.407 11.346 0.717 0.553 11.425 1.014 0.272 11.224 0.438Fixed Effects branch branch branch branch branch branch branch branch branchObservations 896 274 895 437 175 436 459 99 459Adjusted R2 0.094 0.125 0.112 0.134 0.163 0.146 0.062 0.138 0.078

Note: Loans Extensive margin: a dummy=1 if the individual has ever taken a loan from the BRAC saving group by 2015).# Loans: The number of loans taken with BRAC group by 2015, including zero loans. Loans Intensive margin: log value ofloans taken with BRAC group by 2015 conditional on having taken a loan. The mean value of the dependent variable for thediscussion treatment. is displayed below the table. Standard errors in brackets are clustered at the group level (level of rand-omization). Regressions include endline sample weights, and all regressions include branch fixed effects.* p<0.1, ** p<0.05, *** p<0.01.


Table

2.8:Savingsand

loansin

2015;New

mem

bersonly

Dep.var.:

SavingsLoans

Extensive

margin

Intensivemargin

Extensive

margin

#Loans

Intensivemargin

Anyw

hereBRAC

Anyw

hereBRAC

Anyw

hereBRAC

BRAC

BRAC

(1)(2)

(3)(4)

(5)(6)

(7)(8)

Vote

treatm-0.060 ∗∗

-0.0140.309 ∗

0.309-0.017

-0.018-0.273 ∗

-0.051[0.028]

[0.044][0.177]

[0.210][0.045]

[0.057][0.152]

[0.133]

Discuss.m

ean0.945

0.86910.896

10.4340705

0.6231.285

11.663Fixed

Effects

branchbranch

branchbranch

branchbranch

branchbranch

Observations

383383

240240

401401

401181

Adjusted

R2

0.1230.156

0.2120.269

0.1770.216

0.1260.160

Note:

SavingExtensive

margin

:adum

my=

1ifthe

individualhas

savingsin

2015,anyw

here(colum

n1)

orwith

BRAC

(column2).

SavingsIntensive

margin

:log

valueof

savingsin

2015,anyw

here(colum

n3)

orwith

BRAC

(column4),

conditionalon

havingnonzero

savings.Loans

Extensive

margin

:adum

my=

1ifthe

individualhas

evertaken

aloan

by2015,

anywhere

(column5)

orwith

BRAC

(column6).#

LoansThe

number

ofloans

takenwith

BRAC

by2015,

includingzero

loans.Loans

Intensivemargin

:log

valueof

loanstaken

with

BRAC

by2015

conditionalon

havingtaken

aloan.

Wedo

nothave

surveyinform

ationabout

theam

ountof

loansnor

ofthe

number

ofloans

takentaken

outsideof

theBRAC

group.The

mean

valueof

thedependent

variablefor

thediscussion

treatment.

isdisplayed

belowthe

table.Standard

errorsin

bracketsare

clusteredat

thegroup

level(level

ofrandom


includeendline

sample

weights,

andall

regressionsinclude

branchfixed

effects.*p<0.1,

**p<0.05,

***p<0.01.


Tab

le2.9:

Loan

allocation

amon

gstayers,

measuredin

2015

Dep.v

ar.:

Loan

dummy

#Lo

ansgiven

Total

logloan

amou

ntPoverty

proxy:

Poo

rNoincome

Noloan

Poo

rNoincome

Noloan

Poo

rNoincome

Noloan

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Cha

racteristic*Vote

0.207∗

0.060

0.296∗

0.411∗

0.013

0.550

1.915

1.423

3.792∗

[0.116]

[0.123]

[0.151]

[0.216]

[0.303]

[0.416]

[1.373]

[1.526]

[1.967]

Cha

racteristic

-0.093

-0.040

-0.272∗∗

-0.351∗∗

-0.038

-0.639∗∗

-1.018

-0.796

-3.003∗

[0.091]

[0.086]

[0.121]

[0.176]

[0.263]

[0.308]

[1.098]

[1.149]

[1.630]

Vote

-0.133∗

-0.111

-0.343∗∗

-0.306∗

-0.297∗

-0.729∗

-1.372∗

-1.449

-4.353∗∗

[0.067]

[0.074]

[0.143]

[0.166]

[0.169]

[0.395]

[0.809]

[0.913]

[1.893]

Discussionmean

0.564

0.576

0.650

1.053

1.104

1.350

5.630

5.731

6.827

Fixed

Effe

cts

bran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chObservation

s359

347

311

358

346

310

310

297

267

AdjustedR

20.131

0.132

0.189

0.136

0.142

0.191

0.129

0.131

0.178

Note:

Loan

dummy=

1ifthemem

berha

sever

takenaloan

intheBRAC

IGF

grou

p#

Loan

sgiven:Num

berof

loan

sever

takenfrom

theIG

Fgrou

pTotal

loan

amou

nt:T

helogam

ount

oftheloan

staken,

totalm

argin.

Discussionmean:M

eanvalueof

thedepe

ndentvariab

leam

ongmem

bers

inthediscussion

grou

pswithpo

vertyprox

y=0.

Stan

dard

errors

inbrackets

areclusteredat

grou

plevel.*p<

0.1,

**p<

0.05,***p<

0.01


Table 2.10: Default on loans, reported in 2015

Dep. var.: Dummy=1 if ever defaulted on IGF loan

Sample: All old members Stayers only

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

Vote 0.0140 0.014 0.004 0.024[0.0408] [0.056] [0.051] [0.071]

Poor -0.069 -0.059[0.069] [0.121]

Poor*Vote 0.048 0.041[0.094] [0.149]

Discussion mean 0.102 0.102 0.111 0.111Fixed Effects branch branch branch branchObservations 354 278 249 204Adjusted R2 0.002 0.017 0.048 0.0739

Note: Note: Default loan: A dummy =1 if the member has ever defaultedon a loan taken in the group. The sample is restricted to members whohave ever taken a loan from the group. Discussion mean: Mean value ofthe dependent variable among non-poor members in the discussion groups.Standard errors in brackets are clustered at the group level (level of ran-domization). Regressions include endline sample weights, and all regres-sions include branch fixed effects. * p<0.1, ** p<0.05, *** p<0.01.

Table 2.11: Savings and loans in 2015 for dropouts compared to staying members

Dep var: Loan extensive margin Saving extensive margin Saving intensive margin

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

Dropout -0.189∗∗∗ -0.207∗∗∗ -0.429∗∗∗ -0.466∗∗∗ 0.395∗∗ 0.338[0.035] [0.051] [0.045] [0.066] [0.186] [0.212]

Vote -0.055 -0.079 -0.160[0.054] [0.055] [0.202]

Dropout*Vote 0.034 0.073 0.106[0.069] [0.088] [0.339]

Discussion mean 0.648 0.659 0.808 0.831 10.773 10.794Fixed Effects branch branch branch branch branch branchObservations 896 896 878 878 421 421Adjusted R2 0.136 0.135 0.209 0.210 0.121 0.119

Note: Dropout is a dummy=1 if the respondent has dropped out, 0 if the respondent is still a member of theBRAC saving group Extensive margin: a dummy=1 if the individual has savings (has taken loan) anywhere in2015. Intensive margin: log value of savings in 2015, conditional on having nonzero savings. We do not have surveyinformation about the amount of loans nor of the number of loans taken taken outside of the BRAC group. The meanvalue of the dependent variable for stayers is displayed below the table for columns 1, 3 and 5. Discussion mean:The mean value of the dependent variable for stayers in the discussion treatment is displayed below the table forcolumns 2, 4 and 6. Standard errors in brackets are clustered at the group level (level of randomization). Regressionsinclude endline sample weights, and all regressions include branch fixed effects. * p<0.1, ** p<0.05, *** p<0.01.

APPENDIX 1 59

Appendix 1


Table A.1: Differences between leaders (committee members) across treatments, alternative economicvariables

Wealth score raw Asset score Income score Log income 2011(1) (2) (3) (4)

Vote -3.856∗ -0.438 -0.526 -0.385∗[2.118] [0.331] [0.402] [0.227]

IGA Control 5.418 1.724∗ -1.331 0.265[4.214] [0.903] [1.085] [0.539]

Fixed Effects branch branch branch branchObservations 312 302 304 232Adjusted R2 0.240 0.016 0.009 0.228

Note:Wealth score raw is compiled at the household level using the Uganda Progress out of povertyindex constructed by the Grameen foundation (2011). Contains variables measuring the household’s(inverse) poverty status. Higher value=less poor. Asset (income) score indicates the position (decile)of an individual in the asset(income) distribution of her group at baseline. Log income 2011 is the logvalue of income at baseline conditional on having nonzero income. Robust standard errors in brackets,clustered at the group level (level of randomization). IGA Control : Control for the share that hadreceived income generating activity training from BRAC, included as control in all regressions toaccount for imbalance in this variable across treatments at baseline. and include sample weights forthe fraction of initial committee members interviewed in baseline, and branch fixed effects. * p<0.1,** p<0.05, *** p<0.01.

APPENDIX 1 61

Tab

leA.2:B

aselinelead

ercharacteristicscompa

redto

regu

larmem

bers,c

ontrollin

gforindividu

alba

selin

ewealthscore(sam

ple:

allo

ldmem

bers)

Econo

mic

variab

les

Socioecon

omic

Social

conn

ection

proxies

LogAssets

Has

some

Empl

Migration

Has

Age

atMajority

Sharestrong

ShareCMs

2011

education

Network

expe

rien

cechild

ren

1stbirth

tribe

links

linkedto

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Lead

er0.113

0.077∗

0.123∗∗∗

0.036

0.136∗∗∗

-0.127

-0.024

0.044∗∗∗

0.032

[0.141]

[0.044]

[0.032]

[0.032]

[0.043]

[0.428]

[0.026]

[0.017]

[0.026]

Vote*Le

ader

-0.157

-0.046

-0.120∗∗

-0.152∗∗∗

0.069

-0.451

0.099∗∗

0.002

0.065

[0.196]

[0.065]

[0.047]

[0.050]

[0.058]

[0.489]

[0.039]

[0.019]

[0.040]

Wealthscore2011

0.008∗∗

0.010∗∗∗

0.001

0.005∗∗∗

-0.002

0.028∗∗

-0.002∗∗

0.001∗

-0.000

[0.004]

[0.001]

[0.001]

[0.001]

[0.001]

[0.011]

[0.001]

[0.000]

[0.001]

Lead

er+Vote*Le

ader

-0.044

0.031

0.002

-0.116∗∗∗

0.206∗∗∗

0.579∗∗

0.075∗

0.047∗∗∗

0.097∗∗∗

[0.134]

[0.049]

[0.032]

[0.039]

[0.038]

[0.239]

[0.029]

[0.010]

[0.031]

Discussionmean

13.485

0.487

0.279

0.283

0.531

18.557

0.542

0.135

0.109

Fixed

Effe

cts

grou

pgrou

pgrou

pgrou

pgrou

pgrou

pgrou

pgrou

pgrou

pObservation

s1323

1365

1449

1338

1379

824

1449

1008

1449

Adjusted

R2

0.268

0.250

0.330

0.232

0.230

0.086

0.697

0.184

0.213

Note:

Wealthscoreindicatesthepo

sition

(decile

)of

anindividu

alin

thewealthscoredistribu

tion

ofhergrou

p.The

scoreis

constructedusingtheUgand

aPPI

indexcompiledby

theGrameenfoun

dation

(2011).H

ighervalue=

less

poor.L

ogAssets20

11The

logged

valueof

assets

atba

selin

e,cond

itiona

lonha

ving

nonzero

assets.H

asno

educ.:du

mmy=

1ifindividu

alha

sno

scho

oling.

Employmentnetwork:

dummy=

1ifindividu

alha

din

thepa

styear

gotten

help

from

anyo

neou

tside

closefamily

formatters

ofem

ployment.Thisisaproxyforha

ving

labo

rmarketconn

ection

s.Migration

experience:du

mmy=

1ifindividu

alha

sever

lived

outside

thevilla

ge.Migration

isdo

neforshorttimeworkor

stud

iesan

dis

positively

correlated

withassets

andwealth.

Has

child

ren:du

mmy=

1iftherespon

dent

has

anychild

ren.

Age

at1stbirth:Indicatesindividu

al’s

agewhe

nshefirst

gave

birth.

Low

erageindicatesworse

socio-econ

omic

situation.

Onlyavailableforthose

who

have

child

renMajoritytribe:

dummy=

1iftheindividu

albe

long

sto

themostcommon

tribein

hersaving

grou

p,Sh

arefriend

sin

grou

p:ano

rmalized

degree

centralitymeasure

basedon

how

man

yothermem

bers

named

heram

ongtheir2be

stfriend

sin

thegrou

pat

baselin

e,no

rmalized

bynu

mbe

rof

othermem

bers

ingrou

pat

baselin

eSh

areCMslin

kedto:thenu

mbe

rof

individu

alsna

med

amon

gher4be

stfriend

sat

baselin

ewho

laterbe

camecommitteemem

bers,divided

bythenu

mbe

rof

committeemem

bers6=

therespon

dent.Discussionmean:Meanvalueam

ongno

n-lead

ersin

thediscussion

treatm

ent.

IGA

Con

trol:Con

trol

forthesharethat

hadreceived

incomegenerating

activity

training

from

BRAC,includ

edas

controlin

allregression

sto

accoun

tforim

balancein

this

variab

leacross

treatm

ents

atba

selin

e.Rob

uststan

dard

errors

inbrackets,clusteredat

thegrou

plevel(level

ofrand

omization).Regressions

areweigh

tedforthefraction

ofinitialcommitteemem

bers

interviewed

inba

selin

e,an

dinclud

ebran

chfix

edeff

ects.*p<

0.1,

**p<

0.05,***p<

0.01.


Table

A.3:D

ropoutby

2015related

toeconom

icvariables

atbaseline,allold

mem

bers,extensivemargin

Poor

(bottom25%

Poor

(bottom25%

Has

nomarket

Has

nosavings

Has

noloans

ofwealth

distr.)of

assetdistr.)

incomein

2011in

2011in

2011(1)

(2)(3)

(4)(5)

Characteristic

0.0120.064

0.082 ∗∗0.102

0.057[0.033]

[0.040][0.041]

[0.115][0.058]

Vote*characteristic

-0.076-0.082 ∗

-0.0230.005

-0.040[0.048]

[0.049][0.059]

[0.127][0.083]

Vote

-0.040-0.036

-0.049-0.071

-0.052[0.046]

[0.046][0.051]

[0.044][0.092]

Discussion

mean

0.7180.705

0.6860.714

0.663p-value

Discussion

=Vote

forChar=

10.045

0.0390.192

0.6000.044

p-valueChar0

=Char1

inVote

0.0740.492

0.1490.037

0.769Fixed

Effects

branchbranch

branchbranch

branchIG

AControl

yesyes

yesyes

yesObservations

1,4491,351

1,4101,396

1,310Adjusted

R2

0.2310.231

0.2300.238

0.236

Note:

The

dependentvariable

inallregressions

isan

indicatorof

havingdropped

outby

2015.The

independentvariables

aredum

mies

proxyingfor

povertyin

differentways.T

he2first

columns

indicatebeing

inthe

bottom25%

ofthe

wealth

distributionor

assetdistribution,respectively,of

one’sgroup.H

asno

market

incomeindicates

beingeither

asubsistence

farmer

oradependent

atbaseline.

Has

nosavings

indicatesnot

savingat

thetim

eof

thebaseline

surveyand

Has

noloans

indicatesnot

havingtaken

loansfrom

otherlenders

thanBRAC

(atthe

timeof

baseline,the

BRAC

groupshad

notstarted

givingout

loans).The

resultsare

robustto

reducingthe

sample

onlyto

individualswith

non-missing

valuesfor

theasset,

wealth

score,incom

e,savings

andloans

indicatorvariables.

Discussion

mean

:Share

ofdropouts

inthe

discussiontreatm

entfor

mem

berswith

independentvariable=

0.p-value

Discussion

=Vote

forChar=

1is

fromatest

ofwhether

dropoutrate

isthe

sameacross

treatments

forindividuals

with

povertycharacteristic=

1.p-value

Char0

=Char1

inVote

isfrom

atest

ofwhether

dropoutsare

similar

(interm

sof

eachpoverty

characteristic)to

stayersin

thevote

treatment.

IGA

Control:

theshare

thathad

receivedincom

egenerating

activitytraining

fromBRAC,is

includedas

acontrolin

allregressionsto

accountfor

imbalance

inthis

variableacross

treatments

atbaseline.R

obuststandard

errorsin

brackets,clusteredat

thegroup

level(levelofrandom


includebaseline

sample

weights,and

branchfixed

effects.*p<0.1,

**p<0.05,

***p<0.01.

APPENDIX 1 63

Table A.4: Loans anywhere by 2015, extensive margin, all old members

All Stayers Dropouts(1) (2) (3)

Vote -0.027 -0.062 -0.018[0.047] [0.061] [0.051]

log assets2011 0.001 -0.009 0.011[0.007] [0.007] [0.009]

Fixed Effects branch branch branchObservations 691 337 354Adjusted R2 0.118 0.127 0.098

Note: Dependent variable is a dummy for having taken aloan anywhere by the year 2015. Standard errors in brack-ets are clustered at group level. * p<0.1, ** p<0.05, ***p<0.01


Table

A.5:Loan

allocationam

ongallold

mem

bers,measured

in2015

Dep.var.:

Loandum

my

#Loans

givenTotallog

loanam

ountPoverty

proxy:Poor

Noincom

eNoloan

Poor

Noincom

eNoloan

Poor

Noincom

eNoloan

(1)(2)

(3)(4)

(5)(6)

(7)(8)

(9)

Characteristic*V

ote0.175 ∗∗

0.0330.180 ∗∗

0.347 ∗∗0.119

0.481 ∗∗1.667 ∗∗

0.4302.352 ∗∗

[0.069][0.083]

[0.087][0.147]

[0.204][0.204]

[0.694][0.984]

[1.067]Characteristic

-0.097 ∗0.016

-0.205 ∗∗∗-0.223 ∗∗

-0.008-0.507 ∗∗∗

-1.097 ∗∗0.148

-2.424 ∗∗[0.051]

[0.058][0.070]

[0.092][0.149]

[0.164][0.497]

[0.722][0.925]

Vote

-0.094 ∗-0.055

-0.198 ∗∗-0.167 ∗

-0.118-0.497 ∗∗

-0.915-0.572

-2.555 ∗∗[0.051]

[0.057][0.098]

[0.100][0.104]

[0.209][0.559]

[0.616][1.208]

Discussion

mean

0.4230.406

0.5220.766

0.7591.087

4.1783.887

5.434Fixed

effectsbranch

branchbranch

branchbranch

branchbranch

branchbranch

Observations

714701

639713

700638

650636

582Adjusted

R2

0.1190.121

0.1350.125

0.1320.145

0.1090.110

0.122

Note:

Loandum

my=

1ifthe

mem

berhas

evertaken

aloan

inthe

BRAC

IGFgroup

#Loans

given:N

umber

ofloans

evertaken

fromthe

IGFgroup

Total

loanam

ount:The

logam

ountof

loanstaken,

totalmargin.

Discussion

mean

:Mean

valueof

thedependent

variableam

ongmem

bersin

thediscussion

groupswith

povertyproxy=

0.Standard

errorsin

bracketsare

clusteredat

grouplevel.

*p<0.1,

**p<0.05,

***p<0.01

APPENDIX 1 65

Tab

leA.6:W

elfare

effects

Dep.v

ar.:

Has

loan

s2015

Has

saving

s2015

Savedam

t.2015

Poverty

proxy:

Poo

rNoincome

Noloan

sPoo

rNoincome

Noloan

sPoo

rNoincome

Noloan

s(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Cha

racteristic*Vote

0.153∗∗

0.070

0.137

0.008

-0.018

-0.002

-0.444

0.693∗∗

0.688∗

[0.074]

[0.085]

[0.098]

[0.082]

[0.110]

[0.089]

[0.284]

[0.318]

[0.384]

Cha

racteristic

-0.107∗∗

-0.032

-0.158∗∗

-0.078

-0.052

0.015

0.014

-0.370

-0.674∗∗∗

[0.053]

[0.056]

[0.067]

[0.061]

[0.070]

[0.071]

[0.213]

[0.241]

[0.253]

Vote

-0.077∗

-0.046

-0.150

-0.038

-0.027

0.001

0.037

-0.222

-0.692∗

[0.045]

[0.053]

[0.091]

[0.059]

[0.063]

[0.075]

[0.200]

[0.203]

[0.358]

Cha

racteristic+

[Cha

r.*V

ote]

0.046

0.039

-0.021

-0.069

-0.071

0.013

-0.431

0.322

0.014

[0.053]

[0.065]

[0.077]

[0.055]

[0.087]

[0.075]

[0.190]

[0.233]

[0.355]

Discussionmean

0.557

0.544

0.652

0.588

0.576

0.556

10.91

10.92

11.17

Fixed

Effe

cts

bran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chbran

chObservation

s714

701

639

699

686

627

327

319

286

Adjusted

R2

0.112

0.105

0.130

0.043

0.042

0.049

0.123

0.122

0.130

Note:

The

depe

ndentvariab

lein

columns

1-3is

anindicatorof

having

loan

s,withBRAC

oran

yotherlend

er,in

2015,while

incolumns

4-6thedepe

ndent

variab

leis

anindicatorof

having

saving

s,withBRAC

oran

ywhere

else,in

2015

andthedepe

ndentvariab

lein

columns

7-9is

thelogtotalUGX

valueof

all

saving

sheld

(any

whe

re)in

2015.The

indepe

ndentvariab

lesaredu

mmiesprox

ying

forpo

vertyin

diffe

rent

way

s.Poor(C

olum

ns1,

4an

d7)

indicatesbe

ingin

thebo

ttom

25%

ofthewealthdistribu

tion

ofon

e’sgrou

p.Noincome(colum

ns2,

5an

d8)

indicatesha

ving

nomarketincome,

i.e.be

ingeither

asubsistence

farm

eror

adepe

ndentat

baselin

e.Noloan

s(colum

ns3,

6an

d9)

indicatesno

tha

ving

takenloan

sfrom

otherlend

ersthan

BRAC

atba

selin

e(atthetimeof

baselin

e,theBRAC

grou

psha

dno

tstartedgiving

outloan

s).Discussionmean:Meanvalueof

therelevant

outcom

evariab

lein

thediscussion

treatm

entfor

mem

bers

withindepe

ndentvariab

le=0.

Regressions

includ

eendlinesampleweigh

ts,an

dbran

chfix

edeff

ects.*p<

0.1,

**p<

0.05,***p<

0.01.


Table A.7: Sampling and attrition rates, endline survey 2015

# in census 2015 # sampled Share sampled # responded response rate(Share sampled)

Stayers discussion 247 246 1.00 206 0.84Stayers vote 304 304 1.00 237 0.78Leavers discussion 657 270 0.41 225 0.83Leavers vote 603 248 0.41 228 0,92New discussion 518 221 0.43 208 0.94New vote 441 187 0.42 194 1.04

Total Observ. 2770 1476 0.53 1298 0.88Note: Data was collected using lists of sampled individuals, divided by their member type (stayer, dropout or new member).We sampled all stayers but only about 40% of dropouts and new members respectively. If a dropout (a new member) couldnot be interviewed, the data collectors were instructed to interview the first person in the same branch in a correspondingreplacement list for dropouts (new members). The replacement respondent was often a dropout (new member) in a nearbysaving group but not necessarily in the same group, and was not separated by treatment. This explains why share of newmembers in vote groups exceeds 1, new members in discussion groups that were not located have been replaced by newmembers in vote groups.

Table A.8: Group size 2015: stayers and new members active at census

All groups All groups Active >5 members(1) (2) (3) (4)

Vote treatm 0.994 0.994 1.184 2.550[1.104] [1.102] [1.185] [1.567]

Group size 2011 0.141 0.109 -0.0278[0.124] [0.134] [0.188]

Fixed Effects branch branch branch branchObservations 92 92 84 56Adjusted R2 0.511 0.512 0.452 0.260

Note: Dependent variable in all regressions is Group size at census 2015. Groupsize is computed as the sum of stayers and new members. Robust standard errorsin brackets. * p<0.1, ** p<0.05, *** p<0.01

APPENDIX 2: VARIABLE CONSTRUCTION DETAILS 67

Appendix 2: Variable construction details

Baseline variables

A few baseline variables central in the analysis require some additional explanation.

The wealth score is constructed using the Uganda PPI (Progress out of poverty) score

table (Grameen foundation, 2011, 2015). This is an index combining variables that that

measure the (inverse) poverty level of a household. These variables include measures of

the building materials of the main house of the household, the main source of lighting

used by the household, the education level of the young household members and of

its female household head and basic asset holdings (shoes and clothes of all household

members). The baseline survey did not contain questions about the full set of variables

required to compute the PPI score and we have therefore not converted the score into

likelihood of being below a certain poverty line, which is the ultimate purpose of the

PPI. Instead the score is used in its raw format to compare individual members’ wealth

status. We also construct the wealth score distribution in the group by decile, and use

the member’s position in this distribution as an outcome variable in some regressions.

Assets holdings of the household is constructed using answers to a roster listing a

number of household assets For each listed asset the member reported the quantity

of such assets owned by her household, and assessed their value. Asset value can in

principle be zero if the household owns none of the listed assets, but this is rarely

observed (4.32% of the members at baseline report a zero asset value) . Income of the

member is defined as income from market activities during the past 12 months and is

constructed using answers to a roster listing a number of possible income generating

activities. If a respondent reports having worked in an activity during the past 12

months, she is asked follow up questions including how much she earned in this activity

during the past 12 months. Income can be zero if a person only has non market or in

kind income (i.e. if she is a subsistence farmer/cattle keeper) or if she is not active

on the labor market. 23.33% of the members at baseline report a zero income. To

construct social network variables we make use of a question in the baseline survey


where the member is asked to list her four best friends in the group ("if she had to

choose 4". Using this information we construct two types of network variables for each

member: (i) variables measuring the relative degree centrality of the member, i.e. how

many other members mentioned her among their two or four best friends, normalized

by the number of members in the group interviewed at baseline and (ii) a variable

measuring how many of those who later became committee members that were listed

among a member’s four best friends.

Chapter 3

Preparing for Genocide: Community

Meetings in Rwanda∗

3.1 Introduction

In many civil wars and conflicts, ordinary seemingly unorganized civilians participate

in violence. For example, during the Rwandan Genocide in 1994, around 430,000 Hutu

civilians joined the army and militiamen in killing an estimated 800,000 Tutsis and

"moderate" Hutus in only 100 days.1 Civilian participation in violence often magni-

fies and escalates a given conflict with disastrous effects on the social fabric and the

economy, let alone the human suffering. Thus, it is crucial to understand the causes

of civilian participation in violence. Anecdotal evidence for the Rwandan case sug-

gests that in the years before the genocide, weekly-held community meetings called

Umuganda were used to sensitize and mobilize the civilian Hutu population. While

∗This paper is co-authored with Evelina Bonnier, Jonas Poulsen and Thorsten Rogall. We thankEli Berman, Tom Cunningham, Meilssa Dell, Jonas Hjort, Juanna Joensen, Magnus Johannesson, ErikLindqvist, Andreas Madestam, Eva Mörk, Suresh Naidu, Nathan Nunn, Torsten Persson, CristianPop-Eleches, Marit Rehavi, David Strömberg, Jakob Svensson, Erik Verhoogen and Miguel Urquiola,as well as seminar participants at Harvard, UCSD, IIES, SSE, NEUDC, the ASWEDE Conferenceon Development Economics, the Annual Bank Conference on Africa and Columbia DevelopmentColloquium for many helpful comments. Miri Stryjan is grateful for funding from Handelsbanken’sResearch Foundation

1In 1990, Rwanda had 7.1 million inhabitants out of which 6 million were Hutus.

69

70 PREPARING FOR GENOCIDE: COMMUNITY MEETINGS IN RWANDA

Umuganda was originally designed as mandatory work meetings to improve the village

infrastructure, earlier accounts of the genocide suggest that in the beginning of the

1990s, these meetings were abused by the political elites to spread anti-Tutsi senti-

ments and prepare the population for genocide (Cook, 2004; Straus, 2006; Thomson,

2009; Verwimp 2013).

This paper provides the first empirical analysis of how important local, elite led

community meetings might have been in inducing the civilian population to partici-

pate in violence. Despite the specific focus of this paper, we argue that examining the

possibly negative effect of these community meetings is of more general importance.

There is a widely held belief that community meetings foster social capital, by pro-

viding arenas for people to meet, exchange ideas, solve free-rider problems and create

public goods (Grootaert and van Bastelaer, 2002; Guiso, Sapienza and Zingalez, 2008;

Knack and Keefer, 1997; Putnam, 2000). Consistently, many important development

agencies today increasingly focus on ‘community driven’ development projects in which

deliberative forums and grass root participation play a central role (see Mansuri and

Rao, (2012) for a recent overview). We investigate whether there is a ‘dark side’ to

these community meetings where social capital does not bridge the societal, ethnic

divides but rather enforces bonding within the different ethnic groups, i.e. the Hutu

population in the Rwandan case. Understanding this process is even more important

since Umuganda was formally reintroduced in Rwanda in 2008, and similar practices

have been installed in Burundi and are discussed in the Democratic Republic of Congo

(DRC) and recently in Kenya.2 Identifying the causal effect of these meetings on par-

ticipation in genocide is difficult for two reasons. First, we lack data on the number

of people participating in Umuganda or the number of meetings taking place in a

given locality. Second, even if that data existed, our estimates would likely suffer from

omitted variable bias. On the one hand, village-specific unobservable characteristics

that affect both genocide violence and Umuganda intensity, for instance local leader

quality, could produce a spurious positive correlation between the two, thus biasing

2For details about the Kenyan case, see Daily Nation (March 2016).


the estimate upwards. On the other hand, if Umuganda meetings were strategically

used in areas where genocide participation would have been unobservably low, the

estimate would be downward biased.

To overcome these data and endogeneity issues, we use exogenous rainfall variation

to estimate the effect of Umuganda meetings on participation in civil conflict. The idea

is simple: we expect the meetings to be less enjoyable when it rains and furthermore

to be cancelled altogether under heavy rains. The fact that the community work only

took place on Saturdays makes it possible to isolate the Umuganda effect from general

rainfall effects (e.g. rainfall affecting income through agriculture) by only using the

variation in Saturday rainfall while controlling for average daily rainfall. We use the

number of Saturdays with heavy rainfall during the 3.5 year pre-genocide period (from

October 1990, the outbreak of the civil war, to March 1994, the eve of the genocide)

as our variable of interest. After the start of the civil war in October 1990, tensions

between Hutu and Tutsi intensified and the Hutu-dominated government became more

aggressive towards the Tutsi minority, finally culminating in the genocide. To control

for local characteristics, we include 142 commune fixed effects. Furthermore, we can

provide a first placebo check by controlling for heavy rainfall on all other six week-

days. Thus, we ensure that identification only stems from local variation in rainfall on

Saturdays, which is arguably exogenous and should only affect genocide participation

through its effect on Umuganda meeting intensity.

However, there is one major concern regarding the exclusion restriction. In partic-

ular, the effect we estimate might not be due to the political element of Umuganda

per se, but merely a consequence of people getting together in general. We will argue

in great detail why this concern is unwarranted. In particular, we will show that nei-

ther rainfall on Sundays, a church day where people traditionally meet, nor rainfall on

public holidays, affect participation. Moreover, the estimates are robust to excluding

Kigali, the Rwandan capital, where one might expect major outdoor events to take

place on weekends.

We proxy for genocide violence by the number of people prosecuted in the Gacaca


courts, normalized by sector Hutu population.3 About 10,000 local Gacaca courts were

set up all over the country to prosecute the crimes committed during the genocide.

Importantly, these courts distinguished between civilian perpetrators and perpetrators

belonging to an organized group such as militia gangs, the national army or local

police. Using prosecution instead of actual participation rates might introduce some

bias. However, the Gacaca data is strongly correlated with other measures of violence

from other various sources and we also present a number of additional tests to rule

our that systematic errors are biasing our results.

Our reduced-form results indicate a negative relationship between Umuganda inten-

sity and civilian participation in genocide: one additional rainy Saturday is associated

with a 5 percent decrease in the civilian participation rate. Interestingly, this negative

relationship is entirely driven by sectors that are ruled by the pro-genocide Hutu par-

ties. In places with pro-Tutsi parties in power the effects are reversed, suggesting that

meetings may have been used to create bonds between the two ethnicities.

All effects are similar although statistically weaker for organized participation. This

is not surprising since militia and army men would often not have been affected by

pre-genocide rainfall in the village where they were prosecuted (they moved around

during the genocide and did not necessarily commit their crimes in their hometowns).

Our results have important policy implications and are also relevant for other coun-

tries. In 2008, the Rwandan government reintroduced Umuganda. Our results show

that these meetings can easily be abused and that caution is warranted, in particu-

lar since tension between the Tutsi and the Hutu still exist in Rwanda. Furthermore,

similar practices have been implemented in Burundi and are being discussed in the

Democratic Republic of Congo (DRC) and Kenya. These countries all have histories of

violent conflict along ethnic lines, which once more calls for caution when establishing

an institution such as mandatory community meetings.

Our paper contributes to the literature in several ways. First of all, it adds to the

vast conflict literature. Blattman and Miguel (2010) give an excellent review of this3A sector corresponds to the Rwandan administrative unit of a sector with an average size of 14

square kilometers and 4,900 inhabitants.


literature, vehemently calling for well-identified studies on the roots of individual par-

ticipation in violent conflict. This paper adds to the conflict literature by providing

novel evidence on the strong effects of local community meetings, controlled by the

political elite, on civilian participation in violence. Recent studies on the determinants

of conflict and participation in violence consider institutions, government policy, in-

come and foreign aid (Besley and Persson, 2011; Dell, 2012; Dube and Vargas, 2013;

Mitra and Ray, 2014; Nunn and Qian, 2014, respectively). Furthermore, our paper

complements the literature on the Rwandan Genocide (Rogall, 2014; Straus, 2004;

Verpoorten, 2012a-c; Verwimp, 2003, 2005, 2006; Yanagizawa-Drott, 2014) by provid-

ing novel evidence on its careful preparation.

On the methodology side, our findings speak to the recent discussion on the effects

of rainfall on conflict other than through the income channel (Iyer and Topalova, 2014;

Rogall, 2014; Sarsons, 2011). Prominent studies that use rainfall as an instrument for

income in Africa include Brückner and Ciccone (2010), Chaney (2013) and Miguel,

Satyanath and Sergenti (2004). Our results suggest that rainfall might have negative

direct effects on conflict.

Finally, our results are in line with Satyanath, Voigtlaender and Voth (2015) who

speak for a "dark side" of social capital, in contrast to several contributions high-

lighting its positive effects (Grootaert and van Bastelaer, 2002; Guiso, Sapienza and

Zingalez, 2008; Knack and Keefer, 1997).

The remainder of the paper is organized as follows. Section 2 provides some back-

ground information on the Rwandan genocide. Section 3 presents the data used for

the analysis and Section 4 lays out our empirical strategy. Section 5 presents the main

results and assesses their robustness and Section 6 discusses mechanisms and channels.

Section 7 concludes with possible policy implications.


3.2 Background

A History of Conflict A long chain of events lead up to the 1994 Rwandan geno-

cide.4 German and Belgian colonizers promoted the Tutsi minority and gave them

supremacy over the Hutu majority.5 This division created a strong tension between

the two groups and culminated in the Rwandan revolution, or the Social revolution of

1959, where the Tutsi monarchy was dissolved in favor of a republic led by the Hutus.

Many Tutsi civilians were killed; others fled Rwanda for neighboring countries such as

Burundi, Tanzania and, in particular, Uganda. During the following decade, the coun-

try faced several attacks from exiled Tutsi rebel groups with following Hutu retaliation.

In 1974 - paramount to the introduction of a modern version of Umuganda - Juvénal

Habyarimana took power in Rwanda through a coup d’état. His subsequent rule was

based on a pro-Hutu ideology ("Hutu power"), further discussed in the next section. In

October 1990, the RPF - a rebel army mostly composed of Tutsi exiles eager to replace

the Hutu-led government - invaded Rwanda from Uganda, starting the Rwandan civil

war. Fighting between the Hutu-led government and the Tutsi rebels continued until

the Arusha Accords were signed in August 1993.6 While a multi-party system had

been installed in the early phase of the peace talks, this had little (or no) effect on

reducing societal tension and conflict. On April 6 1994, the airplane carrying president

Habyarimana was shot down. Responsibility for the attack is still disputed today, but

within only a few hours of the attack, extremists within the Hutu-dominated parties

managed to take over key positions of government and initiated a 100-day period of

ethnic cleansing throughout Rwanda. Estimates suggest that around 800,000 people,

mostly Tutsi and "moderate" Hutus (i.e. Hutus believed to side with the Tutsi), were

killed. The mass killings ended in mid-July, when the RPF rebels, who in the meantime

4For insightful accounts of this period, see for example Prunier (1995), Gouveritch (1998), DesForges (1999), Dallaire (2003), Hatzfeld (2005, 2006) and Straus (2006).

5In the 1991 census data used in this paper, the average reported share of Hutus per commune is87%.

6The essence of this treaty was a power-sharing government, including representatives from bothsides of the conflict.

3.2. BACKGROUND 75

renewed the civil war, defeated the Rwandan Hutu army and the various militia groups.

A large number of Hutu civilians participated in the genocide violence, directed

by the interim government (Dallaire, 2003). In our sample, there are approximately

416,000 civilian perpetrators.7

Umuganda The practice of Umuganda dates back to pre-colonial times. During a

day of community service, villagers would get together to build houses for those not

able to do this themselves, or help each other out on the fields in times of economic

hardship (Mukarubuga, 2006). Rather than being mandatory, Umuganda was initially

considered a social obligation (Melvern, 2000). This changed during the the colonial

period, when the Belgian colonizers used Umuganda for organizing compulsory work.

Consistently, the local term for Umuganda was then uburetwa, or forced labor (IRDP,

2003). All men had to provide communal work 60 days per year. Most of the manual

labor was hereby carried out by members of the ethnic Hutu majority under the

supervision of Tutsi chiefs (Pottier, 2006): a first sign of Umuganda’s potential to

create a division between the two ethnic groups.

During the post-colonial era from 1974 onwards, the meaning of Umuganda changed

again when the newly elected Hutu president Habyarimana turned it into a political

doctrine (Mamdani, 2001). Verwimp (2000, p. 344) cites Habyarimana:

"The doctrine of our movement [Movement for Development, MRND] is

that Rwanda will only be developed by the sum of the efforts of its people.

That is why it has judged the collective work for development a necessary

obligation for all inhabitants of the country."

The program combined a practical motivation - achieving development objectives

with weak state finances - with a strong ideological element. Participation was again

compulsory through government coercion and failure to participate usually involved

paying a fine.8 The local leaders of the neighborhood who presided over a group of7For more information, see for example Dallaire (2003), Des Forges (1999), Gouveritch (1998),

Hatzfeld (2005, 2006), Prunier (1995) and Straus (2006).8In today’s Rwanda, the fine for not participating in Umuganda is slightly less than $10.


ten households were responsible for the weekly Umugandas and had the power to

decide who were to participate and to demand fines from those failing to do so (Ver-

wimp, 2000). The state chose the projects on which at least one adult male per family

had to work every Saturday morning (Uvin, 1998). According to a report from 1986:

56 percent of the work performed during Umuganda included various types of anti-

erosion measures, such as terracing and digging ditches; 15 percent were construction

of communal buildings; 21 percent consisted of maintenance work of communal roads;

3 percent were related to construction of water supply systems, and another 3 per-

cent were related to agriculture. In this period, Umuganda substantially contributed

to Rwanda’s GDP (Guichaoua, 1991).

Habyarimana’s ideology stressed the importance of the cultivator as the true Rwan-

dan (Straus, 2006). This view clearly embraced the Hutu population with their history

as cultivators, as opposed to the Tutsi who were said to be pastoralists. During the

period leading up to the genocide, Umuganda was used to strengthen group cohe-

sion within the "indigenous" ba-Hutu and marginalize the "non-indigenous" ba-Tutsi

(Lawrence and Uwimbabazi, 2013). The patriotic focus of Umuganda became par-

ticularly salient in the early 1990’s when "government propaganda gave no choice to

Rwandans other than to attend Umuganda for political mobilization" (Lawrence and

Uwimbabazi 2013, p. 253). Furthermore, " (...) those who could not attend were re-

garded as enemies of the country who ran the risk of being brutalised and killed."

(ibid.).

Although little is known about the link between participation in Umuganda before

the genocide and participation in violence during the genocide - a link which we hope

to shed some new light on in this paper - anecdotal evidence speaks to the importance

of Umuganda as an instrument for local party and state officials to mobilize the peasant

population. The fact that all Rwandans of working age, be it farmers or intellectuals,

were required to participate in Umuganda (Guichaoua, 1991) made it a potential

arena for reaching the entire population. Straus (2006) shows that 88 percent of the

perpetrators he interviewed regularly participated in Umuganda before the genocide

3.3. DATA 77

broke out.9 Verwimp (2013, p. 40) notes:

"Umuganda gave the local party and state officials knowledge and experi-

ence in the mobilization and control of the labor of the peasant population.

A skill that [would] prove deadly during the genocide."

Umuganda was also used during the genocide itself, with the new name gukorn

akazi, or "do the work", which meant the killing of Tutsis (Verwimp, 2013). Other

slogans related to Umuganda used before the genocide such as "clearing bushes and

removing bad weeds" now had a completely altered connotation (Lawrence and Uwim-

babazi, 2013). By equating the participation in genocide violence with participation

in Umuganda, the Hutu elite could signal that participation in genocide violence, just

like participation in Umuganda, was a social obligation for all ’true’ Rwandans.

In 2008, the Tutsi-led government re-introduced Umuganda in Rwanda with the

general aim to promote development and reduce poverty in the aftermath of the geno-

cide (Uwimbabazi, 2012). Participation is again mandatory for all able-bodied individ-

uals between 18 and 65 years of age, and typical tasks include cleaning streets, cutting

grass and trimming bushes along roads, repairing public facilities or building houses

for vulnerable individuals on the last Saturday of every month.

3.3 Data

We combine several datasets from different sources to construct our final dataset, which

comprises 1433 Rwandan sectors. Sectors are the second smallest administrative level,

and the level for which the outcome data on the perpetrators is available. Table 3.1

reports the summary statistics for our variables.

Participation Rates Our two key measures are participation in civilian violence

and organized violence. Ideally, we would like to have a direct measure of participation

in the genocide. Since no such data exists, we follow the literature and use prosecution9These findings are, however, based on correlation studies and it is not possible to claim causality.


rates for crimes committed during the genocide as a proxy (Friedman, 2013; Heldring,

2014; Rogall, 2014; Yanagizawa-Drott, 2014). This data are taken from a nation-wide

sector-level dataset, provided by the government agency "National Service of Gacaca

Jurisdiction", which records the outcome of the almost 10,000 Gacaca courts set up

all over the country. The released sector level data classifies the accused into one of

two categories, depending on the role played by the accused and the severity of the

crime.

The first category which we refer to as "organized participants" concerns: (i) plan-

ners, organizers, instigators, supervisors of the genocide; (ii) leaders at the national,

provincial or district level, within political parties, army, religious denominations or

militia; (iii) the well-known murderer who distinguished himself because of the zeal

which characterized him in the killings or the excessive wickedness with which killings

were carried out; (iv) people who committed rape or acts of sexual torture. These per-

petrators mostly belonged to the army and militia or were local leaders. Approximately

77,000 people were prosecuted in this category.10

The second category which we refer to as "civilian participants" concerns: (i) au-

thors, co-authors, accomplices of deliberate homicides, or of serious attacks that caused

someone’s death; (ii) the person who - with the intention of killing - caused injuries or

committed other serious violence, but without actually causing death; (iii) the person

who committed criminal acts or became the accomplice of serious attacks, without

the intention of causing death. People accused in this category are not members of

any of the organized groups mentioned for the first category and are thus considered

to be civilians. Approximately 430,000 people were prosecuted in this category. As

mentioned, the second category is our main outcome variable since civilian participa-

tion in the killings is more likely to have been affected by Umuganda than organized

participation.

The reliability of the prosecution data is a key issue for the analysis. One concern

when using prosecution data instead of actual participation is the presence of survival10Since we lose some observations for category 1 and category 2 in the matching process, our sample

consists of 415,935 category 2 perpetrators and 74,168 category 1 perpetrators.

3.3. DATA 79

bias: in those sectors with high participation rates, the violence might have been

so widespread that no witnesses were left, or the few remaining were too scared to

identify and accuse the perpetrators, thus resulting in low prosecution rates. However,

this concern is unlikely to be warranted: Friedman (2013) shows that the Gacaca data

is positively correlated with several other measures of violence from three different

sources.11 Furthermore, Friedman (2013, pp. 19-20) notes that "the Gacaca courts

have been very thorough in investigating, and reports of those afraid to speak are rare,

so this data is likely to be a good proxy for the number of participants in each area."

Nevertheless, to be cautious, in the following analysis we will show that our results are

robust to dropping those sectors with mass graves (an indication of high death rates)

and also to an alternative specification using the presence of a mass grave directly as

a dependent variable proxying for violence.

Another concern is that some of those people prosecuted in the Gacaca courts might

have committed their crimes not during the genocide, but rather during the period

of civil war preceding the genocide (October 1990 until August 1993). In particular,

we cannot rule out that (a) some perpetrators may, in fact, have been accused of

participation in massacres and other violence during the civil war (and not during

the genocide) and (b) that individuals who had previously participated in violence

during the civil war were more likely to have been recognized and trialed for genocide

crimes than individuals who "only" participated in the genocide. In order to mitigate

this concern, we exclude communes with violence against the Tutsi during the period

from October 1990 to March 1994 (Viret, 2010). Importantly, violence directed against

Hutus was not trialed in the Gacaca courts (Human Rights Watch, 2011; Longman,

2009).

Rainfall Data We use the National Oceanic and Atmospheric Administration (NOAA)

database of daily rainfall estimates, which stretches back to 1983, as a source of ex-

11These sources are a 1996 report from the Ministry of Higher Education, Scientific Research, andCulture (Kapiteni, 1996); the PRIO/Uppsala data on violent conflicts (Gleditsch et al, 2002); and adatabase of timing and lethality of conflict from Davenport and Stam (2009).


ogenous weather variation. The NOAA data relies on a combination of actual weather

station gauge measures and satellite information on the density of cloud cover to de-

rive rainfall estimates at 0.1-degree (∼ 11 kilometers at the equator) latitude-longitude

intervals. Considering the small size of Rwanda, this high spatial resolution data, to

our knowledge the only such data available, is crucial for obtaining reasonable rain-

fall variation. Furthermore, the high temporal resolution, i.e. daily estimates, allows

us to confine variation in rainfall to the exact days of Umuganda. Since Rwanda is

a very hilly country, there is considerable local variation in rainfall. Moreover, these

sectors criss cross the various rainfall grids and each sector polygon is likely to overlap

with more than one rainfall grid. The overall rainfall in each sector is thus obtained

through a weighted average of the grids, where the weights are given by the relative

areas covered by each grid.

Village Boundary, Road and City Data The Center for Geographic Information

Systems and Remote Sensing of the National University of Rwanda (CGIS-NUR)

in Butare provides a sector boundary map, importantly with additional information

on both recent and old administrative groupings. Since Rwandan sectors have been

regrouped under different higher administrative units a number of times after the

genocide, this information allows us to match sectors across different datasets (e.g. the

1991 census and the Gacaca records).

Africover provides maps with the location of major roads and cities derived from

satellite imagery. We use these maps to calculate the sector area, as well as to calculate

various distance measures, such as the distance of the sector centroid to the closest

main road, to the closest city, to the borders of the country and to Kigali and Nyanza,

the recent capital and the old Tutsi Kingdom capital, respectively.

Additional Data The remaining data is drawn from Genodynamics and the IPUMS

International census data base. This data includes population, ethnicity and radio

3.4. EMPIRICAL STRATEGY 81

ownership from 1991.12 Except for population, all these variables are only available

at the commune level. Ethnicity is defined as the fraction of people that are Hutu

or Tutsi, respectively. About 10 percent of the population are Tutsi. Importantly, the

Tutsi minority is spread out across the entire country. We calculate the Tutsi minority

share used in the analysis as the fraction of Tutsi normalized by the fraction of Hutu.

Verpoorten (2012c) provides data on the location of mass graves which she con-

structs using satellite maps from the Yale Genocide Studies Program. Guichaoua

(1991) provides information on the party affiliation of the commune leaders (called

burgomasters) at the eve of the genocide.

Matching of data and summary statistics The different data sets are matched

by sector names within communes. A commune is an administrative unit above the

sector. There were 142 communes in total, which were, in turn, grouped into 11

provinces. Unfortunately, the matching is imperfect, since some sectors either have

different names in different data sources, or use alternate spelling. It is not uncommon

for two or more sectors within a commune to have identical names, and this prevents

successful matching. However, overall, only about 5 percent of the sectors do not have a

clear match across all sources. Furthermore, as these issues are idiosyncratic, the main

implication for our analysis is lower precision in the estimates than would otherwise

have been the case.

3.4 Empirical Strategy

To identify the effect of Umuganda meetings on participation in genocide violence,

we use local variation in rainfall as a proxy. Since we lack data on the number of

people participating in Umuganda, we focus on the reduced form effect. Our identifi-

cation strategy thus rests on two assumptions. First, sectors with heavier rainfall on

12This data is only available for 1991. However, mobility was extremely limited because of gov-ernmental restrictions and land markets were also strongly controlled (Andre and Platteau, 1998;Prunier, 1995).


Saturdays experienced fewer or less intensive Umuganda meetings (first stage). Sec-

ond, conditional on our control variables, rainfall on Saturdays does not have a direct

effect on genocide violence other than through the Umuganda meetings (exclusion

restriction).

First Stage Ideally, we would like to directly test the first-stage relationship using

data on the number of people participating in Umuganda before the genocide. Since

such data does not exist, we will instead provide indirect evidence for expecting a

strong first stage.

Several other studies have documented and exploited negative relationships be-

tween rainfall and participation in open-air events. One of the first examples is Collins

and Margo (2007) who use rainfall in April 1968 as an instrument for participation

in the US riots after the death of Martin Luther King. More recent examples include

Madestam et al. (2013) and Madestam and Yanagizawa-Drott (2011). Similarly, sev-

eral other studies use rainfall and other weather phenomena for exogenous variation

in voter turnout on election days (Eisinga et al., 2012; Fraga and Hersh, 2011; Gomez

et al., 2012; Hansford and Gomez, 2010; Horiuchi and Saito, 2009).

However, in all these cases, rain has an effect both on the direct cost of attending

the open-air event and the opportunity cost of attending. For example, Lind (2014)

finds that voter turnout in Norway increases when it rains on election day because bad

weather reduces the opportunity cost of going to the polling station. Since Umuganda

was mandatory, the opportunity cost mechanism is however unlikely to play any role

in our case, however. Instead, rainfall likely made the meetings and the work less

productive, or even lead to cancellations. Still, the true functional form between rainfall

and participation in mandatory community work is unknown. To make progress, we

reasonably assume that the typical Umuganda tasks, exclusively outdoor work, become

difficult or impossible to perform once a certain rainfall threshold has been reached.13

Following Harari and La Ferrara (2013) who define an extreme weather shock as

13The typical Umuganda tasks took place outside and, as mentioned above, included landscaping,road maintenance, construction, and agriculture (Guichaoua, 1991).

3.4. EMPIRICAL STRATEGY 83

two standard deviations from the long-term average, we choose this threshold to be

10mm.14 Thus, we will use the number of Saturdays from October 1990 to March

1994, that each sector received more than 10 mm of rainfall as a our main explanatory

variable.15 Furthermore, in the appendix, Table A.1, we show that our results are also

robust to using average daily rainfall on Saturdays and all other weekdays as our main

explanatory variables. Figure 3.3 shows the distribution of Saturday rainfall across

Rwanda in our period of interest.

Furthermore, to better understand whether rainfall affected the extensive or the

intensive margin of Umuganda meetings, we can vary these thresholds. More specif-

ically, we will also use thresholds of 6 mm, 8 mm, 9 mm and 12 mm, respectively.16

If we see effects already at low thresholds, it speaks to less enjoyable meetings or an

effect at the intensive margin. If the effects begin only at higher levels, cancellations

are more likely to be driving the results, i.e. affecting the extensive margin. Average

daily rainfall in Rwanda is low, however (see Table 3.1), which means that for very

high thresholds, the variation will be too small to detect any effects.

Exclusion Restriction Once more, our empirical strategy relies on the counterfac-

tual assumption that, absent the Umuganda meetings, rainfall on Saturdays had no

effect on genocide violence. This is unlikely to be true without further precautions.

Rainfall on Saturdays, like all other weekdays, is likely to affect rain-fed production

and is therefore correlated with income. Income, in return, potentially affects genocide

participation as the reasons for participating were often driven by material incentives

14The long-term average daily rainfall in Rwanda from 1984 to 1994 was 2.6 mm with a standarddeviation of 3.8 mm. We calculate this number taking the average across all sectors and all days from1984 to 1994. Two standard deviations from the long-term average corresponds to 10.24 mm.

15Madestam et al. (2013) use a threshold of 0.1 inches (2.5 millimeter) of rainfall, a light drizzle,to predict participation in the Tea Party Tax Day rally in the US. While a 2.5 mm threshold maybe appropriate to capture participation in a voluntary rally in the US, we believe that our case, i.e.mandatory meetings, requires a higher threshold. Madestam et al. (2013) also use 0.35 inches (≈ 9mm) as a robustness check for a higher threshold of rainfall. In Table 3.3, we show that our resultsare robust also to using this threshold.

16The 8 mm and 12 mm thresholds correspond to the average of the 95th and the 99th percentileof daily rainfall in Rwanda over the period from 1984 to 1994. Here, we follow Dyson (2009) who,in order to understand the characteristics of rainfall in South Africa, defines heavy and very heavyrainfall as the average of the 95th and 99th percentile of daily rainfall, respectively.


and genocide perpetrators were given the opportunity to loot the property of the vic-

tims, or could bribe themselves out of participating (Hatzfeld, 2005). Besides affecting

agricultural outcomes, heavy rainfall might destroy infrastructure such as roads or

housing, which is also likely to affect households’ economic well-being and therefore

participation in conflict.

To address this problem, and to solely isolate the Saturday rainfall effect, we control

for average daily rainfall from January 1984 to September 1990 and our period of

interest from October 1990 to March 1994. Furthermore, we control for rainfall on all

other six weekdays. The absence of systematic, significant effects for days other than

Saturdays serves as a first placebo test. To account for local characteristics, we also

add 142 commune fixed effects.

At this point, we still need to argue that no other events potentially happening

parallel to Umuganda on Saturdays could be driving our results. In particular, one

might be concerned that people meeting and interacting in general might affect the

participation in genocide violence. Although we cannot directly test for this, we will

provide several indirect tests alleviating this concern.

Specifications We run the following reduced-form regression to estimate the effect

of Umuganda meetings on participation in genocide violence

Gic

Hic= α +β #Saturdays(Rain f all > t mm)+Xicπ + γc + εic (3.1)

where Gic is the number of Hutu prosecuted in either category 1 or category 2, i.e.

our proxy for genocide violence and Hic is the Hutu population in sector i in commune

c. #Saturdays(Rain f all > t mm) is our explanatory variable of interest: the number

of Saturdays from October 1990 to March 1994 with rainfall above t mm. Our main

specification uses 10 mm as a measure of heavy rainfall, but our results are robust

to using other rainfall thresholds. Xic is a vector of sector-specific controls, including

average daily rainfall from January 1984 to September 1990, average daily rainfall

from October 1990 to March 1994 and the number of all other weekdays with rainfall

3.5. RESULTS 85

above t mm during our period of interest, October 1990 to March 1994. Finally, γc are

commune fixed effects, and εic is the error term. We allow error terms to be correlated

across sectors within the same commune by clustering the standard errors at the

commune level. For the sake of robustness, we also allow error terms to be correlated

across sectors within a 25, 50 and 75 km radius (Conley, 1999).17 Moreover, since

the prosecution rates are heavily skewed to the right, we weight our observations by

total sector population size, but our results do not rely on this weighting scheme. The

coefficient of interest β captures the percentage point change in genocide participation

following an additional Saturday with rainfall above t mm.

3.5 Results

Main Effects The reduced-form relationship between the number of civilian perpe-

trators per Hutu and the number of Saturdays with rainfall above 10 mm is strongly

negative and statistically significant at the 99 percent confidence level (column 1 in Ta-

ble 3.2) and this relationship holds up when adding 142 commune fixed effects (column

2) and the number of other weekdays with rainfall above 10 mm (column 3). Regarding

magnitude, the point estimate of 0.409 (column 3 with all controls) suggests that one

additional rainy Saturday reduces the civilian participation rate by 0.409 percentage

points (note that the civilian participation rate is measured in percent). If we assume

a one-to-one relationship between the number of rainy Saturdays and the number of

cancelled Umuganda meetings, an additional cancelled meeting reduces the average

civilian participation rate by 5.4 percent (interpreted at the mean of civilian perpe-

trators per Hutu, which is 7.7 percent). Reassuringly, none of the other weekdays is

systematically and significantly related to civilian violence (we cannot reject the null

that all coefficients are equal to zero, p-value 0.937).

The results for organized perpetrators are statistically weaker, they are significant

at the 90 percent confidence level (columns 4 to 6). This is not surprising: since the

17The results are reported in Table A.2 in the appendix.


category of organized perpetrators mostly consists of members of the militia, it is

unclear whether the sector where they committed their genocide crimes (and were

subsequently prosecuted) is the same as the one where they lived before the genocide

(October 1990 to March 1994). Thus, they will not have been exposed to the same

number of Umugandas as the inhabitants of that sector. If this is the case, our data

is likely to suffer from measurement error increasing standard errors. Since the main

focus in this paper is to examine if Umuganda can explain civilian violence, we ex-

clude organized violence from our main analysis. However, we report the corresponding

results for organized perpetrators in the appendix (Tables A.3 to A.6).

To understand whether rainfall led to cancellations, or rather made the Umuganda

meetings less enjoyable, we vary the threshold in increments of 2 mm: from 6 mm to 12

mm.18 Table 3.3 reports the results. Heavy rainfall on Saturdays is negatively related

to civilian participation for all thresholds and significant at least at the 90 percent

confidence level for all thresholds above 6 mm. Importantly, we find the strongest

effects for thresholds above 9mm, suggesting that it was rather cancellations that led to

a decrease in violence. Once more, we find no significant effects for other weekdays and,

consistently, we cannot reject the null hypothesis that the non-Saturday coefficients

are jointly equal to zero (the p-values range from 0.34 to 0.97).

Robustness Checks Next, we perform a number of robustness checks and placebo

tests, reported in Table 3.4. The potential survival bias in the prosecution data is un-

likely to matter: the reduced form point estimates are virtually identical to the baseline

results and similarly significant at the 99 percent confidence level when dropping sec-

tors with at least one mass grave (indicating high death rates, column 1). Furthermore,

we can also use the presence of a mass grave directly as a dependent variable. Consis-

tently, columns 7 and 8 show that sectors with many rainy Saturdays are less likely to

have a mass grave site altogether. The point estimate of -0.013 (standard error 0.004,

column 8), significant again at the 99 percent level, suggests that a sector is 26 percent

18To be consistent with Madestam et al. (2013), we also use 0.35 inches (which corresponds to 9mm) as a threshold.

3.5. RESULTS 87

less likely to have a mass grave site, given an additional rainy Saturday.

One might also be concerned that the UN troops which were stationed in Kigali,

although few, affected the Umuganda meetings, thus biasing our estimates. But once

more, the results are robust to dropping sectors in Kigali city (column 2). Furthermore,

the results are robust to dropping all main cities and sectors close to them (column

3).

The results are also unaffected by adding a number of additional controls that

potentially affect civilian participation in violence (column 4). These include distance

to the border, distance to major cities, distance to Kigali and distance to Nyanza

as well as population density. To illustrate this, being close to the border potentially

made it easier for the Tutsi or for those Hutu unwilling to participate in the killings

to leave the country. Distance to cities, in particular the capital Kigali, is likely to be

correlated with urbanization and public goods provision (economic activity). Nyanza

was the old Tutsi Kingdom capital and sectors further away from it do still, on average,

exhibit lower Tutsi shares. Population density eventually captures social pressure as

well as food pressure, both said to be important reasons for the genocide (Boudreaux

2009; Diamond, 2005; Verpoorten, 2012b).19

Finally, as a placebo check, we re-estimate the reduced-form regressions, instead

using the number of Saturdays (and other weekdays) with high rainfall during the

period October 1994 to March 1998 instead (from here on denoted post-genocide

period). To account for possible seasonality in the rainfall data, we chose the same

calendar period as our period of interest, i.e. October 1, 1994 to March 31, 1998.

Reassuringly, the coefficient for high rainfall on Saturdays in the post-genocide period

is small and insignificant (-0.012, standard error 0.106, column 5) and the same is true

for the coefficients on all other weekdays of the post-genocide period (except Monday).

These small and insignificant point estimates furthermore remain unchanged when

adding rainfall by weekday during our period of interest, October 1990 to March 1994

19The food pressure argument essentially assumes a Malthusian type of model: a fixed amount ofland to grow crops feeds a growing population (fertilizers were seldom used in Rwanda (Percival andHomer-Dixon, 2001)).


(column 6).

As another placebo check, we rerun the main specification for both militia and civil-

ian violence using Saturday rainfall during the 3.5 year pre-genocide period (October

1, YEAR to March 31, YEAR+4) where YEAR ∈ {1983,2013}. To illustrate this, we

begin with the period from October 1, 1984 until March 31, 1988 and end with the

period from October 1, 2009 until March 31, 2013. As expected, the two distributions

of the resulting 20 coefficients are both somewhat centered around 0 and, reassuringly,

the coefficient on Saturday rainfall from 1990 to 1994, the actual pre-genocide period,

is an extreme outlier to the left in both cases: None of the other point estimates is

larger in absolute value (the results are shown in Figures 3.1 and 3.2).

Exclusion Restriction After demonstrating a strong and robust effect of high Sat-

urday rainfall on civilian participation in genocide, we still have to argue that this

effect results from people participating in Umuganda together.

Most importantly, since major outdoor events, such as music festivals or soccer

games, usually take place on weekends, potentially affected by rainfall, one might be

concerned that people meeting and interacting in general could affect the participa-

tion in genocide violence. However, recalling our main result in Table 3.2, we find

no significant effect for Sunday rainfall. Since people traditionally attend church on

Sundays, this is the first piece of evidence speaking against the effects being driven by

people meeting in general. Besides, as seen above, our results are robust to dropping

the capital Kigali and other major cities in the sample; places where one might expect

these major outdoor events to predominantly take place.

In a similar vein, heavy rainfall on public holidays, another occasion for people to

meet, does not seem to matter: the point estimate on the number of public holidays

with rainfall above 10mm is statistically insignificant and small, when expressed in

standard deviations (column 1 in Table 3.5).20 The same is true when adding religious

20Note that we exclude holidays that fall on a Saturday since these might still have been subjectto Umuganda.

3.6. CHANNELS 89

and non-religious holidays separately to the regression (column 2).21

Throughout our period of interest from 1990 to 1994, violent acts against Tutsi

and "moderate" Hutu were already taking place. If these pre-genocide perpetrators

are included in the Gacaca data, and there is a relationship between rainfall before

the genocide and targeted violence during that period, for instance through transport

costs, our estimates might be biased. To rule out this possibility, we drop communes

where violence against the Tutsi took place before the genocide (Viret, 2010). Reas-

suringly, our results for civilian participation are robust (column 3).

To provide further evidence that the effects we measure above result from the

political elites abusing Umuganda meetings, we split the sample of sectors into those

located in communes with local pro-genocide Hutu party leaders and those located in

communes with pro-Tutsi party leaders22. Interestingly, the negative relationship from

above seems to be entirely driven by the pro-genocide Hutu-governed sectors. The

point estimate on Saturday rainfall is -0.466 (standard error 0.123, column 5), slightly

larger than our main effect and again highly significant at the 99 percent confidence

level. The opposite is true in pro-Tutsi sectors: the point estimate on Saturday rainfall

is large and positive, albeit given the small sample of only 161 sectors, it is insignificant

(0.706, standard error 0.896, in column 6 and 0.399, standard error 0.796, in column

7 with all other weekday controls). The numbers suggest that in these sectors, the

meetings were used to create bonds between the two ethnicities.

3.6 Channels

In the following section, we try to better understand the channels and mechanisms

through which Umuganda worked. Since the mechanisms in Hutu-governed sectors

and pro-Tutsi-governed sectors are likely to differ, we always analyze the two sub-

samples separately. All results are reported in Table 3.6.

21Religious holidays are, for instance, Easter and Christmas, non-religious holidays in Rwanda are,for instance, Independence Day and Labor Day.

22This data is available from the introduction of a multiparty system in January 1992.


Interaction Effects Starting with the Hutu-run sectors, a natural first question is

whether the political Hutu elites mostly spread propaganda and informed civilians

about the views of the government - something a radio broadcast might have done

just as well - or whether the local elites rather brought civilians together, practicing

mobilization, something that certainly would have required physical presence in the

locality. Importantly, there were two radio stations in Rwanda (Radio Rwanda and

Radio RTLM, the former having national coverage), which relentlessly informed lis-

teners about the pro-genocide view of the government. This hints at a way of testing

the initial question: if the Umuganda meetings mostly worked through information,

then the effect of the Umuganda meetings, i.e. Saturday rainfall, should be smaller

(in absolute values) in sectors that were already informed, i.e. exhibited high levels of

radio ownership. Thus, we should observe a positive interaction effect of Saturday rain-

fall with radio ownership among the Hutu population in the data. The point estimate

on the interaction term is positive (0.659, column 1); however with a standard error

of 0.786, it is clearly insignificant. Furthermore, when we replace the radio ownership

variable by a dummy taking on the value of 1 if radio ownership lies above the median,

the interaction effect is essentially zero (the result is not shown). Thus, it seems to be

the case, that Umuganda worked beyond information and propaganda.

Rather, consistent with the local elites using Umuganda to bring people together,

the interaction effect of Saturday rainfall with population density is positive and highly

significant at the 99 percent confidence level. The point estimate of 0.134 (standard

error 0.023, column 2) suggests that a one standard-deviation increase in population

density reduces the effects of Umuganda by about 28 percent. Thus, Umuganda was

particularly effective in less densely populated areas - bringing people together.

The effectiveness of Umuganda might also depend on the size of the Tutsi minor-

ity. Large Tutsi minorities might boycott or hinder the meetings. However, the data

suggests that this is not the case. The point estimate on the interaction between Sat-

urday rainfall and the Tutsi population size is insignificant and, if anything, negative

(column 3). This is once more unsurprising: since the Tutsi were the clear minority

3.7. DISCUSSION AND CONCLUSION 91

in Rwanda, never holding the majority in any sector, the Hutu elites were not very

concerned about their presence. In fact, taken at face value, the negative point esti-

mate of -1.090 (standard error 1.526) suggests that the meetings were more successful

in sectors with larger Tutsi minorities. The perceived Tutsi threat might have been

more salient in these sectors and the enemy easier to point out. All results are robust

to adding all three heterogeneous effects at once (column 4). The opposite is true in

sectors run by Tutsi party elites. In these sectors, it seems that the local elites had

to use the Umuganda meetings to compensate for the anti-Tutsi propaganda spread

on the radio. The interaction effect of Saturday rainfall with radio ownership among

the Hutu is negative and significant at the 90 percent confidence level. The point esti-

mate of -12.925 (standard error 6.542) suggests that the positive effect of Umuganda

is about 26 percent lower in places with a radio ownership level of one standard devi-

ation, as compared to places with no radio ownership at all (column 5). Furthermore,

the local pro-Tutsi elites seemed to have been more effective in sectors with fewer

Hutu inhabitants. The interaction effect of Saturday rainfall with the size of the Tutsi

minority is positive and almost statistically significant (p-value 0.124) in column 7.

This is consistent with the Tutsi elites having to overcome a potential pro-genocide

bias in the Hutu population. However, population density did not seem to matter in

these sectors. The interaction effect of Saturday rainfall with population density in

column 6 is insignificant and, if anything, positive (0.355, standard error 0.494). Thus,

contrary to the Hutu-run sectors, Umuganda in pro-Tutsi sectors was more successful

in highly populated areas. The above results are once more robust to controlling for

all three heterogeneous effects at once (column 8).

3.7 Discussion and Conclusion

Our results show that the local Hutu elites used mandatory community meetings

to mobilize the civilian population for genocide. Using exogenous variation in heavy

rainfall on the day of the mandatory community-work meetings Umuganda, we find


that one additional rainy community day decreased the share of civilian perpetrators in

the Rwandan genocide by around 5 percent. Interestingly, this negative effect becomes

positive in sectors run by anti-genocide pro-Tutsi parties. Thus, in these sectors, the

meetings may have been used to compensate for the various forms of Hutu propaganda

on the radio and bridge the differences between the two ethnic groups. Our findings

are important for several reasons.

First, a large number of civilians participated in the killings during the Rwandan

genocide. While it is a common understanding that the genocide was centrally planned

and organized, little is known about the link between the planning and the wide accep-

tance of the genocide among the civilian population. Our paper suggests that weekly

held community meetings played a major role in this preparation and mobilization pro-

cess. Second, people getting together during community meetings is commonly said

to foster a sense of belonging and create social capital, generally viewed as positive

for development and community building (see, for example, Knack and Keefer, 1997;

Grootaert and van Bastelaer, 2002; Guiso, Sapienza and Zingalez, 2008). As empha-

sized by Putnam (2000), social capital can bridge the divides in a society. However,

we show that there is a ‘dark side’ to these community meetings. Thus, although the

institution of Umuganda may have the potential to act as a community building force,

our results show that when placed in the wrong hands, the effects can become dis-

astrous. However, somewhat comfortingly, our results also suggest that in Tutsi-led

sectors, Umuganda was used to work against propaganda and overcome hatred.

The more optimistic view of this type of institution might explain why the current

Rwandan government reinstalled Umuganda in 2008. Indeed, official statements about

Umuganda today emphasize values such as "solidarity" and "reconciliation" and the

practice is said to foster a sense of community. These mandatory work days are now

held monthly, on the last Saturday of every month. A similar practice is also present

in Burundi and is being discussed in DR Congo and Kenya. Our analysis clearly shows

that these meetings are powerful instruments and caution is warranted, especially in

countries with histories of ethnic tension.

REFERENCES 93

References

Andre, C. and J.-P. Platteau. 1998. Land relations under unbearable stress:

Rwanda caught in the Malthusian trap, Journal of Economic Behavior and Orga-

nization, 34(1), pp. 1-47.

Besley, T. and T. Persson. 2011. The Logic of Political Violence, Quarterly Journal

of Economics, 126(3), pp. 1411-1445.

Blattman, C. and E. Miguel. 2010. Civil War, Journal of Economic Literature,

48(1), pp. 3-57.

Boudreaux, K. 2009. Land Conflict and Genocide in Rwanda, The Electronic Journal

of Sustainable Development, 1(3), pp. 86-95.

Brückner M. and A. Ciccone. 2010. Rain and the Democratic Window of Oppor-

tunity, Econometrica, 79(3), pp. 923-947.

Chaney, E. 2013. Revolt on the Nile: Economic Shocks, Religion, and Political Power,

Econometrica, 81(5), pp. 2033-2053.

Collins W. J. and R. A. Margo. 2007. The Economic Aftermath of the 1960s

Riots in American Cities: Evidence from Property Values. The Journal of Economic

History, 67(4), pp. 849-883.

Conley, T. G. 1999. GMM Estimation with cross sectional Dependence, Journal of

Econometrics, 92(1), pp. 1-45.

Cook, S. E. (Ed.). 2004. Genocide in Cambodia and Rwanda: new perspectives, No.

1, Transaction Publishers.

Dallaire, R. 2003. Shake hands with the devil, Random House Canada, Toronto.

Daily Nation (retrieved on March 5, 2016). Emuhaya MP Seeks to Introduce Manda-


Davenport, C. and A. Stam. 2009. Rwandan Political Violence in Space and Time,

mimeo


Dell, M. 2012. Trafficking Networks and the Mexican Drug War, mimeo.

Des Forges, A. 1999. Leave None to Tell the Story: Genocide in Rwanda, Human

Rights Watch and the International Federation of Human Rights Leagues, New York,

NY, USA. www.hrw.org/legacy/reports/1999/rwanda/.

Diamond, J. 2005. Collapse: How societies choose to succeed or fail, Viking Penguin,

New York.

Dube, O. and J. F. Vargas. 2013. Commodity Price Shocks and Civil Conflict:

Evidence from Colombia, Review of Economic Studies, 80(4), pp. 1384-1421.

Dyson, L. L. 2009. Heavy Daily Rainfall Characteristics in the Guateng Province,

Water Research Council, 2009-10.

Eisinga, R., M. Grotenhuis and B. Pelzer. 2012. Weather Conditions and Politi-

cal Party Vote Share in Dutch National Parliament Election, 1971-2010, International

Journal of Biometeorology, 56(6), pp. 1161-1165.

Fraga, B. and E. Hersh 2011. Voting Costs and Voter Turnout in Competitive

Elections, Quarterly Journal of Political Science, 5(4), pp. 339-356.

Friedman, W. 2013. Local Economic Conditions and Participation in the Rwandan

Genocide, mimeo.

Gleditsch, N. P., Wallensteen, P., Eriksson, M., Sollenberg, M. and H.

Strand. 2002. Armed Conflict 1946-2001: A New Dataset, Journal of Peace Research,

29(5), pp. 615-637.

Gomez D. P., Aronow, P. M. and M. C. McGrath. 2012. Field Experiments and

the Study of Voter Turnout, Journal of Elections, Public Opinion & Parties, 23(1),

pp. 27-48.

Gourevitch, P. 1998. We wish to inform you that tomorrow we will be killed with our

families, Farrar, Straus & Giroux, New York.



REFERENCES 95





295-320.

Hansford, T. G. and B. T. Gomez. 2010. Estimating the Electoral Effects of Voter

Turnout, American Political Science Review, 104(2), pp. 268-288.

Harari, M. and E. La Ferrara. 2013. Conflict, Climate and Cells: A Disaggregated

Analysis, C.E.P.R. Discussion Papers 9277.

Hatzfeld, J. 2005. Machete season: The Killers in Rwanda speak, Picador, New York.

Hatzfeld, J. 2006. Life laid bare: The Survivors in Rwanda speak, Other Press, New

York.

Heldring, L. 2014. State Capacity and Violence: Evidence form the Rwandan Geno-

cide, CSAE Working Paper, WPS/2014-08.

Horiuchi, Y. and J. Saito. 2009. Rain, Election, and Money: The Impact of Voter

Turnout on Distributive Policy Outcomes, Yale, mimeo

Human Rights Watch. 2011. Justice Compromised The Legacy of Rwanda’s Community-

Based Gacaca Courts, Report, May 2011.

IRDP. 2003. Sustaining Peace in Rwanda: Voice of the People, Institut de Recherche

et de Dialogue pour la Paix.

Iyer, L. and P. Topalova. 2014. Poverty and Crime: Evidence from Rainfall and

Trade Shocks in India, mimeo.

Kapiteni, A. 1996. La Première Estimation du Homre des Victims du Genocide du

Rwanda de 1994 Commune par Commune en Fev 1996, Report of the Ministry of

Higher Education, Scientific Research, and Culture.

Kirschke, L. 1996. Broadcasting genocide: censorship, propaganda & state-sponsored

violence in Rwanda 1990-1994, Article 19.




Lawrence, R. and P. Uwimbabazi. 2013. Indigenous Practice, Power and Social

Control: The Paradox of the Practice of Umuganda in Rwanda in Race, Power and

Indigenous Knowledge Systems, Interdisciplinary Journal for the Study of the Arts

and Humanities in Southern Africa, 20(1), pp. 248-272.

Lind, J. T. 2014. Rainy Day Politics: An Instrumental Variables Approach to the

Effect of Parties on Political Outcomes, University of Oslo, mimeo.

Longman, T. 2009. An Assessment of Rwanda’s Gacaca Courts, Peace Review, 21(3),

pp. 304-312.

Madestam A., Shoag, D., Veuger S. and D. Yanagizawa-Drott. 2013. Do

Political Protests Matter? Evidence from the Tea Party Movement, Quarterly Journal

of Economics, 128(4), pp. 1633-1685.

Madestam A. and D. Yanagizawa-Drott. 2011. Shaping the Nation: The Effect

of Fourth of July on Political Preferences and Behavior in the United States, mimeo.

Mamdani, M. 2001. When Victims Become Killers: Colonialism, Nativism, and the

Genocide in Rwanda, Fountain, Kampala.

Mansuri, G. and V. Rao. 2012. Localizing development: does participation work?,


Melvern, L. 2000. A People Betrayed: The Role of the West in Rwanda’s Genocide,

Cape Town.

Miguel, E., Satyanath, S. and E. Sergenti. 2004. Economic Shocks and Civil

Conflict: An Instrumental Variables Approach, Journal of Political Economy, 112(4),

pp. 725-753.

Mitra, A. and D. Ray. 2014. Implications of an Economic Theory of Conflict:

Hindu-Muslim Violence in India, Journal of Political Economy, 122(4), pp. 719-765.

REFERENCES 97

Mukarubuga, C. 2006. The Experience of Social Forums Against Poverty: The Case

of Rwanda, Agency for Co-Operation and Research in Development.

Nunn, N. and N. Qian. 2014. U.S. Food Aid on Civil Conflict, American Economic

Review, 104(6), pp. 1630-1666.

Percival, V. and T. Homer-Dixon. Environmental Scarcity and Violent Conflict:

The Case of Rwanda, The Journal of Environment and Development, 5(3), pp. 270-291.


Pottier, J. 2006. Land Reform for Peace: Rwanda’s 2005 Land Law in Context,

Journal of Agrarian Change, 6(4), pp. 509-537.

Prunier, G. 1995. The Rwanda Crisis: History of a Genocide, Hurst and Company,

London.

Rogall, T. 2014. Mobilizing the Masses for Genocide, IIES, Stockholm University,

mimeo.

Sarsons, H. 2011. Rainfall and Conflict, mimeo.



Straus, S. 2004. How Many Perpetrators Were There in the Rwandan Genocide? An

Estimate, Journal of Genocide Research, 6(1), pp. 85-98.



Thomson, S. M. 2009. Resisting Reconciliation: State Power and Everyday Life in

Post-Genocide Rwanda, PhD dissertation, Halifax, Nova Scotia, Dalhousie University.

Uvin, P. 1998. Aiding Violence: the Development Enterprise in Rwanda, Kumarian

Press.

Uwimbabazi, P. 2012. An Analysis of Umuganda: The Policy and Practice of Com-

munityWork in Rwanda, College of Humanities, KwaZulu-Natal, South Africa, Septem-

ber.


Verpoorten, M. 2012a. The intensity of the Rwandan genocide: Fine measures from

the gacaca records, Peace Economics, Peace Science and Public Policy, 18(1), pp. 1-26.

Verpoorten, M. 2012b. Leave None to Claim the Land: A Malthusian Catastrophe

in Rwanda?, Journal of Peace Research, 49(4), pp. 547-563.

Verpoorten, M. 2012c. Detecting Hidden Violence: The Spatial Distribution of Ex-

cess Mortality in Rwanda. Political Geography. 31(1), pp. 44-56.

Verwimp, P. 2000. Development Ideology, the Peasantry, and Genocide: Rwanda

Represented in Habyarimana’s Speeches, Journal of Genocide Research, 2(3), pp. 325-

361.

Verwimp, P. 2003. Testing the Double-Genocide Thesis for Central and Southern

Rwanda, Journal of Conflict Resolution, 47(4), pp. 423-442.

Verwimp, P. 2005. An Economic Profile of Peasant Perpetrators of Genocide: Micro-

level Evidence from Rwanda, Journal of Development Economics, 77(2), pp. 297- 323.

Verwimp, P. 2006. Machetes and Firearms: the Organization of Massacres in Rwanda,

Journal of Peace Research, 43(1), pp. 5-22.



Viret, E. 2010. Rwanda - A Chronology (1867-1994), Online Encyclopedia of Mass

Violence, published on 1 March 2010, ISSN 1961-9898., accessed 27 October 2014 from

http://www.massviolence.org/Rwanda-A-Chronology

Yanagizawa-Drott, D. 2014. Propaganda and Conflict: Evidence from the Rwandan

Genocide, Quarterly Journal of Economics, 129(4), pp. 1947-1994.


Figures and Tables


Figure 3.1: Placebo Check: Civilian Partcipation

Figure 3.2: Placebo Check: Organized Participation

Notes: The figures shows the distribution of coefficients on Saturday Rainfall for civilian violence (Figure 3.1) andorganized violence (Figure 3.2) when using Saturday rainfall during the 3.5 years of the pre-genocide period (October1, YEAR to March 31, YEAR+4) from the years 1984 to 2013.


Figure 3.3: Rainfall map

Note: Map of Rwanda showing the number of rainy Saturdays in 1990-1994 by sector. Communeborders are indicated on the map as grey lines.


Figure 3.4: Prosecuted civilians

Note: Map of Rwanda showing the number of prosecuted civilians by sector.


Table 3.1: Summary Statistics

Mean Std.dev. Obs.

A. Violence & Population

# Civilian Perpetrators 290.25 286.43 1433# Organized Perpetrators 51.76 70.51 1433% Civilian Perpetrators per Hutu (p.H.) 7.66 7.93 1433% Organized Perpetrators per Hutu (p.H.) 1.40 2.09 1433Pre-Genocide Violence against Tutsi, dummy 0.15 0.36 1433Mass Grave found in Sector, dummy 0.05 0.21 1432Population in Sector, ’000 4.88 2.48 1433Hutu Population in Sector, ’000 4.26 2.17 1433Population Density, ’000 0.50 0.85 1433

B. Rainfall

# Sat(Rain>10mm) 18.25 4.24 1433# Sun(Rain>10mm) 15.14 5.19 1433# Mon(Rain>10mm) 15.13 4.22 1433# Tue(Rain>10mm) 18.10 3.52 1433# Wed(Rain>10mm) 20.51 4.76 1433# Thu(Rain>10mm) 21.53 3.97 1433# Fri(Rain>10mm) 17.02 4.75 1433Average Daily Rainfall, 1980s 2.58 0.48 1433Average Daily Rainfall, 1990s 2.44 0.55 1433# Pub. Holidays(Rain>10mm) 0.85 0.20 1433# Non-Rel. Holidays(Rain>10mm) 1.56 0.21 1433# Rel. Holidays(Rain>10mm) 1.00 0.11 1433

C. Other Variables

Fraction of Hutu with Radio 0.33 0.09 1433Tutsi Minority Share 0.10 0.13 1433Distance to Kigali (km) 3.99 0.64 1433Distance to Main City (km) 2.91 0.71 1433Distance to Nyanza (km) 4.00 0.66 1433Distance to the Main Road (km) 1.41 1.23 1433Distance to the Border (km) 2.82 0.91 1433

Note: The # prosecuted militiamen is crime category 1: prosecutions against organizers, leaders, armyand militia; # prosecuted civilians is crime category 2: prosecutions against civilians. The per Hutu(p.H.) variables are expressed in percent. Pre-Genocide Violence against Tutsi is a dummy takingthe value of 1 if the sector experienced violence against Tutsi in the pre-genocide period. The twoaverage daily rainfall variables are measured in millimeters. The distance variables are measured inkilometers. Population is the population number in the sector and Population Density is populationper square kilometers, from the 1991 census. Radio ownership and ethnicity data are taken from the1991 census, available only at the commune level. There are 142 communes in the sample. The TutsiMinority Share is defined as the fraction of Tutsi normalized by the fraction of Hutu.


Table 3.2: Main Effects

Dependent Variable: % Civilian Perpetrators, p.H. % Militiamen, p.H.

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

# Sat(Rainfall>10mm) -0.580∗∗∗ -0.425∗∗∗ -0.409∗∗∗ -0.115∗∗∗ -0.065∗ -0.057∗[0.118] [0.125] [0.128] [0.033] [0.033] [0.030]

# Sun(Rainfall>10mm) 0.041 -0.037[0.102] [0.031]

# Mon(Rainfall>10mm) 0.080 0.100∗∗∗[0.112] [0.031]

# Tue(Rainfall>10mm) 0.023 -0.046[0.084] [0.030]

# Wed(Rainfall>10mm) 0.031 0.007[0.111] [0.028]

# Thu(Rainfall>10mm) -0.007 -0.064[0.134] [0.041]

# Fri(Rainfall>10mm) -0.057 0.006[0.099] [0.027]

Standard Controls yes yes yes yes yes yesCommune Effects no yes yes no yes yesR2 0.15 0.52 0.52 0.07 0.36 0.37N 1433 1433 1433 1433 1433 1433

Note: # of Sat(Rainfall>10mm) is the number of Saturdays with rainfall above 10mm during the periodOctober 1990 to March 1994 (and similarly for all other weekdays). % Civilian Perpetrators per Hutu (p.H)and % Militiamen per Hutu are measured in percent. Standard Controls include average daily rainfall forJanuary 1984 to September 1990 and average daily rainfall for October 1990 to March 1994. All regressions arerun using weighted least squares (WLS) estimation with population size as weights. There are 142 communesin the sample. Standard errors are clustered at the commune level. *significant at 10 percent, **significant at5 percent, ***significant at 1 percent.


Table 3.3: Different Rainfall Thresholds

Dependent variable: % Civilian Perpetrators, p.H.

Rainfall Threshold x: 6 mm 8 mm 9 mm 10 mm 12 mm

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

# Sat(Rainfall > x mm) −0.142 −0.241∗ −0.250∗ −0.409∗∗∗ −0.385∗∗∗

[0.115] [0.140] [0.141] [0.128] [0.132]# Sun(Rainfall > x mm) 0.080 0.068 0.073 0.041 −0.043

[0.084] [0.124] [0.117] [0.102] [0.137]# Mon(Rainfall > x mm) 0.069 0.009 0.079 0.080 −0.053

[0.088] [0.118] [0.117] [0.112] [0.120]# Tue(Rainfall > x mm) −0.020 0.000 0.069 0.023 0.135

[0.128] [0.123] [0.099] [0.084] [0.123]# Wed(Rainfall > x mm) 0.003 0.043 −0.065 0.031 −0.058

[0.093] [0.111] [0.106] [0.111] [0.118]# Thu(Rainfall > x mm) 0.129 0.004 0.140 −0.007 −0.233∗∗

[0.096] [0.107] [0.123] [0.134] [0.107]# Fri(Rainfall > x mm) −0.048 0.106 −0.079 −0.057 −0.216

[0.086] [0.094] [0.086] [0.099] [0.137]Standard Controls yes yes yes yes yesCommune Effects yes yes yes yes yesR2 0.51 0.51 0.51 0.52 0.52N 1433 1433 1433 1433 1433

Note: # of Sat(Rainfall>x mm) is the number of Saturdays with rainfall above x mm during the period October 1990 toMarch 1994 (and similarly for all other weekdays). The value of x is given in the column header. % Civilian Perpetratorsper Hutu (p.H) is measured in percent. Standard Controls include average daily rainfall for January 1984 to September1990 and average daily rainfall for October 1990 to March 1994. All regressions are run using weighted least squares(WLS) estimation with population size as weights. There are 142 communes in the sample. Standard errors are clusteredat the commune level. *significant at 10 percent, **significant at 5 percent, ***significant at 1 percent.

106 PREPARING FOR GENOCIDE: COMMUNITY MEETINGS IN RWANDATable

3.4:Robustness

andPlacebo

Tests

Dependent

variable:%

Civilian

Perpetrators,

p.H.

Massgrave

inVillage

Without

Without

Without

Additional

FutureAlternative

Mass

Graves

Kigali

Major

Cities

Controls

Rainfall

Dep.

Var.

(1)(2)

(3)(4)

(5)(6)

(7)(8)

#Sat(R

ainfall>10m

m)

−0.403 ∗∗∗

−0.426 ∗∗∗

−0.425 ∗∗∗

−0.368 ∗∗∗

−0.452 ∗∗∗

−0.015 ∗∗∗

−0.013 ∗∗∗

[0.129][0.131]

[0.133][0.125]

[0.126][0.004]

[0.004]

#Sun(R

ainfall>10m

m)

0.0540.048

0.0650.033

0.0510.002

[0.103][0.106]

[0.108][0.107]

[0.110][0.004]

#Mon(R

ainfall>10m

m)

0.0560.076

0.0630.113

0.108−

0.003[0.109]

[0.113][0.115]

[0.114][0.101]

[0.004]#

Tue(R

ainfall>10m

m)

0.0740.028

0.0490.021

0.0620.008

[0.086][0.089]

[0.089][0.080]

[0.086][0.004] ∗

#Wed(R

ainfall>10m

m)

0.0290.015

0.0350.033

0.0550.006

[0.107][0.121]

[0.127][0.105]

[0.133][0.004]

#Thu(R

ainfall>10m

m)

−0.011

0.0210.018

0.0190.059

−0.003

[0.128][0.136]

[0.140][0.126]

[0.144][0.004]

#Fri(R

ainfall>10m

m)

−0.014

−0.054

−0.034

−0.006

−0.034

−0.009 ∗∗

[0.097][0.099]

[0.104][0.098]

[0.103][0.003]

#Sat(R

ainfall>10m

m),94-98

−0.012

0.008[0.106]

[0.111]#

Sun(Rainfall>

10mm),94-98

0.1300.111

[0.114][0.112]

#Mon(R

ainfall>10m

m),94-98

−0.279 ∗∗

−0.323 ∗∗

[0.118][0.139]

#Tue(R

ainfall>10m

m),94-98

−0.153

−0.099

[0.109][0.105]

#Wed(R

ainfall>10m

m),94-98

−0.168

−0.231

[0.152][0.155]

#Thu(R

ainfall>10m

m),94-98

−0.123

−0.113

[0.128][0.131]

#Fri(R

ainfall>10m

m),94-98

0.1240.200 ∗

[0.110][0.103]

StandardControls

yesyes

yesyes

yesyes

yesyes

Additional

Controls

nono

noyes

nono

nono

Com

mune

Effects

yesyes

yesyes

yesyes

yesyes

R2

0.510.51

0.510.52

0.510.52

0.160.16

N1367

14221358

14331433

14331432

1432

Note:#

ofSat(R

ainfall>10m

m)isthe

number

ofSaturdayswith

rainfallabove10m

mduring

theperiod

October

1990to

March

1994(and

similarly

forall

otherweekdays).

%Civilian

Perpetrators

perHutu

(p.H)ismeasured

inpercent.

Incolum

n1wedrop

sectorswith

mass

graves(indicating

highdeath

rates).In

column2wedrop

sectorsin

thecapital

Kigali

andin

column3wedrop

allsectors

inand

closeto

themain

cities.In

column

4weadd

additionalcontrols.These

arepopulation

density,distanceto

Kigali,N

yanza,theborder,the

closestmain

roadand

theclosest

main

city.In

columns

5and

6,wealso

controlfor

futurerainfall.

Incolum

ns7and

8,weuse

adum

myindicating

whether

amass

gravewas

foundin

thesector

asan

alternativedependent

variable.StandardControls

includeaverage

dailyrainfallfor

January1984

toSeptem

ber1990

andaverage

dailyrainfall

forOctober

1990to

March

1994.Allregressions

arerun

usingweighted

leastsquares

(WLS)

estimation

with

populationsize

asweights.

There

are142

communes

inthe

sample.

Standarderrors

areclustered

atthe

commune

level.*significantat

10percent,**significant

at5percent,

***significantat

1percent.


Tab

le3.5:

Exclusion

Restriction

Dep

endent

variab

le:

%Civilian

Perpe

trators,

p.H.

Excl.Pre-

LocalH

utu

LocalT

utsi

Pub

licHoliday

sViolence

Lead

ers

Lead

ers

(1)

(2)

(3)

(4)

(5)

(6)

(7)

#Sa

t(Rainfall>

10mm)

−0.

407∗∗∗

−0.

393∗∗∗

−0.

483∗∗∗

−0.

481∗∗∗

−0.

466∗∗∗

0.70

60.

399

[0.1

24]

[0.1

22]

[0.1

49]

[0.1

20]

[0.1

23]

[0.8

96]

[0.7

96]

#Su

n(Rainfall>

10mm)

0.04

50.

049

0.05

30.

043

0.58

3[0

.103]

[0.1

06]

[0.1

21]

[0.0

98]

[0.8

96]

#Mon

(Rainfall>

10mm)

0.08

20.

074

0.11

60.

045

0.08

7[0

.110]

[0.1

11]

[0.1

31]

[0.1

10]

[0.3

87]

#Tue(R

ainfall>

10mm)

0.03

10.

040

0.00

20.

037

0.00

2[0

.095]

[0.0

94]

[0.0

95]

[0.0

82]

[0.4

24]

#Wed(R

ainfall>

10mm)

0.03

00.

028

−0.

025

−0.

002

0.67

3∗

[0.1

13]

[0.1

10]

[0.1

16]

[0.1

16]

[0.3

49]

#Thu

(Rainfall>

10mm)

−0.

006

−0.

000

0.12

6−

0.07

20.

814

[0.1

34]

[0.1

35]

[0.1

51]

[0.1

27]

[0.6

81]

#Fri(Rainfall>

10mm)

−0.

050

0.00

5−

0.01

60.

010

−0.

193

[0.1

17]

[0.1

32]

[0.1

18]

[0.0

92]

[0.4

00]

#Pub

.Holidays(Rainfall>

10mm)

−0.

597

[2.3

69]

#Non

-Rel.H

olidays(Rainfall>

10mm)

−1.

439

[1.7

59]

#Rel.H

olidays(Rainfall>

10mm)

−5.

424

[3.8

71]

Stan

dard

Con

trols

yes

yes

yes

yes

yes

yes

yes

Com

mun

eEffe

cts

yes

yes

yes

yes

yes

yes

yes

R2

0.52

0.52

0.49

0.54

0.55

0.32

0.35

N1433

1433

1213

1272

1272

161

161

Note:

#of

Sat(Rainfall>10

mm)is

thenu

mbe

rof

Saturdayswithrainfallab

ove10mm

during

thepe

riod

Octob

er1990

toMarch

1994

(and

simila

rlyfor

allo

ther

weekd

ays).%

Civilian

PerpetratorsperHutu(p.H

)ismeasuredin

percent.In

columns

1an

d2wealso

controlfor

thenu

mbe

rof

public

holid

ays

(separated

into

relig

ious

andno

n-relig

ious

holid

aysin

column2)

withrainfallab

ove10mm.In

column3wedrop

sectorswhere

violence

againstTutsi

took

placebe

fore

thegeno

cide.In

columns

4an

d5thesampleis

restricted

tosectorswithpro-geno

cide

partiesrulin

gthecommun

e.In

columns

6an

d7,

thesampleis

restricted

tosectorswithan

ti-genocidepa

rtiesrulin

gthecommun

e.Stan

dard

Con

trolsinclud

eaverageda

ilyrainfallforJa

nuary1984

toSeptem

ber1990

andaverageda

ilyrainfallforOctob

er1990

toMarch

1994.Allregression

sarerunusingweigh

tedleastsqua

res(W

LS)

estimation

withpo

pulation

size

asweigh

ts.There

are14

2commun

esin

thesample.

Stan

dard

errors

areclusteredat

thecommun

elevel.*significan

tat

10pe

rcent,

**sign

ificant

at5pe

rcent,***significan

tat

1pe

rcent.


Table

3.6:Channels

-Interaction

Effects

Dependent

variable:%

Civilian

Perpetrators,p.H

.

LocalHutu

LeadersLocalT

utsiLeaders

(1)(2)

(3)(4)

(5)(6)

(7)(8)

#Sat(R

ainfall>10m

m)

−0.695 ∗∗

−0.533 ∗∗∗

−0.378 ∗∗∗

−0.541 ∗

4.292 ∗0.295

−1.463

1.820[0.294]

[0.124][0.140]

[0.309][2.396]

[0.801][0.870]

[1.470]#

Sat(Rainfall>

10mm)x

Radio

Own-

ership0.659

0.414−

12.925 ∗−

11.722 ∗

[0.786][0.895]

[6.543][6.572]

#Sat(R

ainfall>10m

m)xPopulation

Density

0.134 ∗∗∗0.125 ∗∗∗

0.3550.279

[0.023][0.032]

[0.494][0.540]

#Sat(R

ainfall>10m

m)x

Tutsi

Mi-

norityShare

−1.090

−1.505

10.01610.834 ∗

[1.526][1.533]

[6.208][6.133]

StandardControls

yesyes

yesyes

yesyes

yesyes

Other

Weekday

Controls

yesyes

yesyes

yesyes

yesyes

Com

mune

Effects

yesyes

yesyes

yesyes

yesyes

R2

0.550.55

0.550.55

0.360.36

0.360.38

N1272

12721272

1272161

161161

161

Note:

#of

Sat(Rainfall>

10mm)is

thenum

berof

Saturdayswith

rainfallabove

10mm

duringthe

periodOctober

1990to

March

1994.%

Civilian

Perpetrators

perHutu

(p.H)is

measured

inpercent.

Radio

Ownership

isthe

fractionof

Hutu

owning

aradio.

Tutsi

Minority

Shareis

thefraction

ofTutsi

dividedby

thefraction

ofHutu.

Incolum

ns1to

4the

sample

isrestricted

tovillages

with

pro-genocideHutu

partiesruling

thecom

mune.

Incolum

ns5to

8the

sample

isrestricted

tovillages

with

anti-genocidepro-T

utsiparties

rulingthe

commune.

StandardControls

includeaverage

dailyrainfall

forJanuary

1984to

September

1990and

averagedaily

rainfallfor

October

1990to

March

1994.Other

Weekday

Controls

includethe

number

ofSun/M

on/Tue/W

ed/Thu/Fri

with

rainfallabove

10mm

duringthe

periodOctober

1990to

March

1994.Allregressions

arerun

usingweighted

leastsquares

(WLS)

estimation

with

populationsize

asweights.

There

are142

communes

inthe

sample.

Standarderrors

areclustered

atthe

commune

level.*significant

at10

percent,**significant

at5percent,

***significantat

1percent.

APPENDIX 109

Appendix

Table A.1: Main Effects - Linear Specification


(1) (2)

Average Rainfall Sat −4.093∗∗ −0.569[1.635] [0.355]

Average Rainfall Sun 0.541 −0.175[1.783] [0.480]

Average Rainfall Mon 0.306 0.507[1.543] [0.432]

Average Rainfall Tue 1.378 0.287[1.351] [0.422]

Average Rainfall Wed 1.703 −0.107[1.454] [0.303]

Average Rainfall Thu −0.082 −0.057[1.113] [0.298]

Average Rainfall Fri 0.262 −0.001[0.983] [0.235]

Standard Controls yes yesCommune Effects yes yesR2 0.51 0.36N 1433 1433

Note: Average Rainfall Sat is the average daily Saturday rainfall during the period from October1990 to March 1994 (and similarly for all other weekdays). % Civilian Perpetrators per Hutu (p.H) ismeasured in percent. Standard Controls include average daily rainfall for January 1984 to September1990. All regressions are run using weighted least squares (WLS) estimation with population sizeas weights. There are 142 communes in the sample. Standard errors are clustered at the communelevel. *significant at 10 percent, **significant at 5 percent, ***significant at 1 percent.


Table A.2: Conley Standard Errors


25 km 50 km 75 km 25 km 50 km 75 km

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

# Sat(Rainfall>10mm) −0.391∗∗∗ −0.391∗∗∗ −0.391∗∗∗ −0.053 −0.053∗ −0.053∗∗∗

[0.129] [0.134] [0.132] [0.035] [0.029] [0.026]∗∗

Number Sun>10 0.054 0.054 0.054 −0.031 −0.031 −0.031[0.097] [0.088] [0.078] [0.032] [0.034] [0.032]

Number Mon>10 0.129 0.129 0.129 0.121∗∗∗ 0.121∗∗∗ 0.121∗∗∗

[0.096] [0.108] [0.108] [0.036] [0.038] [0.041]Number Tue>10 0.062 0.062 0.062 −0.049∗ −0.049∗ −0.049∗

[0.104] [0.119] [0.116] [0.029] [0.027] [0.026]Number Wed>10 0.071 0.071 0.071 0.010 0.010 0.010

[0.105] [0.09] [0.087] [0.028] [0.024] [0.021]Number Thu>10 0.025 0.025 0.025 −0.046 −0.046 −0.046

[0.128] [0.14] [0.155] [0.036] [0.036] [0.035]Number Fri>10 −0.057 −0.057 −0.057 0.003 0.003 0.003

[0.11] [0.103] [0.09] [0.027] [0.03] [0.03]Standard Controls yes yes yes yes yes yesCommune Effects yes yes yes yes yes yesR2 0.48 0.48 0.48 0.37 0.37 0.37N 1433 1433 1433 1433 1433 1433

Note: # of Sat(Rainfall>10mm) is the number of Saturdays with rainfall above 10mm during the period October 1990to March 1994 (and similarly for all other weekdays). % Civilian Perpetrators per Hutu (p.H) and % Militiamen perHutu are measured in percent. Standard Controls include average daily rainfall for January 1984 to September 1990and average daily rainfall for October 1990 to March 1994. There are 142 communes in the sample. Standard errorscorrecting for spatial correlation within a radius of 25km, 50km and 75km are in square brackets, Conley (1999). Theradius used in each regression is given in the column header. *significant at 10 percent, **significant at 5 percent,***significant at 1 percent.

APPENDIX 111

Table A.3: Different Rainfall Thresholds

Dependent variable: % Militiamen, p.H.

Rainfall Threshold x: 6 mm 8 mm 9 mm 10 mm 12 mm

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

# Sat(Rainfall > x mm) 0.020 −0.037 −0.041 −0.057∗ −0.072∗∗

[0.030] [0.032] [0.030] [0.030] [0.034]# Sun(Rainfall > x mm) −0.022 0.005 −0.009 −0.037 −0.014

[0.027] [0.034] [0.033] [0.031] [0.038]# Mon(Rainfall > x mm) 0.021 0.024 0.05∗ 0.100∗∗∗ 0.060∗

[0.039] [0.034] [0.027] [0.031] [0.031]# Tue(Rainfall > x mm) 0.020 0.022 0.011 −0.046 0.022

[0.029] [0.031] [0.026] [0.030] [0.037]# Wed(Rainfall > x mm) −0.001 −0.027 −0.021 0.007 −0.016

[0.027] [0.037] [0.035] [0.028] [0.036]# Thu(Rainfall > x mm) −0.002 −0.015 −0.014 −0.064 −0.045

[0.024] [0.036] [0.039] [0.041] [0.042]# Fri(Rainfall > x mm) 0.023 0.049 −0.008 0.006 −0.026

[0.031] [0.035] [0.030] [0.027] [0.035]Standard Controls yes yes yes yes yesCommune Effects yes yes yes yes yesR2 0.36 0.36 0.36 0.37 0.36N 1433 1433 1433 1433 1433

Note: # of Sat(Rainfall>x mm) is the number of Saturdays with rainfall above x mm during the period October 1990to March 1994 (and similarly for all other weekdays). The value of x is given in the column header. % Militiamen perHutu (p.H) is measured in percent. Standard Controls include average daily rainfall for January 1984 to September 1990and average daily rainfall for October 1990 to March 1994. All regressions are run using weighted least squares (WLS)estimation with population size as weights. There are 142 communes in the sample. Standard errors are clustered atthe commune level. *significant at 10 percent, **significant at 5 percent, ***significant at 1 percent.

112 PREPARING FOR GENOCIDE: COMMUNITY MEETINGS IN RWANDATable

A.4:R

obustnessand

Placebo

Tests

Dependent

variable:%

Militiam

en,p.H

.Massgrave

inVillage

Without

Without

Without

Additional

FutureAlternative

Mass

Graves

Kigali

Major

Cities

Controls

Rainfall

Dep.

Var.

(1)(2)

(3)(4)

(5)(6)

(7)(8)

#Sat(R

ainfall>10m

m)

−0.044

−0.061 ∗∗

−0.069 ∗∗

−0.052 ∗

−0.069 ∗∗

−0.015 ∗∗∗

−0.013 ∗∗∗

[0.028][0.031]

[0.031][0.030]

[0.030][0.004]

[0.004]

#Sun(R

ainfall>10m

m)

−0.026

−0.037

−0.030

−0.036

−0.034

0.002[0.031]

[0.032][0.031]

[0.032][0.030]

[0.004]#

Mon(R

ainfall>10m

m)

0.089 ∗∗∗0.098 ∗∗∗

0.097 ∗∗∗0.112 ∗∗∗

0.104 ∗∗∗−

0.003[0.029]

[0.031][0.031]

[0.031][0.032]

[0.004]#

Tue(R

ainfall>10m

m)

−0.035

−0.048

−0.041

−0.057 ∗

−0.038

0.008 ∗

[0.030][0.031]

[0.031][0.029]

[0.029][0.004]

#Wed(R

ainfall>10m

m)

0.0120.001

0.0120.002

0.0090.006

[0.027][0.030]

[0.031][0.028]

[0.027][0.004]

#Thu(R

ainfall>10m

m)

−0.062

−0.062

−0.073 ∗

−0.052

−0.054

−0.003

[0.041][0.042]

[0.043][0.036]

[0.041][0.004]

#Fri(R

ainfall>10m

m)

0.0160.004

0.0040.015

0.017−

0.009 ∗∗

[0.026][0.027]

[0.029][0.025]

[0.026][0.003]

#Sat(R

ainfall>10m

m),94-98

0.0020.014

[0.028][0.028]

#Sun(R

ainfall>10m

m),94-98

0.078 ∗∗0.075 ∗∗

[0.033][0.031]

#Mon(R

ainfall>10m

m),94-98

−0.063 ∗

−0.060 ∗∗

[0.032][0.030]

#Tue(R

ainfall>10m

m),94-98

0.0390.043

[0.039][0.036]

#Wed(R

ainfall>10m

m),94-98

−0.022

−0.033

[0.037][0.033]

#Thu(R

ainfall>10m

m),94-98

−0.021

−0.025

[0.035][0.035]

#Fri(R

ainfall>10m

m),94-98

0.0390.050 ∗∗

[0.027][0.025]

StandardControls

yesyes

yesyes

yesyes

yesyes

Additional

Controls

nono

noyes

nono

nono

Com

mune

Effects

yesyes

yesyes

yesyes

yesyes

R2

0.370.37

0.380.38

0.360.38

0.160.16

N1367

14221358

14331433

14331432

1432

Note:#

ofSat(R

ainfall>10m

m)isthe

number

ofSaturdayswith

rainfallabove10m

mduring

theperiod

October

1990to

March

1994(and

similarly

forall

otherweekdays).

%Militiam

enper

Hutu

(p.H)is

measured

inpercent.

Incolum

n1wedrop

sectorswith

mass

graves(indicating

highdeath

rates).In

column2wedrop

sectorsin

thecapital

Kigali

andin

column3wedrop

allsectors

inand

closeto

themain

cities.In

column4

weadd

additionalcontrols.

These

arepopulation

density,distance

toKigali,

Nyanza,

theborder,

theclosest

main

roadand

theclosest

main

city.In

columns

5and

6wealso

controlfor

futurerainfall.

Incolum

ns7and

8weuse

adum

myindicating

whether

amass

gravewas

foundin

thesector

asan

alternativedependent

variable.StandardControls

includeaverage

dailyrainfallfor

January1984

toSeptem

ber1990

andaverage

dailyrainfall

forOctober

1990to

March

1994.Allregressions

arerun

usingweighted

leastsquares

(WLS)

estimation

with

populationsize

asweights.

There

are142

communes

inthe

sample.

Standarderrors

areclustered

atthe

commune

level.*significantat

10percent,**significant

at5percent,

***significantat

1percent.

APPENDIX 113

Tab

leA.5:E

xclusion

Restriction

Dep

endent

variab

le:

%Militiam

en,p

.H.

Excl.Pre-

LocalH

utu

LocalT

utsi

Pub

licHoliday

sViolence

Lead

ers

Lead

ers

(1)

(2)

(3)

(4)

(5)

(6)

(7)

#Sa

t[Rainfall>

10mm)

−0.

052∗

−0.

054∗

−0.

048

−0.

092∗∗∗

−0.

083∗∗∗

0.38

60.

214

[0.0

29]

[0.0

29]

[0.0

32]

[0.0

31]

[0.0

28]

[0.2

23]

[0.1

86]

#Su

n(Rainfall>

10mm)

−0.

030

−0.

033

−0.

019

−0.

034

−0.

017

[0.0

30]

[0.0

30]

[0.0

38]

[0.0

31]

[0.1

80]

#Mon

(Rainfall>

10mm)

0.10

4∗∗∗

0.10

4∗∗∗

0.10

7∗∗∗

0.06

6∗∗

0.41

9∗∗∗

[0.0

31]

[0.0

31]

[0.0

33]

[0.0

28]

[0.1

40]

#Tue(R

ainfall>

10mm)

−0.

032

−0.

033

−0.

057∗

−0.

035

−0.

214∗

[0.0

29]

[0.0

31]

[0.0

30]

[0.0

31]

[0.1

20]

#Wed(R

ainfall>

10mm)

0.00

40.

005

0.00

50.

007

−0.

101

[0.0

30]

[0.0

29]

[0.0

27]

[0.0

29]

[0.1

66]

#Thu

(Rainfall>

10mm)

−0.

061

−0.

060

−0.

030

−0.

080∗

0.04

5[0

.042]

[0.0

41]

[0.0

40]

[0.0

41]

[0.1

73]

#Fri(Rainfall>

10mm)

0.01

80.

026

0.03

30.

010

0.06

4[0

.028]

[0.0

30]

[0.0

27]

[0.0

28]

[0.1

53]

#Pub

.Holidays(Rainfall>

10mm)

−1.

055∗

[0.5

57]

#Non

-Rel.H

olidays(Rainfall>

10mm)

−0.

777∗∗

[0.3

59]

#Rel.H

olidays(Rainfall>

10mm)

−0.

837

[0.8

66]

Stan

dard

Con

trols

yes

yes

yes

yes

yes

yes

yes

Com

mun

eEffe

cts

yes

yes

yes

yes

yes

yes

yes

R2

0.37

0.37

0.36

0.36

0.37

0.33

0.37

N1433

1433

1213

1272

1272

161

161

Note:#

ofSa

t(Rainfall>10

mm)isthenu

mbe

rof

Saturdayswithrainfallab

ove10mm

during

thepe

riod

Octob

er1990

toMarch

1994

(and

simila

rlyforall

otherweekd

ays).%

Militiam

enperHutu(p.H

)is

measuredin

percent.

Incolumns

1an

d2,

wealso

controlforthenu

mbe

rof

public

holid

ays(separated

into

relig

ious

andno

n-relig

ious

holid

aysin

column2)

withrainfallab

ove10mm.Incolumn3,

wedrop

sectorswereviolence

againstTutsitook

placebe

fore

thegeno

cide.In

columns

4an

d5,

thesampleis

restricted

tosectorswithpro-geno

cide

partiesrulin

gthecommun

e.In

columns

6an

d7,

thesampleis

restricted

tosectorswithan

ti-genocidepa

rtiesrulin

gthecommun

e.Stan

dard

Con

trolsinclud

eaverageda

ilyrainfallforJa

nuary1984

toSeptem

ber1990

andaverageda

ilyrainfallforOctob

er1990

toMarch

1994.Allregression

sarerunusingweigh

tedleastsqua

res(W

LS)

estimationwithpo

pulation

size

asweigh

ts.T

here

are14

2commun

esin

thesample.Stan

dard

errors

areclusteredat

thecommun

elevel.*significan

tat

10pe

rcent,**sign

ificant

at5pe

rcent,

***significan

tat

1pe

rcent.


Table

A.6:C

hannels-Interaction

Effects

Dependent

variable:%

Militiam

en,p.H.

LocalHutu

LeadersLocalT

utsiLeaders

(1)(2)

(3)(4)

(5)(6)

(7)(8)

#Sat(R

ainfall>10m

m)

−0.185 ∗∗

−0.105 ∗∗∗

−0.045

−0.089

−0.281

0.0760.011

−0.958

[0.079][0.029]

[0.038][0.083]

[1.055][0.179]

[0.406][1.479]

#Sat(R

ainfall>10m

m)x

Radio

Own-

ership0.292

0.1071.646

3.081

[0.216][0.265]

[3.661][4.640]

#Sat(R

ainfall>10m

m)xPopulation

Density

0.034 ∗∗∗0.040 ∗∗∗

0.326 ∗∗0.307 ∗∗

[0.006][0.008]

[0.114][0.123]

#Sat(R

ainfall>10m

m)x

Tutsi

Mi-

norityShare

−0.447

−0.647

1.0930.664

[0.535][0.606]

[1.870][1.605]

StandardControls

yesyes

yesyes

yesyes

yesyes

Other

Weekday

Controls

yesyes

yesyes

yesyes

yesyes

Com

mune

Effects

yesyes

yesyes

yesyes

yesyes

R2

0.370.37

0.370.37

0.370.39

0.370.40

N1272

12721272

1272161

161161

161

Note:

#of

Sat(Rainfall>

10mm)isthe

number

ofSaturdays

with

rainfallabove

10mm

duringthe

periodOctober

1990to

March

1994.%

Militiam

enper

Hutu

(p.H)

ismeasured

inpercent.

Radio

Ownership

isthe

fractionof

Hutu

owning

aradio.

Tutsi

Minority

Shareis

thefraction

ofTutsi

dividedby

thefraction

ofHutu.

Incolum

ns1to

4,the

sample

isrestricted

tovillages

with

pro-genocideHutu

partiesruling

thecom

mune.

Incolum

ns5to

8,the

sample

isrestricted

tovillages

with

anti-genocidepro-T

utsiparties

rulingthe

commune.

StandardControls

includeaverage

dailyrainfall

forJanuary

1984to

September

1990and

averagedaily

rainfallfor

October

1990to

March

1994.Other

Weekday

Controls

includethe

number

ofSun/M

on/Tue/W

ed/Thu/Fri

with

rainfallabove

10mm

duringthe

periodOctober

1990to

March

1994.Allregressions

arerun

usingweighted

leastsquares

(WLS)

estimation

with

populationsize

asweights.

There

are142

communes

inthe

sample.

Standarderrors

areclustered

atthe

commune

level.*significant

at10

percent,**significant

at5percent,

***significantat

1percent.

Chapter 4

Selection into Borrowing: Survey

Evidence from Uganda∗

4.1 Introduction

Business growth is tightly linked to credit access. Credit rationing caused by infor-

mation asymmetries, weak legal institutions and high transaction costs are, however,

believed to restrain developing credit markets (Amendariz de Aghion and Morduch,

2005). Microfinance emerged in Asia in the 1970’s as a solution to this challenge, and

has attracted much attention and praise during the past decades.1 Much of the opti-

mism around Microfinance stems from the fact that repayment rates were surprisingly

high - lending to the poor provided to be a profitable activity and this held the promise

that credit rationing would decrease with economic growth as a result. Recent studies

have however found the impact of microfinance initiatives to be limited. First of all,

∗I would like to thank the staff at BRAC Uganda Research and Evaluation Unit, in particularMunshi Sulaiman and Paul Sparks, for their collaboration and practical help during field work, andthe Uganda Small Scale Industries Association for helpful input during the survey design process.The paper has benefited from helpful comments from Selim Gulesci, Andreas Madestam, Anna Sand-berg and Jakob Svensson as well as from seminar participants at the IIES and the Stockholm andUppsala Development Group (DSG). Financial support from Handelsbanken’s Research Foundationis gratefully acknowledged.

1Among the pioneers of Microfinance are Grameen Bank of Bangladesh, founded in 1976, that to-gether with its founder Muhammad Yunus received the Nobel Peace price in 2006; and the BangladeshiNGO BRAC that started Microfinance activities in 1974 (BRAC, 2016).

115

116 SELECTION INTO BORROWING

despite large efforts to reach target clients, take-up rates remain low (Banerjee 2013;

Banerjee et al. 2015a). Moreover there seems to be little impact on the likelihood

of business startup and the income and welfare of borrower-households (Banerjee et

al. 2015a). Some experimental studies suggest that take-up and effectiveness of mi-

crofinance may improve if contractual terms are changed (Field et al., 2013; Karlan

and Zinman, 2008). However, the experimental work to date concerns existing clients;

i.e. clients who select into borrowing under prevailing contract terms, and therefore

can not shed light on how contractual changes would affect credit demand through

composition of the borrower pool among micro enterprises.2 Taking selection effects

into account may also lead to different conclusions regarding investment behavior and

success. Lenders may be reluctant to change loan terms out of fear of attracting riskier

clients that have a lower probability of repaying the loan. Therefore, understanding

how selection into borrowing is affected by contract terms is a first order concern when

seeking to understand how loans can be made more effective.

This study offers a first insight into selection effects due to changes in the standard

loan contract structure. I examine loan demand in a representative sample of urban

micro enterprises in Uganda - most without credit experience. Using hypothetical loan

demand questions, I explore how firm owners’ stated interest in taking a loan changes

when contract terms are altered from those of a standard loan contract. Specifically, I

vary the interest rate and the collateral requirement; two aspects of loan contracts that

play a crucial role for selection into borrowing in theoretical models of credit contract

structure and credit rationing (Stiglitz and Weiss, 1981; Wette, 1983; De Meza and

Webb, 1987). With my representative sample of firms, I am able to draw conclusions

about firm owners who normally would not borrow.

The sample, 925 micro and small firms, was drawn from a firm census conducted

specifically for the study in the Kampala metropolitan area in Uganda, and contains

both manufacturing and retail firms.3 The survey collected comprehensive information2In particular, Field et al. (2013) randomly offer a grace period to clients once loan groups are

formed and loans already approved; while Karlan and Zinman (2008) induce random variation in theloan maturity for previous borrowers.

3The census firms are representative for some of the most prevalent business sectors in urban and


about the business activities along with measures of the firm owners risk-attitudes

and measures of the riskiness of the business environment in which the firm operates.

Moreover, respondents were presented a number of loan contracts that varied in core

features such as the interest rate, the collateral level and the flexibility of repayment. I

elicit loan demand based on the responses to these hypothetical questions. Specifically,

I focus on the interaction of loan demand with individual risk aversion, since the risk

aversion/riskiness of the borrower is a central aspect of loan contract theory due to its

relation to the likelihood of making successful investments.

I find that around 14 percent of the surveyed firms in my sample express an interest

in a "standard" loan similar to the business loans currently offered by Microfinance

institutions and NGOs in Uganda. This figure is close to actual take up rates in other

studies of microfinance: Crépon et al. (2015) observe a take-up rate of 16% in North

Africa while Banerjee et al. (2015b) find a rate of 19% in India. It is also similar to

actual borrowing experience of my respondents.4 Interestingly I find hypothetical loan

take up to be highly sensitive to changes in the contract. Moreover, the propensity to

select into borrowing as contract terms are changed from the standard ones depends

on individual risk aversion of the firm owner and the volatility in demand ("riskiness")

of her business environment. Specifically, while firm owners that are not risk averse

become 8-10 percentage points more interested in a loan if the interest rate is lowered

from 25% to 20% annually, firm owners with a risk aversion score above the median

are 16-18 percentage points more likely to start borrowing following this change, i.e.

about twice the effect size. Owners that face a less risky business environment in terms

of unpredictability and fluctuation in sales display a similar pattern: they are 15 per-

centage point more likely to start borrowing following a lower interest rate compared

to an 8 percentage point increase in demand among those with a more risky business

semi-urban areas of Uganda. According to the 2010/2011 Business Registry, welding and carpentryare two of the 3 single largest groups among manufacture in the country, together accounting for 30%of the manufacturing sector. According to the same report motor repair and the retail sectors I focuson are also among the most prevalent in the country, with wholesale of food and beverages being thesingle largest retail sector (Uganda Bureau of Statistics 2011).

4In my sample, 20.6% of respondents had experience of borrowing from semi-formal or formalsources while 9.7% had borrowed in the 2 years preceding the survey.


environment. If the collateral requirement is reduced from 100% to 50% of the loan

value, less risky and less wealthy firm owners start to borrow. Analyzing the results for

manufacturing firms and retail firms separately, I find that fluctuations in the business

environment affects take-up of both contracts substantially for manufacturing firms

while for retail sectors I do not observe differentially higher loan demand among firms

facing more fluctuations than among those with less volatile business environment.

Risk attitudes also affects take-up of the low interest rate-contract more for manufac-

turing firms than for retail firms. The results are robust to using alternative measures

of risk-aversion and risky environment, such as financial vulnerability and absence of

precautionary savings, also when controlling for wealth level of the firm owner, which

differs between sectors. Taken together, these findings suggest that repayment behav-

ior may improve as a result of changing these contractual details, given that loans are

also extended to new prospective clients. To address the concern that hypothetical

questions may lead to overestimation in the willingness to accept a contract, I con-

trol for firm owner specific effects to reduce possible measurement error. I also run

a series of validation tests to make sure that the hypothetical questions are properly

understood by my respondents.

This paper contributes to several strands of the literature. Firstly, it provides a

test of some of the central theoretical results in credit contract theory that seek to

explain the prevalence of credit rationing. (Stiglitz and Weiss, 1981; Wette, 1983; De

Meza and Webb, 1987). Empirical tests of these models are complicated by the fact

that one normally only observes borrowing patterns among those who have selected

in to the credit market. With my representative random sample of firms I am able to

observe firm owners who normally would not borrow and therefore I can test whether

credit markets are characterized by adverse or advantageous selection in my setting.

The paper also adds to the growing literature on credit access and use in developing

country contexts, including evaluations of Microfinance and other forms of semi-formal

credit. A handful of recently published studies provide the first larger scale randomized

evaluations of microfinance initiatives (Attanasio et al., 2015; Angelucci et al., 2015;


Augsburg et al., 2015; Banerjee et al. 2015b; Crépon et al., 2015; Tarozzi et al., 2015).

Taken together, these studies do not find evidence of transformative effects of microfi-

nance on the lives of the poor, nor do they find positive effects on the extensive margin

of business ownership (startups). They do however find modestly positive effects on

business outcomes for already existing micro-businesses (Banerjee et al., 2015a). At

the same time, none of the studies find significant increases in household income or

consumption following this business growth. This is explained by microfinance merely

offering more freedom in decision making about occupational choice within the house-

hold, leading to a reshuffling of resources and labor towards the micro-enterprises

at the expense of other activities (ibid.). However, none of these studies is able to

deal with selection effects: they study households or businesses that have selected into

borrowing.

Within this literature, I more specifically contribute to the limited work on the role

of loan contract structure for the profitability and use of loans. In developing country

contexts I am aware of two recent empirical studies of how changes in the contract

terms affect loan demand and loan use (Field et al., 2013; Karlan and Zinman, 2008).

Just as the abovementioned work, these two studies are also restricted to exploring

intensive margin demand and loan use among existing or former borrowers. Mine is not

an experimental study and the results are therefore more suggestive in nature. Because

we know little about selection into borrowing and the loan attitudes of non-borrowers,

my results are nevertheless interesting.

By focusing on businesses rather than households, this study also contributes to

the literature on small business growth. In developing countries both in Africa and

Asia there is a large group of small businesses (both formal and informal) while very

few businesses grow beyond medium size. Recent studies have tried to understand the

determinants of and obstacles to business growth in developing countries by offering

cash grants (de Mel et al., 2008; Fafchamps et al., 2014), business training (Karlan

and Valdivia, 2011) and combinations of the two (Bandiera et al., 2016; Berge et al.,

2014; Fiala, 2013). The comprehensive survey data that I collect on firm characteristics


related to the firms’ production function, and the fact that I also have detailed data

on owner households, allows me to study the interaction of the production function

and the loan contract, and also to capture heterogenous effects, which have been

shown to matter in related work: Fiala (2013) finds small positive effects of cash

grants on business growth in northern Uganda, but only among high ability males.

de Mel et al. (2009) find that the propensity to innovate is strongly linked to firm

owners’ individual characteristics. Compared to previous studies my sample consists

to a larger share of manufacturing firms as opposed to retailers. Manufacturing firms

have different investment possibilities than retailers and my sample therefore allows for

investigating new aspects of investment in labor and capital compared to the previous

related literature on small business growth.

The remainder of the paper is structured as follows. In the next section, I provide

some institutional background regarding the geographic context and the credit prod-

ucts relevant in my studied setting. In section 3, I outline my hypotheses. Section 4

describes the sampling and survey methodology and the survey data I collected. Sec-

tion 5 presents and discusses the results along with the empirical specifications and

section 6 presents validation checks. Section 7 concludes the paper.

4.2 Institutional Background

In this section, I briefly describe the background of the study in terms of the market

in which the small firms are active and the type of credit available in this setting. This

lays the foundation for my hypotheses about selection connected to amendments to

the interest rate or collateral requirements of a credit contract.

4.2.1 Labor market and small businesses in Sub Saharan Africa

Unemployment and lack of formal sector jobs are significant problems in Sub-Saharan

Africa. Setting up a small business is a common livelihood strategy in urban and semi

urban regions. In 2013, it was estimated that 90% of Uganda’s private sector consisted

4.2. INSTITUTIONAL BACKGROUND 121

of micro, small or medium sized businesses (MSMEs)5 and around 28% of the labor

force was active in such businesses (World Bank, 2013).6

4.2.2 Standard loan contract

Previous studies have shown that the market for lending to the poor suffers from mar-

ket imperfections and is characterized by credit rationing (See e.g. Banerjee (2001)).

The type of credit available to businesses in my setting is most commonly offered

by Microfinance institutions including both NGOs and commercial credit institutions.

Contracts offered by this type of lenders typically display the following features:

1. Constant repayments starting early. This is a typical feature of microfinance

contracts, believed to decrease transaction and monitoring costs of the lender

while fostering a habit to make repayments on time.

2. Limited loan size Due to asymmetric information issues, the loan size is limited

both by the collateral held by the borrower and by the relation between borrower

and lender. Often repeat loans (with the same lender) are allowed to be larger.

3. High interest rates Studies in other developing country contexts report typical

rates around 20-25% APR. In the two studies most closely related to the current

one, interest rates (APR) are 22% (Field et al., 2013) and 200% (Karlan and

Zinman, 2008). In Uganda at the time of this study, annual interest rates offered

by large MFIs were also in the 20-25% interval.7

4. Collateral requirements, often in the form of land titles.8

5Uganda Investment Authority (2013).6According to the World bank’s Uganda Economic Update, 14% pf Uganda’s workforce were

employed in the non agriculture informal enterprise sector in 2013, while another 14% of the workforcewas employed in formal business, most of which was in the private sector. Both of these categorieswould correspond to micro or small enterprises.

7For example, one of the biggest MFIs in Uganda at the time of the study; PRIDE Microfinancein Uganda offered loans with 26% APR and demanding a 100% collateral (Fiala (2013). A "mysteryshopper" investigation carried out by the research team at the time of my data collection revealedsimilar conditions among other prominent MFIs in the Kampala area.

8While not a typical feature of the classical microcredit contract, that instead sometimes relies onthe "social collateral", business loans in Uganda, typically do require collateral. Such loans are larger


Prevailing contract terms offered by Micro finance institutions are standardized in

order to simplify operations for both loan officers and borrowers, and make it possible

for lenders to extend borrowing activities to a large number of clients. While designed

to minimize risk and lower the monitoring cost of the lenders (Banerjee, 2001) these

contract terms may however also tilt investment towards investments that are smaller

and involve less learning than optimal. When designing business loans, as opposed to

individual micro loans, it may therefore be of particular interest to test the optimality

of the prevailing loan terms.

4.3 Hypotheses

In this section, I list the hypotheses tested in this paper. The first set of hypotheses

are about firm owner responses to lowering the interest rate on business loans. I in-

vestigate whether a lower interest rate leads to adverse or advantageous selection into

the borrower pool. In the theoretical literature, Stiglitz and Weiss (1981) showed that

credit rationing can prevail in equilibrium since increasing the interest rate is associ-

ated with a borrower pool more dominated by risky borrowers/projects (with a high

risk of failure). In their model, when the interest rate is raised beyond a certain level,

extending credit to riskier borrowers by further increasing the interest rate is associ-

ated with a decrease in profits for the lender. This results in an equilibrium with excess

demand of credit: risky borrowers are willing to pay higher interest rates but lenders

are not willing to extend such loans. Conversely, De Meza and Webb, (1987) show that

under different assumptions about the distribution of project returns, lower interest

is associated with excess investment in risky projects. In their model, increasing the

interest rate can be a way to curb the over investment that would occur under a lower

interest rate. I do not have a direct measure of the project that borrowers invest in

and I therefore focus on the riskiness of borrowers, proxied by their self reported risk

and often have a longer time horizon than loans that can be accessed through microcredit groups.

4.3. HYPOTHESES 123

aversion and the volatility of their business environment.9

H01: Firm owners that are more risk averse or operate in a less risky environment

are more likely to select into borrowing when the interest rate is lowered (Stiglitz and

Weiss, 1981).

HA1: Firm owners that are more risk averse or operate in a less risky environment are

less likely to select into borrowing when the interest rate is lowered (De Meza and

Webb, 1987).

The next hypotheses are about firm owner responses to lowering the required collat-

eral. Stiglitz and Weiss (1981) show that just as increases in the interest rate attracts

riskier projects, so does increases in the collateral, keeping interest rates constant, if

borrowers are risk averse. Wette (1983) shows that the latter result of Stiglitz and

Weiss holds also for risk neutral borrowers. Keeping the interest fixed and instead

raising the collateral can lead to similar adverse selection into borrowing, lowering the

expected return of the lender. Boucher et al. (2008) more directly discuss how credit

rationing through the collateral channel is affected by borrower wealth. They distin-

guish between two types of credit rationing that collateral requirements may entail in

developing country credit markets: quantity rationing where the inability to provide

collateral excludes poorer borrowers from credit markets, and risk rationing, where

borrowers with collateral refrain from borrowing due to the risk of losing the collat-

eral. Contrary to quantity rationing - which unambiguously excludes poorer potential

borrowers from credit, risk rationing may prevent wealthier people from borrowing if

the wealth is in liquid assets, while results are ambiguous if wealth is in the form of

land.

9As in Stiglitz and Weiss (1981) and De Meza and Webb (1987) I vary the interest rate whilekeeping the collateral fixed. An equilibrium in which no credit rationing prevails is characterized byBester (1985) whose model allows both interest and collateral to change at the same time. Such amodel is less relevant for my microfinance setting where credit rationing is a stylized fact and therigidity of credit contracts prohibit lenders from tailoring loan agreements to specific borrowers.


H02: Firm owners that are more risk averse or operate in a less risky environment are

more likely to select into borrowing when the collateral requirement is lowered (Stiglitz

and Weiss, 1981; Wette, 1983).

HA2: Firm owners that are more risk averse or operate in a less risky environment are

less likely to select into borrowing when the collateral requirement is lowered.

In addition to testing these hypotheses, I also examine if and how the responses to

changes in the loan contract affect firm owners in different types of sectors differently.

In particular, I estimate results separately for retain sectors and for manufacturing

sectors. The success probability of a project is tied to the type of investment implied

by the project. Risk attitudes and riskiness of the business environment is likely to

be of less importance for certain types of investments (e.g. in already familiar tech-

nologies) than for others. Since investment options are very different between retail

and manufacturing firms, this sectoral distinction is likely to affect loan demand and

borrower behavior.

4.4 Survey methodology and Data

The survey data was collected in the Kampala metropolitan area in March and April

2013. Fieldwork was carried out in collaboration with the Research and Evaluation

Unit of the NGO BRAC Uganda, managed and supervised by the author together

with local research officers and carried out by a team of enumerators (interviewers)

recruited specifically for the project. The businesses surveyed are a random sample

drawn from a larger pool of businesses whose contact details were collected in a census

preceding the survey. Details about the sampling and data collection are described

below.

4.4. SURVEY METHODOLOGY AND DATA 125

4.4.1 Census and Sample selection

The census was conducted in January and February 2013. Firms were chosen on the

basis of their business sector and geographic location. Business sectors were selected

as to represent the main sectors in urban and semi-urban Uganda within both man-

ufacture and retail, and to ensure that both female and male business owners would

be represented. The specific sectors included are broadly grouped into retail sectors

and manufacturing sectors, with the former category including supermarkets, smaller

food retail shops, food and beverages wholesale and hardware shops, and the latter

category including carpentry, welding/metal works and motor repair workshops (both

of cars and of motorcycles). The enumerators were instructed to approach all firms

in selected business sectors, with some restrictions on the size and type of business

structure. The lower bound set on firm size depended on the sector: to be included

in the census; manufacturing firms (including motor repair) were required to have at

least 1 employee (formal or informal) in addition to the owner, while firms in retail

were required to have a permanent business location and a well-stocked shop. The

upper bound was set at 15 employees (formal or informal) regardless of the business

sector. The sample thus consists of larger and more established firms than the micro

businesses and households that are the focus of most related studies, and has a heav-

ier emphasis on manufacturing sectors. According to the Ugandan Business registry

2010-11, 98% of all businesses in the country had less than 10 employees and thus

were classified as micro, small or medium sized businesses, and 87% of the workers in

the private sector were working in a business with less than 50 employees (Uganda

Bureau of Statistics 2011).10 I study loan attitudes among owners of micro- and small

10The official definition of Micro, small, medium and large businesses in Uganda is the following:Micro businesses were those with an annual turnover of less than 5 million shillings irrespective ofthe number of employees, while small businesses were those with an annual turnover of between 5and 10 million shillings irrespective of the number of employees. Medium businesses on the otherhand were those with an annual turnover of more than 10 million shillings but employing less than50 persons while the large businesses were those with an annual turnover of more than 10 millionshillings and employing at least 50 persons (Uganda Bureau of Statistics 2011). While definitionsdiffer substantially between countries, an international standard definition has been created by theInternational Labor Organization. This definition states that a micro business is an enterprise withup to ten employees, while small enterprises are those that have 10-100 employees, and medium-sized


businesses in sectors that make up the bulk of the urban private sector in Uganda.11

1,353 businesses were interviewed for the census. Importantly, most of them had

no previous loan experience. The enumerators approached the businesses with a script

saying that they were part of a research project by researchers at Stockholm University

(in Europe), about business growth in "enterprises like yours" and "learning about the

difficulties and opportunities for growth of firms in your sector", and that the data

would be treated with anonymity. Since BRAC is well known as a Microfinance insti-

tution, in order not to prompt respondents to think about loans, potentially deterring

loan averse individuals from taking part in the survey the name of "BRAC" was not

mentioned to respondents.

Based on this census listing a random sample was drawn, stratified by business sec-

tor, of 985 businesses to be interviewed in the main survey .12 Female owned businesses

were over-sampled. The response rate was 94% resulting in 925 businesses participating

in the survey.

4.4.2 Survey data

The survey provided detailed information on firms’ input and investment choices and

their demand for credit under different hypothetical loan contracts. By examining the

choices made by firm owners, the survey allows me to explore how possible take up of

credit may be affected by changes in the cost of lending and collateral levels.

Specifically; the survey was designed so that the hypothetical contracts presented

to the respondents reveal the effect of credit constraints caused by interest rate and

enterprises have 100 to 250 employees (International Labor Organization 2015).11According to the 2010/2011 Business Registry published by the Uganda Bureau of statistics

(UBOS), welding and carpentry are among the 3 single largest groups among manufacture in thecountry and together account for 30% of the manufacturing sector. According to the same reportmotor repair and the retail sectors I focus on are also among the most prevalent in the country, withwholesale of food and beverages being the single largest retail sector (Uganda Bureau of Statistics2011).

12Based on statistics from the census it was decided not to stratify according to geography orborrowing experience since the distribution of sectors was similar across geographical locations andsince loan experience was very limited in all sectors.


collateral requirements.13 In addition, the survey includes modules on firms’ employ-

ees, assets, costs and revenues, seasonality of sales, vulnerability to shocks, credit

history, types of interactions with other businesses, the business-owners’ background,

education, financial literacy, risk attitudes and his/her household’s demographics. A

few sections of the survey require additional explanation, as they are central for my

analysis. These are described in more detail below.

4.4.2.1 Measures of loan demand

To learn more about selection into borrowing, and to investigate which firm charac-

teristics are particularly relevant for loan demand, I included a module eliciting hy-

pothetical loan demand. This section began by offering firms a generic contract with

terms and amounts similar to those of the contract offered by BRAC Small Enterprises

Program and other MFIs in urban Uganda, and then went on to present additional

contracts that amended the contractual aspects to measure the relevance of various

constraints. The benefit of this approach is twofold. First, by exploring within-subject

responses, I partially circumvent the critique that respondents tend to over- or under-

state hypothetical demand compared to the real willingness to accept a contract (See

e.g. Neill et al. (1994)). Also, by presenting each contract unconditionally of the stan-

dard one, the survey overcomes the problem of firms self-selecting on the generic loan,

which could potentially bias the responses to the remaining contracts. In my case, all

925 firms were offered the option of the generic and the perturbed contracts. Loan

officers at BRAC provided input about the phrasing of the contract description for

the hypothetical contracts, to make sure that the loan contracts would be adequately

explained to respondents with varying degrees of loan experience and financial literacy.

The benchmark, "standard" contract is described as follows:

"Imagine you were offered the opportunity to take a loan. If you decide to take this loan,

you can borrow up to 8 million Shillings. You would need to repay this amount plus a 25%

13The survey also included contract amendments designed to study constraints such as uncertainand back-loaded return paths as well as the constraints large fixed costs may pose.


interest within one year. The repayments have to be done in equal monthly repayment install-

ments over the year. [Here, the enumerator was urged to show an example to the respondent].

The lender requests security (collateral) in the form of land. That is, in order to borrow a

certain amount, for example, 3 million14, you need to have formal property rights to land

valued to 3 million and in case you fail to repay, the lender will claim the 3 million in terms

of your land."

The respondent is then asked if they would take such an offer, how much they

would borrow and for what main use. Thereafter, other contracts are described to the

respondent. Here my focus is on the following contract variations15:

• Low interest rate contract: the APR is lowered from 25% to 20%.

• Low collateral contract: The collateral requirement is lowered from amounting

to 100% of the loan size value to only 50% of the value. The collateral is still in

the form of land.

The difference between the standard contract and each amended contract was made

salient and an example was used to show the repayment structure and the size of each

installment with the low interest rate contract, and to show the size of the collateral

with the low collateral contract. Thereafter the respondent is asked if they would take

a loan under those contract terms. The exact wording of the contract variations can

be seen in the Appendix.

4.4.2.2 Measures of risk attitudes and riskiness

The literature has emphasized risk attitudes of the borrower as an important factor

affecting credit demand and investment behavior. Moreover, the riskiness of the project

that the loan is used to finance is also a central component in the decision of the

lender of whether or not to approve a loan, and in the determination of interest rate14Using the 2013World Bank PPP adjusted exchange rate for Uganda (1,014 UGX/USD), 3,000,000

corresponds to 2960 USD. Using the nominal exchange rate of April 1, 2013 (2585 UGX/USD),3,000,000 UGX corresponds to 1161 USD.

15In the validation checks section where I examine the quality of the hypothetical loan contracts Ialso discuss an additional contract amendment in which the collateral type was changed.


and collateral levels. I collect data on risk attitudes. Since I do not observe actual

borrowers, I do not have a perfect proxy for the riskiness of projects. Instead I construct

measures of the risk aversion of the owner and the riskiness of the firm owner’s business

environment. These variables are explained in the next paragraphs.

Risk attitudes: as my measure of risk aversion, I use a survey question where the

respondent was asked to make a judgement of their own willingness to take risks. More

specifically, I ask them to place themselves on a 0−10 scale between "Not at all willing

to take risks" and "Very willing to take risks". I define "risk aversion" as a dummy

variable taking the value 1 if the respondent is below median on this self-reported risk

taking scale, and 0 otherwise. This measure is taken from the German Socio-Economic

Panel and validated by Dohmen et al. (2011) to be predictive of financial risk. Unlike

other commonly used methods of eliciting risk preferences it involves no computations

and should therefore be appropriate for my setting of less educated respondents.16

Riskiness: In addition to a measure of risk attitudes I am also interested in a measure

of the riskiness of the enterprises’ business environment and activities. To capture this,

I construct two indices based on the responses to a list of statements about possible

reasons why loan repayment may be hard, where the respondent indicates to what

extent they agree with each statement. The measures obtained here are thus directly

related to the business practices and environment of the enterprise. The first index

that I construct is the riskiness index which is higher if the respondent agrees that

fluctuations and uncertainty are important constraints for repaying a loan.17 To ensure

16The survey also included a section with a lottery-type elicitation of risk attitudes. Here therespondents answered five question in which they choose between (a) a sure amount and (b) a coinflip between 0 and a gradually increasing amount. The seminal work for measuring risk attitudes usinga list of choices the Binswanger (1981) protocol where the probabilities of the lottery are varied. Ichose not to use it because the concept of probabilities would be too difficult to explain to my poolof less educated respondents. Our questions are instead an adaptation of the protocol used in Holtand Laury (2002), with simplified probabilities, as in Abdellaoui et al. (2011) and Bosch-Domènechand Silvestre (2013). Related studies in developing country contexts have also used similar methodto measure risk preferences (see e.g. de Mel et al. (2008)). Despite these simplification of the list ofchoices, responses turned out to vary with the interviewer. For this reason the 11 point scale (forwhich answers were not correlated with interviewer) provides the basis of the risk averseness measureused in my final analysis.

17More specifically, the respondent was asked to respond on a 4-point scale between strongly agreeand strongly disagree to the following statements: (1) It is difficult to make loan repayments on time


that I do not just capture a general unwillingness to borrow, and that my obtained

measure is not correlated with the confidence of the respondent (some respondents may

be more inclined to saying that they agree with all the statements because they find

all aspects of repayment equally challenging) I also construct a placebo index based on

responses to three other statements about difficulties with repaying loans; constraints

that are not associated with fluctuation and uncertainty.18 I will also examine three

alternative measures of a risky business environment or behavior, explained in the

next subsection.

4.4.2.3 Alternative ways of measuring risk behavior

In addition to the measures of risk aversion and riskiness described in the previous

subsection, in my robustness tests, I use three alternative ways of proxying for risk.

These are (i) financial vulnerability, (ii) having precautionary savings, and (iii) the

habit of selling on credit. Financial vulnerability is assessed using survey questions

about the respondent’s ability to access a certain amount of money for an emergency. I

asked the questions for 2 amounts: 500,000 Ugandan shillings (193 USD) and 2 Million

Ugandan shillings (775 USD).19 For each value, I create a dummy that takes the value

of one if the respondent says that she would be able to obtain the amount interest

free (i.e. in other ways that borrowing the amount from an MFI or a moneylender).

High financial vulnerability may be correlated with a higher risk of not being able to

repay a loan. Precautionary savings is a dummy that equals one if the respondents

answers yes to a survey question asking whether the respondent is currently saving

anywhere. Having savings is used as a proxy for being a less risky borrower. Finally,

due to sales fluctuations (2) It is difficult to make loan repayments on time because it is hard topredict when sales will be good or bad. The index is the average score for these two questions.

18Specifically, the statements are: (1) It is difficult to get a loan because it is hard to know whereto get the best terms (2) It is difficult to get large enough loans to make good business investments(3) It is difficult to make loan repayments on time because it takes a while to know how to generateprofits from an investment.

19In real terms, 500,000 UGX corresponds to 493 USD (using the 2013 World Bank PPP adjustedexchange rate for Uganda) while 2,000,000 corresponds to 1973 USD. Using the nominal exchangerate of April 1, 2013, 500,000 UGX corresponds to 193 USD while 2,000,000 UGX corresponds to 774USD.


selling on credit is a dummy that equals one if the respondents reports to regularly

sell to customers on credit. This is used as a proxy for risky behavior.

4.4.3 Using hypothetical questions

I measure loan demand using a series of hypothetical questions. The choice of using hy-

pothetical questions is motivated by two factors. First, to understand selection effects

in relation to changing credit contract terms, interviewing a representative sample of

businesses - both borrowers and non borrowers - is of crucial importance. I interview a

representative sample of businesses and cover sectors that are not currently the main

target of MFIs and other semi-formal loan-providers.20 While extending credit to these

businesses is a goal of any MFI, doing so requires learning more about their loan de-

mand and loan use. Hypothetical questions is a first step in building this knowledge.

Second, given the wide-spread reluctance to loan take-up and to MFIs among many

business owners in my setting21, the hypothetical question setup provides a way to

approach business owners who might otherwise refuse to participate. Since this study

aims to address the low efficiency of loans, some of which may be explained by the

low take-up rate of businesses with certain preferences or technologies; I did not want

to deter businesses that were reluctant to think about loans from participating in the

study. Presenting the loan attitude questions as purely hypothetical alleviated the risk

of this happening. For similar reasons, I presented the study as a research project from

Stockholm university rather than from BRAC, which was my partner in carrying out

the fieldwork, since BRAC may be known for its micro finance activities and, again,

prevent those business owners reluctant to loans from participating.

Hypothetical questions are, however, often associated with concerns about misre-

porting and bias (See e.g. Neill et al., 1994). For example, certain respondents may

overestimate their demand for goods while others give estimates that are lower than20These are sectors in manufacturing and services: Carpenters, welders and motor repair workshops

for cars and motorcycles.21Several business owners told me during the piloting that they distrust MFIs since these will trick

people and steal their land, and 71% of the respondents report that they distrust NGO’s/developmentorganizations.


their actual demand. Replies could also be affected by the timing and circumstances

of the interview or by the interaction between the respondent and the interviewer. In

other words, these concerns mainly regard individual- and interview occasion-specific

unobservables that complicate the interpretation of the valuations and would be less

serious if focusing on within subject variation, since the level of misreporting is corre-

lated across responses from the same individual (List and Shogren, 2002). Thus, my

main specification is a within-subject specification where I include firm fixed effects.22

In an alternative specification without firm fixed effects, I include controls for the

interview occasion and interviewer.

4.4.3.1 Sector specific firm performance data

The relation between loan take-up and riskiness is likely to be affected by important

factors within the firms’ production function. Investment choices as well as the types

of risk that affect the success of an investment depends on the sector in which the firm

is active. I collect data that allows me to describe retail and manufacturing sectors

and examine differences between them. I include modules on firms’ employees, assets,

costs and revenues. To obtain a reliable measure of business assets and output, survey

construction required detailed knowledge about the equipment and production meth-

ods commonly used by manufacturing firms. To gain such insight I relied on official

manuals for vocational training in Uganda (used in the BTVET or Business, Tech-

nical, Vocational Education and Training schools). In addition, short surveys about

tools and machines were carried out among the manufacturing businesses during the

piloting phase.23 I also collect information on whether demand for the products sold

in the sector is sensitive to business cycles or seasonal variation, or to various shocks,22Since only one interview was done with each firm owner this fixed effect also captures the interview

occasion and interviewer.23Research assistants visited 20 firms from each of the sectors interviewed and asked them to list

all the machines and tools that they owned or used in their business. This data was then aggregatedand allowed me to identify the most important tools and machines in each sector. I also held twomeetings with experts in a local trade association for carpenters, welders, and motor repair shopsin Uganda (Uganda Small Scale Industries Association, USSIA) to confirm the list of machines andtools and to gain further insight into the manner in which firms of different size have access to suchtechnology.


and about the access to employees and the mode of employing them. Moreover, I col-

lected data on vulnerability to shocks, credit history, business networks and types of

interactions with other businesses.

4.4.4 Summary Statistics

Columns 1-3 of Table 4.1 shows the summary statistics on a number of important

variables for the entire sample of firms. The average firm in my sample is 6.7 years

old and 31% of firms are in manufacturing sectors while the remaining are in retail

sectors. The average firm size is 2.7 workers, including the owner, and the average re-

ported asset value corresponds to about 850 USD24 while average reported stock-value

corresponds to around 5000 USD. Many of the firms can thus be classified as micro

or small enterprises.25 The average level of education is 11.5 years which corresponds

to finished secondary school (O-level). Around 20% of owners have ever taken a loan,

and only 11% have taken a loan in the past 2 years. This is similar to shares found

for loan take-up in related studies of the effectiveness of microfinance, where take-up

rates are observed after microfinance was introduced in a geographical area. About

44% of the firm owners in the sample are landowners, which means they have access

to collateral for a loan.

Compared to the related literature, a few differences are particularly noteworthy.

First, the businesses in my sample are larger on average than those in most related

studies, where the focus has been on household enterprises with no employees. More-

over, there is a higher share of manufacturing businesses in my sample. This may

affect borrowing capacity and demand differently, something I discuss in more detail

in the next paragraph. In addition, the business owners in my sample are much less

used to borrowing than those observed in related studies of small business growth and

24Average reported asset value is 2.2 Million UGX. In real terms, this corresponds to 2170 USD(using the World bank PPP adjusted exchange rate for 2013) and in nominal terms to 845 USD (April1, 2013).

25According to a definition used by Uganda Bureau of Statistics, a micro-enterprise is one in whichthe annual turnover does not exceed 10 Million UGX and number of workers is up to 5, while a smallbusiness employs between 5 and 49 people and total assets between UGX: 10 million-100 Million.


microfinance, mainly due to the fact that they are not sampled from among a set of

previous borrowers or in collaboration with a Microfinance institution as in e.g. Field

et al. (2013), Karlan and Zinman (2008), Fiala (2013), Berge et al. (2014), Karlan and

Valdivia (2011) and Valdivia (2013). Although the differences in borrowing experience

between these and my study could to some extent be explained by differences between

geographical regions; even compared to the studies carried out in East Africa (Fiala,

2013; Berge et al., 2014), the share with loan experience in my sample is considerably

lower.

Columns 4 and 5 of Table 4.1 breaks down the sample by whether a firm is classified

as retail or manufacturing. Given the differences in production function between these

broad categories they are likely to have different investment needs which may affect

their credit demand. From the last 2 columns of the table, it is clear that manufactur-

ing and retail businesses differ on many characteristics. Manufacturing businesses in

my sample are on average 1.4 year older than retail businesses and have more employ-

ees: the average number of workers is 4.12 in manufacturing firms, compared to just

2.14 in retail firms, and 93% of the manufacturing firms have at least one employee

compared to only 46% of retail firms. Both reported profits, asset value and stock value

are significantly lower in manufacture firms than in retail firms. Due to the choice of

sectors in manufacture, the share of owners who are female is much lower than in

retail.26 Owners in manufacturing sectors also have on average 1 year less education

than owners in retail. Retailers are more likely to be banked and to keep books than

owners of manufacturing firms, while the latter - somewhat surprisingly - score higher

on financial literacy questions.27 Retailers are more likely to sell to customers on credit

while manufacture firm owners score higher on the "risk index" that is constructed

out of questions that measure riskiness of the business environment described above.

26Manufacturing sectors with a high share of female workers include tailoring and some types offood processing.

27To measure financial literacy I used the 3 simplest questions from the financial literacy sectionof the American Life Panel survey: Numeracy, Inflation and Money illusion, with amounts adaptedto my context. The questions are designed to measure the firm owner’s numeracy and understandingof the concepts of interest rate and inflation.

4.5. RESULTS 135

There is, however, no difference in the self assessed riskiness score. It is noteworthy

that no differences between manufacture and retail owners appear for factors related

directly to loan experience: the share with loan experience is almost identical, as is the

share who have ever delayed on loan repayment or ever been denied the requested loan

size when applying for a loan. Moreover, the share of owners who own land, the most

common type of collateral for loans, is similar between manufacturing and retail sec-

tors. Finally, the fact that manufacturing firms employ more labor is also reflected in

their reported investment demand, where manufacturing firms are considerably more

likely to report that they would want to hire more labor if they could. The manufac-

turing firms also appear more capital constrained: 76% answer in the affirmative to a

question about wanting more capital, compared to only 41% in the retail categories.28

Taken together, firm owners in retail and manufacturing are similar on loan experi-

ence characteristics, as well as on access to collateral, one prerequisite for borrowing

that has been highlighted by the literature. However, manufacturing firms have lower

levels of capital and turnover than retailers and their owners have lower education -

factors that may make them less suited for borrowing. Meanwhile, their demand for

investments, in both labor and capital, is higher than among owners of retail firms.

4.5 Results

4.5.1 Main Results

Table 4.2 shows an overview of the share of respondents expressing interest in the

standard loan contract as well as in each of the contract variations. The contracts

are presented in the order in which they were asked to respondents in the survey

instrument. 14.14% of the respondents reported that they would take a loan if offered

the standard contract. The take up rate of each of the amended contracts is significantly

higher, with 24.67% saying yes to the low interest contract and 27.84% saying yes to28The precise question asked for demand of labor is "If you could find suitable workers, would you

like to employ more workers than you are currently employing?" and for demand of capital: "Is thereanything you would like to buy for this business (e.g. a machine) but cannot do so?".


the low collateral contract.

4.5.1.1 Hypothetical loan demand - standard contract

To examine demand for the standard loan contract, I perform a simple comparison

of a number of key characteristics, between respondents who say no to the standard

contract and those who express interest in the contract. The results are presented

in Table 4.3. Firm owners who want a loan under the standard contract terms own

businesses that are on average one year older, have about 0.6 years less education,

and score higher on the financial literacy score. Not surprisingly they are also more

likely to have borrowing experience and fulfill formal requirements for borrowing: they

are more likely to own land and to have borrowed from formal or semi formal sources

previously. Conditional on loan experience, they are not more or less likely to have

been denied the loan size they applied for in the past. They face a slightly more

risky environment, in terms of unpredictability in sales (the riskiness index) and are

more likely to report wanting to employ more labor in the firm. There is no difference

between manufacturing and retail sectors in the expressed demand for the standard

loan contract.

4.5.1.2 Hypothetical loan demand - contractual changes

Next I analyze how individuals’ loan demand is affected by changes to the standard

credit contract. In particular I focus on amendments to the interest rate and to the

collateral requirements associated with a loan. To estimate how the demand changes

when amendments are made to the standard loan contract, I estimate two models.

The unit of observation is the contract*individual. First I estimate:

Demandic = α + γContractc +βYi +σ [Contractc×Yi]+Xiθ + εic, (4.1)

I also estimate the following within-subject model:

4.5. RESULTS 137

Demandic = α + γContractc +βYi +σ [Contractc×Yi]+ηi + εic, (4.2)

Here, the outcome variable Demandic is either a dummy=1 if individual i states

that she would take the loan, or the log of the loan amount the respondent reports

wanting to apply for. In each regression the standard contract is the baseline category,

and is compared to one other contract: Contractc ∈ {Low interest rate contract, Low

collateral contract}. Y is a characteristic measuring risk aversion or riskiness in business

environment of respondent i, or the wealth quartile of individual i. X in equation 4.1

is a vector of interviewer and interview time-fixed effects, used in the between subject

specification while ηi in equation 4.2 is an individual fixed effect (in the within-subject

specification, the β coefficient will be absorbed by the individual fixed effect). The

coefficients of interest are γ indicating the difference in take up between the amended

contract c and the standard contract for individual with characteristic Y=0, and σ ,

indicating the additional difference in take-up between the standard contract and the

amended contract if characteristic Y=1. Standard errors are clustered at the firm level

for all specifications.

Panel A of Table 4.4 shows the extensive margin estimation for take-up of the

Low interest rate contract compared to the standard loan contract. Columns 1 and 4

show results from the between-subject specification 4.1 while the remaining columns

show the results for the within-subject specification, without and with controls for

individual wealth. Starting with column 1 and 2, the coefficient on the indicator for the

Low interest rate contract (top row) shows that individuals with a high risk business

environment are 9.8 percentage points more likely to take the low interest contract

compared to the standard contract (for which the mean demand in this group is

14.5%). By examining the interaction term, we see that the corresponding difference

in take up for individuals in a low-risk environment (with a low score on the risk

index) is 18.3 percentage points (9.8+8.5). Among individuals who say "no" to the

standard contract, the effect size is thus almost twice as high among low risk firms

than among high risk firms. When individual fixed effects are added in column 2, the


point estimates change only marginally and remain statistically significant at the 90

percent confidence level.29 Columns 4-6 show that introducing a Low interest rate

contract increases demand for borrowing by 6.5-8 percentage points compared to the

standard contract (i.e. from 13.2% to 20-22%) among non risk-averse respondents,

while firm owners categorized as risk averse are an additional 7.7-8.1 percentage points

more likely to start borrowing when offered the low interest contract. There is no

significant correlation between wealth and take-up of the low interest rate contract, all

point estimates for the wealth controls in columns 3 and 6 are small and statistically

insignificant.30

Panel B of Table 4.4 shows the extensive margin demand for the Low collateral

contract. Again, columns 1 and 4 show results from the between-subject specification

while the remaining columns show the results for the within-subject specification, with-

out and with controls for individual wealth, which is likely to be positively correlated

with the ability to put up collateral for a loan. Also for the low collateral contract,

take-up is higher among firm owners in a low risk business environment (with a low

score on the risk index). The top row shows that individuals with a high risk index are

between 14.6-15.7 percentage points more likely to desire the low collateral contract

than the standard contract. Firm owners that have a low risk index are an additional

8.4-11.2 percentage points more likely to demand the low collateral contract. The point

estimate on the interaction term is significant at the 95 percent confidence level in all

specifications and robust to controlling for wealth.31 The results for self reported risk

aversion are weaker: the point estimate on the interaction term between risk aversion

and take up suggests that risk averse people are 5 percentage points more likely to

start borrowing under the low collateral contract as compared to self reported risk

29As expected, the score on the placebo index has no effect on an individual’s take-up of the lowinterest rate contract, results shown in columns 1-3 of Table A.1 in the Appendix.

30Table A.2 shows the intensive (total) margin estimation for take-up of the Low interest rateloan, margin. Results are similar as for the extensive margin, with take-up being differentially higheramong business owners in a low risk environment, and among those reporting to be risk averse.

31The placebo index is not significantly correlated with take-up of the low collateral contract, pointestimates are low and not statistically significant, results shown in columns 4-6 of Table A.1 in theAppendix.

4.5. RESULTS 139

takers. The point estimate is, however, significant at conventional levels only in the

between subject specification - when adding firm fixed effects in column 8 the p-value

for the interaction term increases to 0.153.32

I now turn to the point estimates for the wealth controls, where the omitted cate-

gory is the highest wealth quartile. When adding wealth controls in columns 3 and 6

we see that respondents with lower wealth are more likely to crowd into borrowing as

the collateral requirement is lowered. The point estimate for the lowest and the 3rd

wealth quartiles are positive and statistically significant while the point estimate for

the 2nd wealth quartile is also positive but not significant. Thus, lowering the collat-

eral affects take-up more for the poorer 75% of potential borrowers than for the richest

quartile. The point estimates on the risk-aversion and low risk index terms are not

affected by controlling for wealth. Using the terminology of Boucher et al. (2008) this

finding is in line with poorer borrowers being quantity rationed (not having access to

sufficient collateral).

Taken together, the main results support the adverse selection results of Stiglitz

and Weiss (1981) and Wette (1983): both lowering the interest rate and lowering

the collateral disproportionately attracts less risky borrowers (and correspondingly

adverse selection would occur if the interest rate/collateral was increased). Wealth has

no differential impact on the hypothetical demand for a lower interest rate contract,

but compared to the most wealthy firm owners in my sample, less wealthy ones are

more likely to increase their loan demand if the collateral is lowered.

4.5.2 Heterogenous effects

As significant differences in characteristics were observed between retail and manu-

facturing firms in Table 4.1, and those differences appear to be related to investment

choices, estimating loan demand separately for manufacturing businesses and retail

businesses can shed more light on the determinants of loan demand. In Table 4.5, I

32I did not collect data on loan size for the low interest rate contract, ans can therefore not estimatethe intensive or total margin loan demand for this contract amendment.


present the result from running regressions 4.1 and 4.2 on these two categories sep-

arately for the low interest rate contract. In Panel A we see that the higher take-up

among respondents who are self reportedly risk averse holds in both groups, with

point estimates ranging between 0.072 and 0.090, although the smaller sample within

manufacturing means point estimates are just below conventional significance level

in the within-subject specifications (columns 5 and 6). Turning to Panel B of Table

4.5, we see that, conversely, the higher demand result for the low interest contract

among owners with low risk index (less risky business environment) is entirely driven

by manufacturing firms. While manufacturing firms who face a risky environment are

10-11 percentage points more likely to start borrowing if the interest rate is lowered,

those with a low risk index i.e. that do not face a risky environment are an additional

23 percentage points more likely to do so. For retail firms, the point estimate on the

interaction of low risk index with take up of the low interest contract is much smaller

and statistically insignificant.33 Table 4.6 shows the result from running regression 4.1

and 4.2 on these two categories separately for the low collateral contract. In Panel A

of Table 4.6, I examine whether self stated risk aversion affects a respondent’s take-up

of the low collateral contract when separating between retail (columns 1-3) and man-

ufacturing firms (columns 4-6). The differentially higher take-up of the low collateral

contract among risk averse firms is driven by manufacturing firms. While risk attitude

does not seem to matter for the take up of retail firms, risk averse manufacturing firm

owners are twice as likely as non risk averse manufacturing firm owners to start bor-

rowing when offered such a contract (the coefficient on the interaction term is 10.4-11.3

percentage points). This difference is significant at the 95 % confidence level in the

between subject specification (column 4) and stays significant at the 90 % confidence

level in the within subject specification (columns 5 and 6). Turning to the risk index,

Panel B of Table 4.6 shows the corresponding results. Just as observed for the low

interest rate contract, while among retail firms there is no difference between indi-

33Results for the placebo index are presented in Panel A of Table A.4 in the Appendix. Reassuringlyalso when analyzing manufacturing and retail separately there is no correlation between scores onthe placebo index and expressed take-up of the low interest rate contract.

4.5. RESULTS 141

viduals with a low risk index and those with riskier environments in their likelihood

to "crowd in" when offered a low collateral loan, among manufacturing firms, those

with a low risk index are 21-23 percentage points more likely than firms facing riskier

environments to select into borrowing when the collateral is lowered. This result holds

also when controlling for wealth of the business owner (column 6).34

The results in this subsection nuance the main findings. While more risk averse

owners among both manufacture firms and retail firms express higher demand for the

lower interest rate contract, the differential impact of riskiness on loan demand when

the collateral is lowered is driven by manufacturing firms. Moreover, the risk index,

i.e. the volatility in demand and sales appears much more important for loan demand

among manufacturing firms.

4.5.3 Robustness checks

In this subsection, I show that the main results are robust to excluding current bor-

rowers and to alternative ways of measuring risky behavior and environment of the

firm owner.

For non borrowers, answers to the hypothetical loan questions can be interpreted

as extensive margin loan demand. However, about 10% of my respondents are already

currently borrowing. One might be concerned that these respondents interpret the

hypothetical questions differently than non borrowers do. Since my main focus is on

estimating loan demand and behavior for the non borrowers, this is not a concern

in itself. It is, however, important to rule out that the current borrowers are driving

my results. For this reason, I estimate the main model excluding respondents that

are current borrowers. Table A.3 reassuringly shows that the results for both the low

interest rate contract and the low collateral contract are robust to excluding current

borrowers.

Since I do not have a direct measure of risky loan or investment behavior, I proxied

34Corresponding regression results for the placebo index are presented in the Appendix, Panel Bof Table A.4.


risk with survey measures of risk aversion and of fluctuations in demand. In this section,

estimate equations 4.1 and 4.2, for three alternative measures that can inform us about

a respondent’s risk profile: financial vulnerability, having precautionary savings, and

the habit of selling on credit. Table A.5 shows extensive margin take-up of the low

interest rate contract and focuses on how it differs depending on a firm owners financial

vulnerability. Columns 1-3 show results for the full sample. No clear pattern emerges

between being able to obtain 500,000 UGX and the interest expressed in the low

interest contract. However, when separately analyzing the results for the subgroups

of respondents that do not have access to the highest amount (2 Million UGX) in

columns 4-6, and respondents who are not among the most vulnerable (i.e. who do

have access to at least 500,000 UGX) in columns 7-9, the results suggest that it is

not the most vulnerable group that crowds in when interest rates are lowered. Table

A.6 shows the same for the low collateral contract. Among respondents who can not

obtain 2 million UGX, those who can obtain 500,000 UGX are around 22 ppt more

likely to take up a loan when the collateral value is lowered compared to those who

can not obtain such an amount. Among the non-vulnerable sample: i.e. those who do

have access to at least 500,000 in the case of am emergency, the richer subset, i.e. those

who have access also to 2 million UGX are less likely to crowd into borrowing when

collateral is lowered. This is not surprising given that these "rich" respondents seem

to have access to capital amounting roughly to the value of a loan.

In panel A of Table A.7, I study extensive margin demand for the contract de-

pending on whether a firm owner has saved any money in the past year. Access to

precautionary savings can be seen as a proxy for financial prudence. Columns 1-3

show the results for the low interest rate contract. Having saved last year does not

appear to matter for the hypothetical take-up of the low interest contract. Column 4-6

present results for the low collateral contract. The point estimate on the interaction

term between savings and saying "yes" to the low collateral contract is positive and

significant in all 3 specifications. Respondents who have savings are around 8 ppt more

likely to crowd in to borrowing when offered a low collateral contract than those who

4.6. VALIDATION CHECKS 143

did not have savings. Moreover in column 6 where I control for wealth, respondents in

the lower wealth quartiles are significantly more likely to crowd in when collateral is

lowered as compared to those in the richest quartile. Finally, panel B of Table A.7 I

use a dummy for "Sells to customers on credit" as an alternative measure of risky firm

owner behavior.35 Firm owners that sell on credit are no different in their demand of

the low interest contract from those who do not sell on credit. Such more "risky" own-

ers do however appear to be less interested in the low collateral contract than "safer"

owners (i.e. those who do not sell on credit). The point estimates on the interaction

term between selling on credit and take up of the low collateral loan are negative in all

three specifications and significant in the between subject specification (column 4 of

Panel B in Table A.7) but below conventional significance levels in the within-subject

specifications of columns 5 and 6.

4.6 Validation checks

In this paper, I study loan demand using hypothetical questions rather than actual

choices, something which may lead to concerns regarding to what extent the findings

can predict actual behavior. In this section, I perform a number of validation checks to

confirm that the answers to the hypothetical questions indeed are informative about

the preferences of a respondent.

First, I address the concern that respondents may not understand that my loan

offer questions are hypothetical. If this were the case, I would most likely observe a

negative correlation between being a previous borrower and take-up of any of the hy-

pothetical loan contracts, since those with borrowing experience would be less likely to

say "yes" to the hypothetical loans as they are already involved with other lenders. To

investigate whether this pattern appears, I exploit information from the survey section

on previous loan experience, which comes before the section of hypothetical demand in

35Note that selling on credit is a proxy for risky behavior, unlike having precautionary savings,which is a proxy for safe behavior. One would thus expect any results to work in the oppositedirection here.


the questionnaire: A dummy for whether the respondent is currently borrowing, and

the answers to the following two questions: "Are you planning to take a loan in the

next 2 years to use (mainly) in your business?" and "would you be able to take another

loan from the same loan provider if you wanted to?" (the latter question was asked

only to those who had borrowing experience). Respondents who are current borrowers

or plan to take a loan in the next two years and respondents who say that they would

be able to borrow with their previous lender again are not less likely to say "yes" to

the hypothetical offers. Instead, they are significantly more likely to express interest

in at least one of the hypothetical contracts (Table A.8 in the Appendix). Hence this

concern appears to be unwarranted.

Next, I study the internal consistency between answers to different survey question

about loans. Here, I focus on the respondents who say no to the question "Are you

planning to take a loan in the next 2 years to use (mainly) in your business?" (which is

asked in the section about loan experience, before the hypothetical question section).

The respondents who state that they are not planning to take a loan are asked to spec-

ify the reason for not planning to borrow. An overview of the most common stated

reasons is presented in Table A.9. I focus here on the stated reasons for not planning a

loan that are most closely related to the mechanisms that my contract variations tar-

get, and examine the correlation between, on one hand, stated reasons such as (a) high

cost (interest) of the loan (b) lack of collateral (c) fear of losing the collateral and (d)

the repayment structure, and, on the other hand, the respondents’ expressed interest

in hypothetical contracts that address these specific types of borrowing constraints. I

would expect those who say the interest rate is too high to be more convinced by the

low interest contract, those who have no collateral or who fear losing their collateral to

be more affected by changed collateral contracts, etc. As the upper panel of Table A.10

shows, this is precisely what we see. Looking at the correlation between the stated rea-

sons of not wanting a loan and a dummy for saying no to the standard loan but saying

yes to contract i where i ∈ {low interest, low collateral, any-asset collateral} I find

that respondents who say the interest rate is too high are significantly more likely to

4.6. VALIDATION CHECKS 145

switch to (hypothetical) borrowing when offered the low-interest contract while those

who have no collateral are not affected by a lower interest rate but are more likely to

switch to (hypothetical) borrowing if collateral is lowered or if any asset can be used as

collateral. Those who fear losing their collateral are however not convinced by any of

the contract amendments. Reassuringly, those reporting to be constrained mainly by

high interest rates are not systematically more likely to take up the contracts where

the collateral is changed, nor are those constrained by collateral likely to crowd in

when interest rate is amended.

Within-subject designs can be sensitive to carry over bias and range effects (Char-

ness et al., 2012). Although I did not vary the order in which loan contracts were

presented in the survey instrument, and can thus not present a perfect test for this,

it is still possible to examine the data for simple patterns. A pattern in which stated

demand steadily increases as we move through the list of hypothetical questions would

be a symptom of such bias. This is however not the case - demand is not monotonically

increasing between one contract amendment and the next.

An additional possible concern is that respondents without borrowing experience

may be less informed than previous borrowers about their real preferences regarding

loans. To address this, I restrict the sample to previous borrowers and compare, in this

subsample, the characteristics between those who do and those who do not "take up"

the various hypothetical contracts. While this approach suffers from similar selection

issues that the hypothetical setup is designed to avoid; under the assumption that

the directions of correlations are similar although level effects may be different, this

exercise can still tell us something about whether the effects I find are sensible. I ran

the most basic specifications of the same regressions in this subsample, and I also

examine simple pairwise correlations, shown in the lower panel of Table A.10 in the

Appendix. The patterns are consistent with the regression results of the within-subject

analysis of hypothetical take-up.

Taken together, people’s reasons for not planning a loan, stated in earlier sections

of the questionnaire, are consistent with how they actually respond to the hypothetical


contracts that are described to them later in the questionnaire. I also find evidence

that borrowers’ previous loan experience shapes their hypothetical demand. Results for

the subgroup of respondents who have loan experience, and can therefore be thought

more informed about their real loan preferences go in the same direction as overall

results, and it appears as if the respondents understand the hypothetical nature of the

questions.

4.7 Discussion and Conclusion

In this paper I study how changes in the standard credit contract available to small

entrepreneurs in Uganda affects the demand for the contract. In addition to studying

demand at the intensive margin, my representative sample of business owners allows

me to capture the changes in demand also at the extensive margin, i.e. to shed light

on the selection into borrowing.

In a sample of respondents where 20% have previous loan experience, I find that

around 14% express interest in a standard loan contract of the type offered by NGOs

and Micro Finance Institutions in the area. While there is no difference across business

sectors in the demand for the standard loan contract, firms that demand the standard

contract are more likely to have previous loan experience and to own land than those

who are not interested in the contract. They are also operating in a riskier business

environment and are less educated.

Furthermore, my results indicate that demand for loans is affected by contractual

changes. Lowering the annual interest rate from 25% to 20% attracts clients that are

more risk averse, and that operate in a business environment with less fluctuations

and less demand uncertainties. The same thing is true for loans with lower collateral

requirement, although here the result for the risk aversion variable is significant only in

the subsample of manufacturing firm owners. Importantly, the results for the low col-

lateral contract are unaffected by controlling for wealth of the firm owner. I show that

the results are robust to using alternative measures of firm (owner) riskiness, such as

4.7. DISCUSSION AND CONCLUSION 147

financial vulnerability and the absence of precautionary savings. Since sectoral differ-

ences affect the type of investment options available to a firm, and since fluctuations in

sales and demand appear more important for manufacturing businesses, I also analyze

the results separately for firms in manufacture and retail sectors and find that there

are noteworthy differences. In particular, I find that among manufacturing firms, de-

mand for both contracts is more affected by a risky business environment than among

retail firms. When it comes to the low collateral contract, the higher demand observed

among risk averse owners is driven by manufacturing firms. Given the differences in

production functions between these two categories of firms there are several possible

explanations for this differential effect of risk exposure and riskiness on loan demand

between the sectors. Manufacturing firms in my sample employ more labor and may

therefore be slower to adapt their expenditures to lower demand. Unlike food retailers,

who make up the bulk of retail firms in our sample, the type of inputs held by man-

ufacturing firms are not consumable for the firm owner, making manufacturers more

vulnerable in the case of unexpected drops in demand. The literature on small firm

business growth has placed a heavy emphasis on retail firms and on micro-enterprises

with no employees apart from the owner, and has therefore not been able to shed light

on such sectoral differences.

Taken together, the results for both the low interest rate contract and the low

collateral contract indicate that these contracts attract safer borrowers in terms of

their current business activity and situation, and their stated risk preferences. Since

I elicit hypothetical loan demand and do not observe any actual investments, I am

however not able to analyze the effect on the riskiness of the project.

Due to the hypothetical nature of the loan contract changes, these results should be

interpreted with caution and further studies of actual loan contract changes are needed

to confirm the selection into borrowing. The fact that previous studies of small firms

and credit has studied only clients that have self-selected into borrowing nevertheless

means that this study fills a gap in the literature both on credit markets in general

and on the effectiveness of microfinance. An ideal follow up to this project would be


to examine actual selection in an experiment where the price of credit or the collateral

requirements are randomly lowered. This would also make it possible to study the

actual investments that loans are used for and follow borrowers over time, thereby

getting a more direct measure of the riskiness of a project.

As a first step, a randomized control trial that builds on this study (Gulesci et al.

(2016)) examines the effect of actual contractual changes on loan use and effectiveness

in a sample of firms that have been approved for borrowing under the standard contract

prior to the experiment. Combining the knowledge from these two studies will allow for

better projections of the intensive and extensive margin demand effects from changes

to the standard microfinance contract.

REFERENCES 149

References

Abdellaoui, M., Driouchi, A. and l’Haridon, O., 2011. Risk aversion elicitation:

reconciling tractability and bias minimization. Theory and Decision, 71(1), pp.63-80.



Angelucci, M., Karlan, D. and Zinman, J. 2015. Microcredit impacts: Evi-

dence from a randomized microcredit program placement experiment by Compartamos

Banco. American Economic Journal: Applied Economics, 7(1), pp.151-182.

Attanasio, O., Augsburg, B., De Haas, R., Fitzsimons, E. and Harmgart, H.

2015. The impacts of microfinance: Evidence from joint-liability lending in Mongolia.

American Economic Journal: Applied Economics, 7(1), pp.90-122.

Augsburg, B, De Haas, R., Harmgart, H. and C. Meghir, 2015. The Impacts

of Microcredit: Evidence from Bosnia and Herzegovina. American Economic Journal:

Applied Economics, 7(1), pp 183-203.

Bandiera, O., R. Burgess, N. Das, S. Gulesci, I. Rasul, and M. Sulaiman

2016. Labor Markets and Poverty in Village Economies, Mimeo, Bocconi University.

Banerjee, A.V. 2001. Contracting constraints, credit markets and economic devel-

opment. Mimeo.

Banerjee, A.V. 2013. Microcredit under the microscope: what have we learned in the

past two decades, and what do we need to know? Annu. Rev. Econ., 5(1), pp.487-519.

Banerjee, A., Karlan, D. and Zinman, J. 2015 a. Six randomized evaluations

of microcredit: introduction and further steps. American Economic Journal: Applied

Economics, 7(1), pp.1-21.

Banerjee, A., E. Duflo, R. Glennerster, and C. Kinnan 2015 b. "The Miracle of

Microfinance? Evidence from a Randomized Evaluation." American Economic Journal:

Applied Economics, 7(1): 22-53


Berge, L.I.O., Bjorvatn, K. and Tungodden, B., 2014. Human and financial

capital for microenterprise development: Evidence from a field and lab experiment.

Management Science, 61(4), pp.707-722.

Bester, H., 1985. Screening vs. rationing in credit markets with imperfect informa-

tion. The American Economic Review, 75(4), pp.850-855.

Binswanger, H.P. 1981. Attitudes toward risk: Theoretical implications of an exper-

iment in rural India. The Economic Journal, 91(364), pp.867-890.

Bosch-Domènech, A. and Silvestre, J., 2013. Measuring risk aversion with lists:

a new bias. Theory and decision, 75(4), pp.465-496.

Boucher, S.R., Carter, M.R. and Guirkinger, C., 2008. Risk rationing and

wealth effects in credit markets: Theory and implications for agricultural development.

American Journal of Agricultural Economics, 90(2), pp.409-423.

BRAC’s website. Accessed on May 2, 2016, http://www.brac.net/microfinance.

Charness, G., Gneezy, U. and Kuhn, M.A., 2012. Experimental methods: Between-

subject and within-subject design. Journal of Economic Behavior Organization, 81(1),

pp.1-8.

Crépon, B., E. Duflo, F. Devoto, W. Pariente 2015. Estimating the impact

of microcredit on those who take it up: Evidence from a randomized experiment in

Morocco, American Economic Journal: Applied Economics, Vol. 7, No. 1.

de Mel, S., McKenzie, D. and Woodruff, C. 2008. Returns to capital in mi-

croenterprises: evidence from a field experiment. The Quarterly Journal of Economics,

pp.1329-1372.

de Mel, S., McKenzie, D.J. and Woodruff, C.M. 2009. Innovative firms or in-

novative owners? Determinants of innovation in micro, small, and medium enterprises.

Determinants of Innovation in Micro, Small, and Medium Enterprises (May 1, 2009).

World Bank Policy Research Working Paper Series, Vol.

REFERENCES 151

De Meza, D. and Webb, D.C., 1987. Too much investment: a problem of asym-

metric information. The quarterly journal of economics, pp.281-292.

Dohmen, T., Falk, A., Huffman, D., Sunde, U., Schupp, J. and Wagner,

G.G. 2011. Individual risk attitudes: Measurement, determinants, and behavioral con-

sequences. Journal of the European Economic Association, 9(3), pp.522-550.

Fafchamps, M., McKenzie, D., Quinn, S. and Woodruff, C., 2014. Microen-

terprise growth and the flypaper effect: Evidence from a randomized experiment in

Ghana. Journal of Development Economics, 106, pp.211-226.

Field, E., Pande, R., Papp, J. and Rigol, N. 2013. Does the classic microfinance

model discourage entrepreneurship among the poor? Experimental evidence from In-

dia. The American Economic Review, 103(6), pp.2196-2226.

Gulesci, S., Loiacono, F., Madestam, A. and Stryjan, M. 2016. Credit con-

tract structure and firm growth: Evidence from a randomized control trial. Mimeo,

Stockholm University.

International Labor Organization, 2015. Small and medium-sized enterprises and

decent and productive employment creation. International Labour Conference, 104th

Session, 2015. Report pp 2.



List, J.A. and J.F. Shogren 2002. Calibration of willingness-to-accept. Journal of

Environmental Economics and Management, 43(2), pp.219-233.

Neill, H.R., Cummings, R.G., Ganderton, P.T., Harrison, G.W. andMcGuckin,

T. 1994. Hypothetical surveys and real economic commitments. Land economics,

pp.145-154.

Stiglitz, J.E. and Weiss, A., 1981. Credit rationing in markets with imperfect

information. The American economic review, 71(3), pp.393-410.


Tarozzi, A., Desai, J. and Johnson, K., 2015. The impacts of microcredit: Evi-

dence from Ethiopia. American Economic Journal: Applied Economics, 7(1), pp.54-89.

Uganda Bureau of Statistics, 2011. Report on the census of business establish-

ments, 2010/11.

Uganda Investment Authority’s website accessed on Dec 12, 2013,

http://www.ugandainvest.go.ug/index.php/sme-policy.

Valdivia, M., 2013. Business training plus for female entrepreneurship? Evidence

from a field experiment in Peru. Unpublished manuscript GRADE, Lima, December.



World Bank 2013 World Development Report 2013: Jobs (Washington, DC: World

Bank)

TABLES 153

Tables


Table

4.1:Summary

statisticsfor

entiresam

ple,andseparately

forManufacture

vs.Retailsecors

Fullsam

ple

mean

St.dev.

NRetailm

eanManufacture

mean

Difference

p-value(N

=636)

(N=289)

difference

Firm

characteristics

Sectorin

manufacturing

0.3120.464

925-

--

-Firm

age6.671

5.277891

6.247.61

-1.37***0.000

Totalno.of

workers

2.7621.771

9252.14

4.12-1.98***

0.000Share

firmswith

employees

0.7340.442

9250.64

0.93-0.29***

0.000Typicalm

onthlyprofit

lastyear

(1000UGX)

997.1541325.759

8821055.506

868.356187.149*

0.052Aggregate

assetvalue

(1000UGX)

2238.3675217.137

9252293.303

2117.471175.832

0.635Value

ofcurrent

stock/inventories(1000

UGX)

12883.21516439.932

81513789.455

10787.0733002.382**

0.017Owner

characteristics

Owner

isfem

ale0.282

0.45925

0.380.06

0.32***0.000

Owner

yearsof

education11.458

3.011889

11.7510.80

0.95***0.000

Owner

owns

land0.442

0.497902

0.450.43

0.020.589

Cred

itexp

erience

andfinan

cialinclu

sionEver

borrowed

fromform

al/semi-form

al0.206

0.405922

0.200.21

-0.010.668

Borrow

edin

last2years

0.0970.296

9210.10

0.100.00

0.986Has

bankaccount

0.7470.435

8940.79

0.650.14***

0.000Keeps

books0.732

0.443899

0.810.57

0.24***0.000

Financialliteracy

score0.565

0.316920

0.540.62

-0.08***0.000

Risk

attitudes

andriskin

essSelf

reportedriskiness

4.4372.704

9104.40

4.51-0.11

0.574Risk

index2.226

0.644919

2.202.28

-0.09*0.063

Placebo

index1.831

0.65921

1.851.80

0.050.314

Savedlast

year2.465

12.46917

2.582.21

0.380.159

Can

obtain500K

UGX

0.7660.424

9130.80

0.690.11***

0.000Can

obtain2M

UGX

0.510.5

8880.56

0.410.15***

0.000Sells

tocustom

erson

credit0.733

0.443925

0.790.62

0.17***0.000

Investment

dem

and

Wants

more

labor0.186

0.389925

0.140.30

-0.16***0.000

Wants

more

capital0.518

0.5925

0.410.76

-0.35***0.000

Notes:

Total

no.of

workers

isthe

totalnum

berof

workers

inafirm

,including

theow

nerand

bothpaid

andunpaid

employees.

Monetary

variablesare

reportedin

1000’sUgandan

Shillings(U

GX).Risk

indexis

compiled

fromquestions

measuring

whether

therespondent

facesabusiness

environment

with

fluctuationsor

unpredictability.The

Placebo

indexis

compiled

fromquestions

aboutthe

businessenvironm

entthat

arenot

relatedto

thesetypes

ofrisks.

Selfreported

riskiness:score

when

therespondent

isasked

torank

herselfon

a0-10

scaleaccording

tohow

much

sheis

willing

totake

risks.*p<0.1,

**p<0.05,

***p<0.01

TABLES 155

Table 4.2: Hypothetical take-up of loan contracts, overview

Full sample Retail sectors Manufacturing sectors

% Yes % Yes # responded %Yes # responded

Yes to standard contract 14.14 13.68 614 15.14 284Yes to low interest contract 24.67 22.54 621 29.27 287Yes to grace period contract 28.98 26.92 624 33.45 287Yes to flexible repayment contract 33.26 32.26 623 35.42 288Yes to low collateral contract 27.84 26.8 582 30.0 280

No to all the contracts 54.27 54.95 586 52.86 280

Note: Wants none of the contracts: Dummy =1 if respondent said no to all of the listed contract variations. In case ofmissing response to one or more of the contracts and the respondent said no to the remaining contracts, the variable iscoded as missing. The contracts are listed in the order that they were presented to the respondents.

Table 4.3: Who says yes to the standard loan contract?

No to Yes to Difference p-valuestandard contract standard contract

Manufacturing sector 0.31 0.34 -0.03 0.560Age of business 6.53 7.53 -0.99* 0.051Owner is female 0.27 0.31 -0.04 0.401Years of education 11.55 10.94 0.61** 0.041Total # of workers 2.73 2.99 -0.26 0.130Has employees apart from owner 0.73 0.77 -0.04 0.299Owns land anywhere 0.43 0.51 -0.08* 0.076Borrowing experience 0.18 0.37 -0.19*** 0.000Denied loan size 0.34 0.45 -0.11 0.309Ever delayed on loan repayment 0.13 0.07 0.06 0.373Sells to customers on credit 0.74 0.69 0.06 0.180Self reported riskiness 4.43 4.51 -0.09 0.742Risk index 2.21 2.32 -0.11* 0.070Placebo index 1.82 1.83 -0.01 0.862Has a bank account 0.74 0.77 -0.03 0.481Currently keeps books for business 0.73 0.74 -0.01 0.895Financial literacy score 0.56 0.61 -0.05* 0.070Monthly profit past year (1000’s UGX) 992.47 970.67 21.81 0.866Aggregate assets (1000’s UGX) 2160.91 2525.09 -364.18 0.448Stock/inventories value (1000’s UGX) 12703.46 11785.09 918.37 0.575Wants more labor 0.17 0.28 -0.11*** 0.002Wants more capital 0.52 0.50 0.02 0.675

Notes: Self reported riskiness: score when asked to rank herself on a 0-10 scale according to how much she is willing to takerisks. Risk index is compiled from questions measuring whether the respondent faces a business environment with fluctuationsor unpredictability. The Placebo index is compiled from questions about the business environment that are not related tothese types of risks. Denied loan size and Ever delayed on loan repayment are dummies and shares are reported conditionalon respondent having taken a loan in the past 2 years. * p<0.1, ** p<0.05, *** p<0.01


Table 4.4: Extensive margin demand, full sample

Panel A: Demand for Low interest rate contract(1) (2) (3) (4) (5) (6)

Low interest 0.098∗∗∗ 0.099∗∗∗ 0.079∗∗∗ 0.080∗∗∗ 0.080∗∗∗ 0.065∗∗[0.012] [0.016] [0.030] [0.013] [0.018] [0.030]

Low interest * risk index low 0.085∗∗∗ 0.087∗ 0.085∗[0.033] [0.046] [0.046]

Low interest * risk averse 0.077∗∗∗ 0.081∗∗ 0.079∗∗[0.023] [0.032] [0.032]

Low interest * bottom wealth q 0.014 0.007[0.041] [0.041]

Low interest * 2nd wealth q 0.035 0.027[0.043] [0.043]

Low interest * 3rd wealth q 0.034 0.030[0.043] [0.044]

Mean demand standard contr. 0.145 0.145 0.132 0.132 0.132 0.134Interviewer& time controls yes no no yes no noFirm FE no yes yes no yes yesObservations 1792 1792 1792 1765 1765 1765Adjusted R2 0.199 0.672 0.627 0.153 0.673 0.672

Panel B: Demand for Low collateral contract(1) (2) (3) (4) (5) (6)

Low collateral 0.157∗∗∗ 0.146∗∗∗ 0.089∗∗∗ 0.150∗∗∗ 0.143∗∗∗ 0.088∗∗∗[0.014] [0.020] [0.033] [0.017] [0.024] [0.034]

Low collateral * risk index low 0.084∗∗ 0.112∗∗ 0.108∗∗[0.039] [0.055] [0.053]

Low collateral * riskaverse 0.052∗ 0.055 0.052[0.027] [0.039] [0.039]

Low collateral * bottom wealth q 0.090∗ 0.087∗[0.051] [0.051]

Low collateral * 2nd wealth q 0.023 0.020[0.047] [0.047]

Low collateral * 3rd wealth q 0.116∗∗ 0.120∗∗[0.052] [0.054]


Notes: Low interest (Low collateral) is a dummy=1 if the contract offered is the low interest (low collateral) contract.Risk index is compiled from questions measuring whether the respondent faces a business environment with fluctuationsor unpredictability. The Placebo index is compiled from questions about the business environment that are not relatedto these types of risks. Risk averse: a dummy variable =1 if the respondents gives herself an above median score whenasked to rank herself on a 0-10 scale according to how much she is willing to take risks. Mean demand standard contr.displayed below the table indicates the mean hypothetical takeup of the standard contract in the base category, i.e.respondents with risk index low=0 in columns 1-3 and with risk aversion=0 in columns 4-6, and, in addition, in wealthquartile=4 in columns 3 and 6. Standard errors in brackets are clustered at the firm level,* p<0.1, ** p<0.05, *** p<0.01

TABLES 157

Table 4.5: Extensive margin demand for Low interest contract, retail vs. manufacturing

Panel A: Self stated risk aversionRetail Manufacturing

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

Low interest 0.064∗∗∗ 0.066∗∗∗ 0.054∗ 0.113∗∗∗ 0.112∗∗∗ 0.090[0.015] [0.020] [0.032] [0.029] [0.038] [0.068]

Low interest * riskaverse 0.072∗∗∗ 0.076∗∗ 0.075∗∗ 0.084∗ 0.085 0.090[0.027] [0.037] [0.037] [0.047] [0.062] [0.061]

Low interest * bottom wealth q 0.027 -0.038[0.048] [0.079]


Low interest * 3rd wealth q -0.014 0.092[0.046] [0.091]


Panel B: Risk indexRetail Manufacturing

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

Low interest 0.090∗∗∗ 0.094∗∗∗ 0.073∗∗ 0.112∗∗∗ 0.111∗∗∗ 0.096[0.014] [0.019] [0.034] [0.022] [0.029] [0.059]

Low interest * risk index low 0.031 0.032 0.034 0.228∗∗∗ 0.229∗∗ 0.227∗∗[0.035] [0.048] [0.047] [0.077] [0.103] [0.104]





Notes: Low interest is a dummy=1 if the contract offered is the low interest contract.Risk averse: a dummyvariable =1 if the respondents gives herself an above median score when asked to rank herself on a 0-10 scaleaccording to how much she is willing to take risks. Risk index is compiled from questions measuring whetherthe respondent faces a business environment with fluctuations or unpredictability. Mean demand standard contr.displayed below the table indicates the mean hypothetical takeup of the standard contract in the relevant basecategory for each column. Standard errors in brackets are clustered at the firm level,* p<0.1, ** p<0.05, *** p<0.01


Table 4.6: Extensive margin demand for Low collateral contract, retail vs. manufacturing

Panel A: Self stated risk aversionRetail Manufacturing

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

Low collateral 0.157∗∗∗ 0.156∗∗∗ 0.103∗∗ 0.129∗∗∗ 0.116∗∗∗ 0.051[0.022] [0.030] [0.042] [0.030] [0.039] [0.057]

Low collateral * riskaverse 0.031 0.026 0.018 0.104∗∗ 0.114∗ 0.113∗[0.034] [0.047] [0.048] [0.050] [0.067] [0.066]

Low collat.* bottom wealth q 0.063 0.138[0.061] [0.095]

Low collateral * 2nd wealth q 0.049 -0.039[0.059] [0.074]

Low collateral * 3rd wealth q 0.121∗ 0.132[0.069] [0.088]


Panel B: Risk indexRetail Manufacturing

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

Low collateral 0.163∗∗∗ 0.153∗∗∗ 0.099∗∗ 0.143∗∗∗ 0.133∗∗∗ 0.072[0.018] [0.025] [0.040] [0.024] [0.032] [0.059]

Low collateral * risk index low 0.030 0.063 0.056 0.215∗∗∗ 0.231∗∗ 0.222∗∗[0.045] [0.062] [0.060] [0.081] [0.109] [0.106]

Low collateral * bottom wealth q 0.065 0.133[0.061] [0.094]

Low collateral 2nd wealth q 0.049 -0.038[0.059] [0.073]

Low collateral 3rd wealth q 0.115∗ 0.122[0.067] [0.086]


Notes: Low collateral is a dummy=1 if the contract offered is the low collateral contract. Risk averse: a dummyvariable =1 if the respondents gives herself an above median score when asked to rank herself on a 0-10 scale accordingto how much she is willing to take risks. Risk index is compiled from questions measuring whether the respondentfaces a business environment with fluctuations or unpredictability. Mean demand standard contr. displayed belowthe table indicates the mean hypothetical takeup of the standard contract in the relevant base category for eachcolumn. Standard errors in brackets are clustered at the firm level, * p<0.1, ** p<0.05, *** p<0.01

APPENDIX 1 159

Appendix 1


Table A.1: Extensive margin demand for Low interest and low collateral contract, Placebo index

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

Low interest 0.116∗∗∗ 0.118∗∗∗ 0.097∗∗∗[0.014] [0.019] [0.030]

Placebo low -0.023 -0.001[0.030] [0.030]

Low interest * placebo low -0.008 -0.007 -0.011[0.023] [0.032] [0.033]

Low interest * bottom wealth q 0.011[0.041]

Low interest* 2nd wealth q 0.035[0.043]

Low interest * 3rd wealth q 0.040[0.044]

Low collateral 0.179∗∗∗ 0.171∗∗∗ 0.112∗∗∗[0.017] [0.024] [0.035]

Low collateral *placebo low -0.022 -0.014 -0.017[0.028] [0.039] [0.039]

Low collateral * bottom wealth q 0.088∗[0.051]

Low collateral * 2nd wealth q 0.026[0.047]

Low collateral * 3rd wealth q 0.127∗∗[0.053]

Interviewer& time controls yes no no yes no noFirm FE no yes yes no yes yesObservations 1786 1786 1786 1706 1706 1706Adjusted R2 0.153 0.671 0.671 0.197 0.566 0.572

Notes: Low interest (Low collateral) is a dummy=1 if the contract offered is the low interest (low collateral) contract.The Placebo index is compiled from answers to questions about the difficulty of repaying loans that are unrelated tosales and demand fluctuations. Standard errors in brackets are clustered at the firm level,* p<0.1, ** p<0.05, *** p<0.01

APPENDIX 1 161

Tab

leA.2:Intensive

(Total)margindeman

dforLo

winterest

contract

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Low

interest

1.476∗∗∗

1.500∗∗∗

1.192∗∗∗

1.798∗∗∗

1.766∗∗∗

1.462∗∗∗

1.240∗∗∗

1.237∗∗∗

1.005∗∗

[0.177]

[0.241]

[0.445]

[0.213]

[0.290]

[0.448]

[0.206]

[0.272]

[0.447]

Low

interest

*risk

indexlow

1.427∗∗∗

1.254∗

1.224∗

[0.496]

[0.692]

[0.694]

Low

interest*p

lacebo

indexlow

-0.195

-0.129

-0.178

[0.347]

[0.480]

[0.487]

Low

interest

*risk

averse

1.135∗∗∗

1.115∗∗

1.087∗∗

[0.348]

[0.478]

[0.480]

Low

interest*b

ottom

wealthq

0.204

0.167

0.113

[0.610]

[0.610]

[0.616]

Low

interest

*2n

dwealthq

0.524

0.518

0.412

[0.653]

[0.657]

[0.653]

Low

interest

*3rdwealthq

0.520

0.599

0.452

[0.655]

[0.664]

[0.667]

Interviewer&

timecontrols

yes

nono

yes

nono

yes

nono

Firm

FE

noyes

yes

noyes

yes

noyes

yes

Observation

s1782

1782

1782

1776

1776

1776

1755

1755

1755

Adjusted

R2

0.166

0.676

0.676

0.165

0.675

0.675

0.165

0.676

0.675

Notes:Lo

winterest

isadu

mmy=

1ifthecontract

offered

isthelow

interest

contract.Stan

dard

errors

inbrackets

areclusteredat

thefirm

level,

*p<

0.1,

**p<

0.05,***p<

0.01


Table A.3: Extensive margin demand, sample: non borrowers

Panel A: Demand for low interest loan, sample: non-borrowers(1) (2) (3) (4) (5) (6)

Low interest 0.073*** 0.079*** 0.065** 0.091*** 0.098*** 0.079***[0.014] [0.018] [0.030] [0.012] [0.016] [0.030]

Low interest*risk index low 0.072* 0.084* 0.083*[0.037] [0.049] [0.050]

Low interest*risk averse 0.071*** 0.078** 0.076**[0.024] [0.033] [0.033]

Low interest*bottom wealth q 0.009 0.017[0.043] [0.042]

Low interest*2nd wealth q 0.025 0.033[0.044] [0.044]

Low interest*3rdwealth q 0.023 0.026[0.045] [0.044]

Interviewer& Time controls yes no no yes no noFirm FE no yes yes no yes yesObservations 1724 1724 1724 1750 1750 1750Adjusted R2 0.148 0.671 0.670 0.149 0.673 0.672

Panel B: Demand for low collateral loan, sample: non-borrowers(1) (2) (3) (4) (5) (6)

Low collateral 0.150*** 0.145*** 0.091*** 0.143*** 0.139*** 0.090**[0.015] [0.020] [0.034] [0.018] [0.025] [0.035]

Low collateral*risk index low 0.082* 0.116* 0.115**[0.042] [0.059] [0.058]

Low collateral*risk averse 0.051* 0.060 0.056[0.028] [0.040] [0.040]

Low collateral*bottom wealth q 0.084 0.078[0.053] [0.053]

Low collateral*2nd wealth q 0.022 0.018[0.049] [0.049]

Low collateral*3rd wealth q 0.110** 0.108*[0.054] [0.056]

Interviewer& Time controls yes no no yes no noFirm FE no yes yes no yes yesObservations 1675 1675 1675 1652 1652 1652Adjusted R2 0.190 0.566 0.571 0.188 0.562 0.567

Notes: Low interest (Low collateral) is a dummy=1 if the contract offered is the low interest (low collateral) contract.Risk index is compiled from questions measuring whether the respondent faces a business environment with fluctuationsor unpredictability. The Placebo index is compiled from questions about the business environment that are not related tothese types of risks. Risk averse: a dummy variable =1 if the respondents gives herself an above median score when askedto rank herself on a 0-10 scale according to how much she is willing to take risks. Standard errors in brackets are clusteredat the firm level, * p<0.1, ** p<0.05, *** p<0.01

APPENDIX 1 163

Table A.4: Extensive margin demand by placebo index, retail vs manufacture

Panel A: Low interest rate contractRetail Manufacturing

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

Low interest 0.100∗∗∗ 0.104∗∗∗ 0.084∗∗ 0.150∗∗∗ 0.149∗∗∗ 0.128∗∗[0.016] [0.022] [0.034] [0.029] [0.038] [0.061]

Low interest * placebo low -0.014 -0.016 -0.014 -0.000 0.002 0.004[0.027] [0.036] [0.037] [0.047] [0.063] [0.064]





Panel B: Low collateral contractRetail Manufacturing

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

Low collateral 0.174∗∗∗ 0.168∗∗∗ 0.114∗∗∗ 0.188∗∗∗ 0.179∗∗∗ 0.109∗[0.020] [0.028] [0.042] [0.031] [0.042] [0.061]

Low collateral*placebo low -0.018 -0.010 -0.016 -0.027 -0.022 -0.027[0.035] [0.049] [0.049] [0.050] [0.067] [0.066]

Low collateral*bottom wealth q 0.061 0.145[0.061] [0.094]

Low collateral*2nd wealth q 0.050 -0.031[0.059] [0.073]

Low collateral*3rd wealth q 0.123∗ 0.136[0.068] [0.089]


Notes: Low interest (Low collateral) is a dummy=1 if the contract offered is the low interest (low collateral)contract. The Placebo index is compiled from answers to questions about the difficulty of repaying loans that areunrelated to sales and demand fluctuations. Standard errors in brackets are clustered at the firm level,* p<0.1, ** p<0.05, *** p<0.01


Table

A.5:E

xtensivemargin

demand

forLow

interestcontract,by

financialvulnerability

Fullsample

Respondents

who

cannotobtain

2M

Respondents

who

canobtain

500K

(1)(2)

(3)(4)

(5)(6)

(7)(8)

(9)

Lowinterest

0.092 ∗∗∗0.096 ∗∗∗

0.066 ∗0.087 ∗∗∗

0.091 ∗∗∗0.062

0.163 ∗∗∗0.157 ∗∗∗

0.144 ∗∗∗[0.021]

[0.029][0.039]

[0.022][0.030]

[0.054][0.025]

[0.034][0.048]

Lowinterest

*Can

obtain500K

0.0260.023

0.0280.074 ∗∗

0.0660.068

[0.025][0.034]

[0.034][0.033]

[0.045][0.045]

Lowinterest

*Can

obtain2M

-0.063 ∗∗-0.054

-0.054[0.029]

[0.040][0.042]

Lowinterest

*bottom

wealth

q0.016

0.011-0.001

[0.041][0.064]

[0.050]Low

interest*2nd

wealth

q0.047

0.0380.027

[0.043][0.068]

[0.054]Low

interest*3rd

wealth

q0.042

0.0650.030

[0.043][0.076]

[0.051]Can

obtain500K

0.022-0.057

[0.034][0.042]

Can

obtain2M

0.128 ∗∗∗[0.031]

Interviewer&

timecontrols

yesno

noyes

nono

yesno

noFirm

FE

noyes

yesno

yesyes

noyes

yesObservations

17741774

1774850

850850

13081308

1308Adjusted

R2

0.1520.672

0.6720.160

0.6220.621

0.1680.648

0.647

Notes:

Lowinterest

isadum

my=

1ifthe

contractoffered

isthe

lowinterest

contract.Can

obtain500K

/2M

:Dum

my=1ifrespondent

saysthat

shecould

obtainsaid

amount

within

amonth

ifsheneeded

itfor

anem

ergency.Colum

ns1-3

displayresults

forthe

fullsample.In

columns

4-6Iexclude

therichest

respondents,i.e.respondentswho

saythat

theycannot

obtain2Million

UGX.In

columns

7-9the

sample

isrestricted

tonon

vulnerablerespondents,

i.e.respondents

who

saythat

theycan

obtain500,000

UGX.Standard

errorsin

bracketsare

clusteredat

thefirm

level,*p<0.1,

**p<0.05,

***p<0.01.

APPENDIX 1 165

Tab

leA.6:E

xtensive

margindeman

dforLo

wcolla

teralc

ontract,by

finan

cial

vulnerab

ility

Fullsample

Respo

ndents

who

cann

otob

tain

2M

Respo

ndents

who

canob

tain

500K

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Low

colla

teral

0.114∗∗∗

0.118∗∗∗

0.040

0.088∗∗∗

0.093∗∗∗

0.040

0.329∗∗∗

0.320∗∗∗

0.267∗∗∗

[0.024]

[0.033]

[0.046]

[0.023]

[0.031]

[0.068]

[0.033]

[0.046]

[0.058]

Low

colla

teral*Can

obtain

500K

0.078∗∗∗

0.066∗

0.073∗

0.242∗∗∗

0.227∗∗∗

0.221∗∗∗

[0.029]

[0.040]

[0.041]

[0.040]

[0.055]

[0.055]

Low

colla

teral*

Can

obtain

2M

-0.214∗∗∗

-0.211∗∗∗

-0.204∗∗∗

[0.037]

[0.052]

[0.052]

Low

colla

teral*

bottom

wealthq

0.106∗∗

0.083

0.074

[0.052]

[0.083]

[0.062]

Low

colla

teral*

2ndwealthq

0.046

-0.007

0.003

[0.047]

[0.081]

[0.059]

Low

colla

teral*

3rdwealthq

0.138∗∗

0.141

0.116∗

[0.054]

[0.095]

[0.061]

Can

obtain

500K

-0.016

-0.122∗∗∗

[0.034]

[0.040]

Can

obtain

2M0.124∗∗∗

[0.031]

Interviewer&

timecontrols

yes

nono

yes

nono

yes

nono

Firm

FE

noyes

yes

noyes

yes

noyes

yes

Observation

s1696

1696

1696

830

830

830

1239

1239

1239

Adjusted

R2

0.200

0.566

0.573

0.241

0.533

0.540

0.233

0.560

0.565

Notes:Lo

wcolla

teralisadu

mmy=

1ifthecontract

offered

isthelow

colla

teralcontract.Can

obtain

500K

/2M

:Dum

my=1ifrespon

dent

says

that

shecouldob

tain

said

amou

ntwithinamon

thifshene

eded

itforan

emergency.

Colum

ns1-3displayresultsforthefullsample.

Incolumns

4-6Iexclud

etherichestrespon

dents,i.e

.respo

ndents

who

saythat

they

cann

otob

tain

2MillionUGX.In

columns

7-9thesampleis

restricted

tono

nvu

lnerab

lerespon

dents,

i.e.respon

dentswho

saythat

they

canob

tain

500,000UGX.Stan

dard

errors

inbrackets

areclusteredat

thefirm

level,*p<

0.1,

**p<

0.05,***p<

0.01.


Table A.7: Extensive margin demand, related to savings or selling on credit

Panel A: Demand related to having precautionary savings(1) (2) (3) (4) (5) (6)

Low interest 0.079∗∗∗ 0.092∗∗ 0.065[0.029] [0.038] [0.050]

Saved last year -0.036 -0.048[0.039] [0.038]

Low interest * saved last year 0.044 0.030 0.033[0.031] [0.042] [0.043]

Low interest * bottom wealth q 0.016[0.042]

Low interest * 2nd wealth q 0.044[0.045]

Low interest * 3rd wealth q 0.042[0.044]

Low collateral 0.097∗∗∗ 0.100∗∗ 0.026[0.031] [0.041] [0.054]

Low collateral *saved last year 0.087∗∗ 0.078∗ 0.088∗[0.034] [0.046] [0.048]

Low collateral * bottom wealth q 0.101∗[0.052]

Low collateral * 2nd wealth q 0.035[0.048]

Low collateral * 3rd wealth q 0.123∗∗[0.054]


Panel B: Demand related to selling on credit(1) (2) (3) (4) (5) (6)

Low interest 0.095∗∗∗ 0.099∗∗∗ 0.080∗∗[0.020] [0.028] [0.035]

Sells on credit -0.056∗∗ -0.030[0.028] [0.028]

Low interest rate* sells on credit 0.026 0.023 0.021[0.024] [0.033] [0.034]

Low interest rate * bottom wealth q 0.010[0.041]

Low interest rate * 2nd wealth q 0.034[0.043]

Low interest rate * 3rd wealth q 0.038[0.044]

Low collateral 0.211∗∗∗ 0.202∗∗∗ 0.139∗∗∗[0.028] [0.039] [0.044]

Low collateral* sells on credit -0.053∗ -0.047 -0.048[0.032] [0.045] [0.044]

Low collateral*bottom wealth q 0.092∗[0.051]

Low collateral* 2nd wealth q 0.034[0.047]

Low collateral*3rd wealth q 0.127∗∗[0.054]


Notes: Columns 1-3 display results for contract 2 (low interest) while columns 4-6 display results for the low collateralcontract. Standard errors in brackets are clustered at the firm level, * p<0.1, ** p<0.05, *** p<0.01.

APPENDIX 1 167

Table A.8: Correlation between having access to borrowing and the stated interest in hypotheticalcontracts

Interested in at least one contract

Correlation coeff. N

Is currently borrowing 0.094∗∗∗ 860Plans loan next 2 years 0.055(*) 842Can borrow from same lender 0.255** 71

Notes: Simple pairwise correlations. The variable Can borrow from same lenderis only asked to respondents who have borrowed in the past 2 years.(*) p<0.15, * p<0.1, ** p<0.05, *** p<0.01.

Table A.9: Stated reasons for not planning to borrow in the next 2 years

Full sample Retail sectors Manufacturing sectors

# % # % # %

Do not need capital 168 26.97 136 31.70 32 16.49Interest rate too high 165 26.48 108 25.17 57 29.38Fear losing collateral 158 25.36 104 24.24 54 27.83Do not have access to collateral 64 10.27 43 10.02 21 10.82Installment too often 24 3.85 12 2.80 12 6.19Other reason 44 7.06 26 6.06 18 9.28

Notes: The table shows the main stated reasons for not planning a loan for the respondents who said thatthey were not planning a loan in the next 2 years.

Table A.10: Correlation between hypothetical demand and stated reasons for not planning to borrow

Crowds in to (contract): Low interest Low collateral Any collateral N

Do not need capital -0.039 -0.022 0.011 684Interest rate too high 0.126*** 0.877 -0.016 684Fear losing collateral -0.078 -0.102*** -0.049 684Do not have access to collateral 0.015 0.172*** 0.136*** 684Installment too often -0.008 0.944 -0.043 684

Risk averse 0.159** 0.058 0.017 183Risk index low 0.229*** 0.130* 0.028 190

Notes: Simple pairwise correlation. In the first 5 rows, the sample is restricted to respondents stating thatthey do not plan to borrow in the next 2 years. In the 2 last rows, the sample is restricted to respondentswith borrowing experience. Any collateral contract is not included in the main analysis, but I include ithere since it provides additional support for the validity of the responses to the hypothetical questions.* p<0.1, ** p<0.05, *** p<0.01.


Appendix 2: Loan contract variations

1. Standard contract. "Imagine you were offered the opportunity to take a loan. If

you decide to take this loan, you can borrow up to 3 million Shillings. You would

need to repay this amount plus a 25% interest within one year. The repayments

have to be done in equal monthly repayment installments over the year. SHOW

EXAMPLE. The lender requests security (collateral) in the form of land. That

is, in order to borrow a certain amount, for example, 3 million, you need to have

formal property rights to land valued to 3 million and in case you fail to repay,

the lender will claim the 3 million in terms of your land." If you were offered

such a loan, would you choose to borrow? If yes, how much would you like to

borrow?

2. Low interest rate contract. "Now think about the loan contract we had above

(remind the respondent about the terms equal monthly repayments starting one

month after the loan is taken, and collateral in the form of land). Suppose all

the terms stay the same except the interest rate on the loan is 20% instead of

25%. SHOW EXAMPLE." Do you think this is a better offer compared to the

previous loan contract you were offered? If you were offered such a loan, would

you choose to borrow? If yes, how much would you like to borrow under this

contract?

3. Low collateral contract. If the collateral/security was land for 50% (=half)

of the value of the loan, would that be better than the very first contract? If you

were offered such a loan, would you choose to borrow?

Chapter 5

Credit Contract Structure and Firm

Growth: Evidence from a Randomized

Control Trial∗

5.1 Introduction

Credit markets in many developing countries are characterized by credit rationing,

which excludes certain prospective borrowers from credit access, thereby hindering

business growth. Another, less recognized, constraint is that key aspects of the most

common and accessible form of financing: debt, may inhibit firm expansion. Expanding

a business requires learning, or training employees, in how to use new inputs or market

new products efficiently and how to build a reputation. To the extent that learning

is mechanical, returns to an investment start low but increase gradually. When in-

troducing new products, there is also often uncertainty about demand, implying that

the timing of returns is uncertain. In addition, starting or expanding a business may

∗This paper is co-authored with Selim Gulesci, Francesco Loiacono and Andreas Madestam. Theauthors would like to thank BRAC Uganda, in particular the SEP program staff and the staff atBRAC Uganda Research and Evaluation Unit, for their collaboration and practical help in the im-plementation of the experiment. We also thank Emanuele Bracanti, Chiara Dall’Aglio, FrancescaVisinoni and Matteo Voltan for excellent research assistance. Financial support from ESRC, theSwedish Research Council and Handelsbanken’s Research Foundation is gratefully acknowledged.

169

170 CREDIT CONTRACT STRUCTURE AND FIRM GROWTH

entail sizable indivisible fixed costs in the form of bulky investments such as machines

or buildings. Meanwhile, most debt contracts available to micro-entrepreneurs are de-

signed to reduce lender risk: they stipulate constant repayments and the loan size is

subject to concerns of asymmetric information. The implication is that investment is

distorted toward technologies that involve less learning and are subject to less aggre-

gate risk than otherwise optimal, and avoid indivisible large costs; hampering firm

growth.

Although recent evaluations have found the effect of microfinance on business

growth to be limited, very few studies have studied how debt structure of micro loans

affects business investments and growth by exploring possible changes to the standard

contract. We implemented a randomized control trial designed to measure to what ex-

tent contractual features of the most common contract inhibit the expansion of firms,

and examine contracts amendments that possibly better support firm growth. Together

with the NGO BRAC Uganda’s Small Enterprise Lending Program (SEP), we study

the effect of credit contract terms on small and medium sized firms’ profits, labor and

capital. Using a randomized-control trial methodology we measure whether standard

contractual terms, such as constant and monthly repayments and small initial loan

amounts, are particularly restrictive for firms with (i) backloaded project returns; (ii)

uncertain project returns and; (iii) large, indivisible fixed costs.

The standard contract offered in BRAC’s SEP stipulates monthly repayments dur-

ing the 12 month loan cycle, starting one month after loan disbursement. To investigate

whether the standard contractual terms are restrictive for firms that face indivisible

costs and/or are characterized by backloaded or uncertain project returns, we im-

plemented the following types of interventions for firms approved for BRAC’s SEP

funding: (i) a grace period treatment;1 (ii) a flexible grace period treatment; and (iii)

a cash subsidy treatment. Treatments (i) and (ii) are intended to distinguish the effects

of backloaded returns from those of uncertain project returns while (iii) is expected

to ease the purchase of indivisible goods. These treatments were implemented in the1We define grace period as an additional period, beyond the one stipulated in the standard loan

contract, between the date of disbursement of the loan and the due date of the first loan repayment.


form of repayment rebates covering the equivalent of 2 out of the 12 repayments in a

standard BRAC SEP loan contract (the control group). In addition, we implemented

a late repayment rebate and an alternative control group that received flat rebates,

equivalent of the same share of repayments as the rebates in treatments (i)-(iii), to

account for the income effect caused by granting rebates.

The full sample size in our study will be 2,340 borrowers (clients). In the current

draft, we present findings from the first 754 firms to have completed the loan cycle.

Therefore, we stress that the results presented in this draft are preliminary and can

only provide suggestive evidence about the effects of our treatments.

We find that firms that were given a 2-month grace period increased their prof-

its and household income relative to firms that received a rebate later in the loan

cycle, and to the control groups. They also increased the number of paid employees,

while decreasing the number of unpaid ones, but wage expenditures did not increase

in accordance. Further, the owner households of Early treatment firms started signifi-

cantly more new household-owned firms than the Late rebate and the control groups.

Firms that were offered a Flexible grace period scheme, in which they were free to

skip repayments in any 2 months of their choice, predominantly chose to use these

rebates in the first months of the loan cycle. These findings provide some support for

backloadedness of returns being a more important constraint than the uncertainty of

returns. Firms that received a cash subsidy at the start of the loan cycle increased their

number of employees relative to the control groups, and they also increased their wage

costs. To the extent that this implies that they hired higher quality workers, which

can be seen as an indivisible investment, this finding provides suggestive evidence for

the importance of indivisible costs hampering investments.

This paper contributes to several strands of literature. First it adds to the grow-

ing literature on credit access and use in developing country contexts. A handful of

recently published studies provide the first larger scale randomized evaluations of mi-

crofinance initiatives (Attanasio et al., 2015; Angelucci et al., 2015; Augsburg et al.,

2015; Banerjee et al. 2015b; Crépon et al., 2015; Tarozzi et al., 2015). Taken together,


these studies find no evidence of transformative effects of microfinance on the lives of

the poor, nor do they find positive effects on the extensive margin of business owner-

ship (startups). They do however find modestly positive effects on business outcomes

for already existing micro-businesses (Banerjee et al., 2015a). At the same time, none

of the studies find significant increases in household income or consumption following

the observed positive effects on business activity.

Within this literature, we more specifically contribute to a limited number of studies

of the role of loan contract structure for the profitability and use of loans. We are aware

of only two empirical studies of how changes in the contract terms affect loan use and

efficiency in developing country contexts (Field et al., 2013; Karlan and Zinman, 2008).

Field et al. (2013) randomly offered Indian microfinance borrowers an initial two-month

grace period on their repayments while Karlan and Zinman (2008) introduce a lower

interest rate among previous borrowers of a micro finance institution. Their results

highlight the importance of debt structure but leave a number of questions unanswered.

While the evidence suggests that the grace period allowed for larger initial investments,

it is difficult to disentangle the importance of start-up cost from other factors, such as

the project return path. Our experiment is designed to distinguish between the main

alternative explanations for the findings of these studies, and rather than focusing on

households we focus on existing businesses.2 The study also offers an empirical test of

how some of the central theoretical results about loan contract structure may interact

with the firm’s production function. We discuss these mechanisms in section 2.

We further contribute to the literature on small business growth and return to

capital in SMEs in low income countries. In developing countries both in Africa and

Asia, a large share of the workforce is employed in (both formal and informal) micro

and small sized businesses (Ayyagari et al., 2011). Meanwhile, very few businesses grow.

Recent studies have strived to understand the determinants of and obstacles to business

growth in developing countries by offering cash grants (de Mel et al., 2008; Fafchamps

et al., 2014), business training (Karlan and Valdivia, 2011) and combinations of the two2Although Field et al. (2013) do study firms, their focus on micro sized household businesses.

Karlan and Zinman (2008) sample households, not businesses.

5.2. CONCEPTUAL FRAMEWORK 173

(Bandiera et al., 2016; Fiala, 2013). We add to this literature by our explicit focus on

credit. Returns to capital in the form of giving cash grants may differ from the returns

when capital is given in the form of a loan, precisely because of the implications of the

loan repayment structure. Moreover, the comprehensive survey data that we collect on

firm characteristics related to the firms’ production function, and the fact that we also

have detailed data on owner households, allows us to analyze the interaction of the

production function and the loan contract. Our study also contributes to understanding

the importance of growth constraints as well as drivers of business growth among a

group of small and medium-sized firms who demand loans of greater size than standard

microfinance loans, yet are not large enough to participate in the formal financial

sector.3

The rest of the paper is structured as follows. Section 2 presents the conceptual

framework. Section 3 outlines the experimental design and implementation and section

4 presents the data we collect. In section 5 we present the results, which are further

discussed in section 6. Section 7 concludes the paper.

5.2 Conceptual framework

In this section, we outline the main theoretical mechanisms that our experiment is de-

signed to test for. In particular, we focus on three main features of the firm’s production

function that are likely to be relevant in the contractual design of loan contracts:

1. Backloaded returns to investments: When a firm invests in a new input or tech-

nique, often returns will take time to be realized. It takes time to learn how to

use new inputs or to train new workers, to market new products and to build

a reputation. To the extent that such learning is mechanical and simply time-

consuming, returns to an investment will start low but increase gradually;

2. Uncertain returns to investments: Introducing new products entails a risk if3see Ayyagari et al., (2011) for an overview of the characteristics of small and medium-sized firms

in low-income countries


there is uncertainty about demand. In the presence of market shocks or under

experimentation with new techniques, the timing of revenues from investments

are likely to be uncertain;

3. Indivisible investment costs: Starting or expanding a business may entail sizable

costs in the form of bulky investments such as machines or buildings.

The standard loan contract in microfinance is characterized by regular repayments

that start early in the loan cycle.4 Introducing a grace period, or in other ways making

the repayment scheme more flexible, affects both firms facing backloaded returns to in-

vestment and firms that suffer from uncertain returns to investment. For a borrower,

a more flexible repayment plan may induce higher-return investments either because

such investments have back-loaded returns that take a while to accumulate (and a

flexible plan gives the firm more time to realize the returns before repayments start)

or because there is uncertainty due to idiosyncratic shocks that affect the firm’s re-

turns, and a more flexible repayment plan enables the business owner to smooth such

risks. For example, a grace period allows more time for returns to the investment to

materialize. This is especially relevant if investments are made in a new technology

that requires learning or training employees in using the technology or in new types of

products that need to be advertised and marketed. At the same time, a grace period

also reduces the risk associated with idiosyncratic shocks to profits (e.g. due to mar-

ket conditions). Previous literature has found that the introduction of a grace period

in loan contracts in microfinance leads to higher investments and profits, and higher

default rates (Field et al., 2013). This evidence is consistent both with backloaded

and uncertain returns.5 Our experimental design, explained in detail in the following

section, is designed to shed light on which of these two mechanism may be at work.

On the question of indivisible investment costs, an influential literature shows that

4In related work by Field et al. (2013), the standard period before repayments started was twoweeks. The business loans of our implementing partner BRAC Uganda has a 1 month period betweendisbursement and the first repayment for business loans, whereafter repayments are made monthly.

5In line with the latter mechanism, the authors find a larger effect of the grace period treatmentamong risk averse borrowers.

5.3. EXPERIMENTAL DESIGN 175

the interaction between financial constraints and indivisible investments may give rise

to poverty traps as fixed costs bar the poor from productive activities (Banerjee and

Newman, 1993; Galor and Zeira, 1993; Aghion and Bolton, 1997). On the other hand,

due to the presence of asymmetric information problems in the credit market, loan

size is often constrained and most lenders, including the microfinance institution that

we partner with, enforce borrowing limits on their clients. Whether these limits indeed

imply binding investment constrains is an empirical question, and the empirical evi-

dence on this issue is not conclusive. McKenzie and Woodruff (2006) find that start-up

costs for Mexican micro-enterprises are low in many industries, leading them to reject

the idea that sizable costs hamper investment. However, their sample consists mainly

of small enterprises and they note that indivisibilities may be more relevant for larger

firms. Also, they find evidence of non-convex production technologies in some sectors.

Evaluating a policy experiment that increased credit access in Thailand, Kaboski and

Townsend (2011) document that the program primarily allowed households to under-

take lumpy investments. While this suggests that fixed costs can be important, it is

not clear whether the results apply to firms. In order to examine the importance of

indivisible costs, we will implement a treatment arm in which the size of the loan paid

out to the client will be supplemented with a cash subsidy. This will enable us to test

if access to a larger amount of capital (beyond the amount provided by the lender)

leads to greater investment and business growth.

5.3 Experimental design

The experiment was carried out in collaboration with the NGO BRAC Uganda. BRAC

is a large non-profit organization founded in Bangladesh in 1974, currently active in

12 developing countries in Asia, Sub-Saharan Africa and the Caribbean. BRAC was

launched in Uganda in 2006 and is currently one of the largest development organiza-

tions and micro finance institutions in the country.6 Its core activity is microfinance6BRAC is the largest MFI in Uganda in number of borrowers and the 4th largest in terms of total

loan volume, the only larger providers being the banks Centenary Bank and Equity Uganda and the


which encompasses both microcredit groups and the Small Enterprise Lending (SEP)

program that targets existing business owners.

The loans offered through BRAC Uganda’s SEP program are individual liability

loans offered to business owners. Loans are larger than those given through typical

microcredit arrangements, with a minimum loan size of 2.5 Million UGX, or approxi-

mately 900 USD, and maximum loan sizes following a ladder where first time borrowers

maximum loan size is 8 Million UGX (2,900 USD) while repeat borrowers are allowed

larger loans. The median loan size at the start of our experiment was 3 Million UGX.7

Prospective borrowers are evaluated by BRAC credit officers at the local office closest

to the borrower’s business location. A borrower needs to be the sole business owner,

registered with the tax authorities, a permanent resident of the branch office area

(s)he is applying in, and is not allowed to have outstanding loans with BRAC or other

lenders. In addition, a borrower needs to provide collateral (land) amounting to the

value of the loan and two guarantors who will be responsible for the repayments in

case the borrower fails to repay the loan. A loan cycle lasts for 1 year during which

the loan is paid off in 12 equal installments at an annual interest rate of 25 percent.

We collaborated with the SEP program in 76 local branch offices in Central, West-

ern, and Eastern Uganda. From November 2014, clients that were approved for a loan

in any of these 76 offices, and belonging to one of the business sectors we had pre-

selected, were enrolled into the experiment. Upon enrollment, local research officers

administered a baseline survey to the client, collecting information about business, in-

dividual and household characteristics. Data was collected electronically using tablets,

and after the baseline interview was submitted and received by our central research

team in Kampala, the client was assigned to one out of the five treatment groups, or

to the control group. The local research officer was then informed of the treatment

and - in case the client was not assigned to the control group - met again with the

credit union/cooperative TBS. (Mixmarket, 2016)7Using the nominal exchange rate at the launch of the experiment (November 1, 2014) the mini-

mum SEP loan size corresponded to 912 USD, the median to 1,100 USD and the maximum for firsttime borrowers corresponded to 2,900 USD. Corresponding real values, using the World Bank PPPadjusted exchange rate for 2014, are 2450 USD, 2,940 USD and 7,840 USD, respectively.


client to explain the altered contract terms.

The borrowers were randomly allocated into one of the following 6 groups:

T1. "Early repayment voucher": Firms in treatment group 1 were allowed to skip

the first 2 repayments (months 1-2 in the 1-year loan cycle);

T2. "Late repayment voucher" firms in treatment group 2 were allowed to skip the

last 2 repayments (months 11-12 in the 1-year loan cycle);

T3. "Flexible repayment voucher" firms in treatment group 3 were allowed to skip

any 2 repayments of their own choosing (any 2 months in the 1-year loan cycle);

T4. "Flat repayment voucher" firms in treatment group 4 received rebates on all

repayments such that the total loan repayment over the 1-year loan cycle is

equivalent to the total repayment in each of treatments T1-T3;

T5. "Subsidy voucher" firms in treatment group 5 received a cash grant equivalent

in value to 2 repayments (one sixth of the principal plus the interest payment).

This grant was paid to the firms on the same day as their loan disbursement (i.e.

at the beginning of their loan cycle);

C. "Control" firms in the control group received the standard BRAC loan contract

as described above.

The experiment was designed to introduce exogenous variation in contractual terms

that would enable us to test for the relative importance of the theoretical mechanisms

discussed in Section 2. In particular, treatment arms 1-3 introduce repayment rebates

at different points in time to disentangle the effect of backloaded returns due to deter-

ministic learning, from that of uncertain returns. Under the assumption of independent

and identically distributed negative sales shocks occurring during the 1-year loan cy-

cle, firms will benefit from rebates at any time if uncertain returns to investment is a


constraint. Meanwhile, firms suffering from problems of backloaded returns to invest-

ments will benefit relatively more from rebates early in the loan cycle. An early rebate

will ease the repayment burden in the beginning, when the firm has not yet started

earning the full returns on the investment. This implies that the firm will be able to

take on more backloaded investments if they are, through the early treatment, offered

a grace period. If we observe greater returns to T1-firms (Early rebates) relative to T2

(Late rebates), this supports the backloaded returns channel.

The firms in T1 start repaying 3 months after the loan is disbursed to them, and

finish repaying in month 12, while firms in T2 start repaying 1 month after disburse-

ment and finish in month 10. This setups allows for a straightforward comparison of

our findings to those of Field et al., (2013), who introduced a 2 month grace period to

microfinance clients, in which all repayments were shifted 2 months ahead in time. Our

discounting adjusted treatments (explained in more detail below) and our additional

treatment arms T3 (Flexible rebate scheme) and T5 (Subsidy), will allow us to better

disentangle the mechanisms behind their findings.

If we observe all firms in T3 (Flexible rebate scheme) opting to use their rebates

for the first two repayments, this would also be in line with a backloaded returns

mechanism since it shows that firms prefer to have the rebates early in the loan cycle.

On the other hand, if we observe T3 firms using the rebates at different points in their

loan cycle (instead of the first two repayments), this would be in line with uncertainty

playing a more important role.8

One challenge with the above argument is that firms in treatment groups T1, T2,

and T3 have an overall lower repayment burden than the control group, due to the

dynamic rebates. The rebates are subsidizing the total amount that the client owes

BRAC by the equivalent of 2 repayments. In order to account for this change in

repayment burden, we introduce a Flat repayment rebate over the full loan cycle in

8Even if we observe firms in T3 using their rebates early on in their loan cycle, this in itself is notsufficient for us to claim that backloaded returns are more important than uncertain returns, sincethis may also be due to the business owners having self-control problems. To further investigate thiswe will, once the full data is available, make use of survey modules on time preferences and householdconsumption.


treatment arm 4. The value of the Flat repayment rebate is equivalent, in its share of

the total loan size, to the rebates in T1-T3. The difference is that in T4, the value of

the rebate is to be subtracted equally from each of the 12 repayments such that the

only difference between T4 and the control group is a constant (level) difference each

month. T4 thus provides an alternative control group that accounts for the income

effects caused by the other treatments.9

The Subsidy treatment, T5, was designed to measure the importance of indivisible

or bulky investment costs. If we observe that firms in T5 make types of investments

that are more bulky in nature than firms in the control group or in T4 (the Flat rebate

scheme), this provides support that indivisible fixed costs is an important constraint.

In addition, by comparing T5 to T1 (the Early rebate scheme) we can separate between

backloaded returns and indivisible investment costs channels.

Every firm in the sample is a BRAC borrower that has been approved by BRAC

credit officers and is eligible for a loan. Upon being eligible, firms are then randomized

into one of the treatment (and control) arms. Randomization was implemented at

the individual client level across 76 branches throughout Uganda. The randomization

was stratified (Bruhn and McKenzie, 2009) by region10, sector (manufacturing vs.

retail), and previous experience with BRAC SEP loans (new vs. repeat borrower).

Furthermore, firms that entered into the sample consecutively within a stratum were

assigned blocks and the randomization was conducted within each block. In this way,

we effectively stratify the sample by the order in which firms enter into our sample

9To test for the effect of discounting, we implemented a cross-cutting treatment involving cashtransfers to cover inflation costs, which we will refer to as the "Discounting adjusted treatment".Within each treatment group, 50% of the firms were randomly selected to receive monthly cashtransfers, making the present discounted value of their rebates equivalent to the Subsidy treatment(T5) where no discounting was necessary (since the cash grant was paid at the beginning of the loancycle) assuming 10% annual discount rate. The clients assigned to receive this additional discountingadjustment were transferred a monthly payment via mobile money, every month, the size of whichdepended on the timing of their rebates. For example, firms in T1 were transferred a cash transfer inmonths 1 and 2, making the present discounted value of the total subsidy they received the same ashaving received the cash grant at the beginning of their loan cycle (as in T5).

10Each local BRAC office ("branch") is served and supervised by a BRAC Area Office and some ofthe lending activities of the local offices in a given Area offices are overseen by the same Area officestaff. The 76 branches in our study are grouped into 15 such Areas.


within an area, sector, and previous loan experience. The sample size (390 firms per

treatment group and 390 for the control group) is based on calculations to detect a

0.20 S.D. shift in key outcomes with 80 percent power under a 95 percent confidence

level. This implies that our current preliminary analysis with only 754 borrowers is

under-powered, making our results suggestive in nature.

The treatments were designed such that the value of the repayment voucher pro-

vided each firm with a subsidy, or rebate equivalent to two monthly installments. In

other words, without discounting, the value of the subsidy provided in T1-T5 is one

sixth of the total amount (principal plus interest) that the firm owed BRAC.

The rebates were implemented by giving the clients vouchers to be used instead of

repayment for certain installments. Client specific vouchers were prepared by the cen-

tral team in Kampala and sent to the local branch office where they were distributed

to the client by our locally based research officer, the week after the baseline interview

had been conducted with a given client. The validity date of the vouchers differed

depending on the client’s disbursement date and the treatment group that the client

was assigned to and their value was set according to each client’s loan size and corre-

sponding installment size. The Subsidy treatment clients (T5) received one voucher,

the Flat treatment clients (T4) received 12 vouchers while treatment groups T1-T3

received two vouchers each.

5.4 Data

5.4.1 Baseline and endline survey

Before assigning them to a treatment, we collected baseline data on each client. The

baseline survey instrument includes detailed information about the firm’s history and

present firm operations, as well as about individual and household characteristics of

the business owner/client. The business operation data includes information about

sales and profits volumes, expenditures, the level and type of business assets held at

baseline, and the number and type of workers employed in the business. Individual

5.4. DATA 181

characteristics of the business owner include gender and education as well as measures

of financial literacy, time and risk preferences. We also collect comprehensive data

on the business owner’s household, including household size, income, data on other

household members and their occupations, and other business activities run by the

household. Upon completion of the client’s loan cycle, 12 months after the baseline

survey, an endline survey was conducted, collecting the same type of information that

was included in the baseline.

5.4.2 Business diaries

In addition to the baseline and endline surveys, bi-monthly business diaries were con-

ducted with all the clients in our study. Apart from providing us with regular data on

sales and profits, the diaries focused on changes made in the business, as well as in

the owner’s household, between each visit of the interviewer. This high-frequency data

collection allows us to observe any adjustments made following the disbursement and

use of the loan, as well as in connection to the use of vouchers. In the current version

of the paper we only focus on the baseline and endline survey data, but once the full

data is available the business diary data will provide additional detail to the analysis.

5.4.3 Summary statistics and randomization balance check

Table 5.1 displays summary statistics and balance checks for the entire sample for

some of the central outcome variables of interest, measured at the baseline, with the

Flat treatment as the reference category. The first column shows the mean and stan-

dard deviation for a variable among the clients in the Flat treatment group, and the

following 5 columns show results from a regression of the variable on indicators for the

other treatment groups, and the control group. All monetary values were measured

in Ugandan Shillings (UGX) and then deflated to their October 2014 value and con-

verted to PPP adjusted US dollars.11 The average firm is around 8 years old and has

11The values were deflated to the October 2014 value using the CPI values from Bank of Uganda.October 2014 was chosen since this was the month when the data collection was started, and the


1.32 employees, and the average asset value lies around 16,000 in PPP adjusted US

dollars. 44% of firm owners are female, and the average education of a firm owner is

10 years which corresponds to the lower of two secondary school degrees in Uganda

(O-level). The businesses in this study are thus larger on average than those in most

related studies that have focused on household enterprises with no employees. Over-

all, the sample looks balanced. The two last rows of Table 5.1 reports the F-statistic

and P-value of a joint test of orthogonality of the full set of variables to each of the

treatment dummies. Reassuringly, none of these tests are significant.

A majority of the firms whose baseline summary statistics are shown in Table 5.1

have entered the sample more recently than the subsample we use for the preliminary

analysis, and we therefore do not have endline data on them yet. In the present version

of the analysis, we will include all business owners who had completed their loan

cycle by May 25, 2016.12 Table A.1 in the Appendix shows summary statistics and

balance checks for this sub-sample. There are 754 such businesses. This preliminary

sub-sample is slightly less balanced across treatments. This is to be expected given

the smaller sample. In particular, there are some significant differences in the types

of employees, with Early treatment firms having, on average, fewer paid employees at

baseline and the Flexible treatment firms having more paid employees than firms in

the Flat treatment category. The Flexible and the Subsidy groups have fewer unpaid

employees than the Flat comparison group. To account for these differences, we control

for the baseline value of the outcome variables in all the regressions. However, as above,

in the joint tests, orthogonality of the full set of variables to each of the treatments

cannot be rejected.

deflation makes baseline surveys carried out in different months and years more comparable. Forconversion to USD we use the 2014 average rate of the World Bank PPP adjusted exchange rate forUganda.

12By completing the loan cycle we mean that 1 year has passed since the disbursement date,implying that the endline interview as well as the final repayment installments should have beencompleted.

5.5. RESULTS 183

5.4.4 Attrition

Table A.5 reports attrition shares by treatment, with attrition defined as not having

been interviewed for the endline, despite being past the due date of performing the

endline survey on May 25, 2016 which is the cutoff date for the current sample. In total,

731 out of the 754 firms completed the endline survey. Table A.6 presents the results

from a regression of the treatments on a dummy variable for attrition. Attrition results

of all treatments are shown relative to the Flat treatment group. None of the treatment

groups, nor the control group have significantly different attrition rates from the Flat

treatment group, with the exception of the Late treatment group where the attrition

rate is significantly lower than for the Flat treatment. The p-values at the lower panel

show that the Late treatment also has significantly lower attrition compared to the

Subsidy treatment and the control groups (p-value of 0.011 and 0.031 respectively).

Results focusing on the Late treatment group should therefore be interpreted with

some caution.

5.5 Results

In this section, we present preliminary results for the first 731 firms that finished their

loan cycle within the experiment. Since the Flat treatment group, contrary to the con-

trol group, adjusts for the income effect introduced by any treatment, our specification

uses the Flat treatment as the reference group. In each result table we display the p-

values of pairwise test of equality between the other groups to facilitate comparison

between any pairs of treatments, including comparisons between each treatment group

and the pure control group clients.

5.5.1 Empirical specification

We estimate the following linear regression model, where the Flat treatment group

(T4) is the reference category:


Yi1 = α +β1T 1+β2T 2+β3T 3+β4T 5+β5C+σYi0+ηs + εi1, (5.1)

The outcome variable Yi1 is a characteristic of the business or of the owner’s house-

hold, such as profit value, asset value or household income, measured at the endline.

Following McKenzie (2012), we control for the baseline level of the outcome, Yi0, in

order to improve the efficiency of the estimated treatment effect. T1, T2, T3 and T5

are the Early, Late, Flexible and Subsidy treatments, respectively, and C is the pure

control group. The omitted treatment category is the Flat treatment (T4). ηs is a

stratum fixed effect. The coefficients of interest are β1 to β4 (and β5), indicating the

effect of treatment T on the outcome variable relative to the Flat treatment group.

All regressions are estimated using robust standard errors.

5.5.2 Key business and household outcomes

In this section, we examine the effect of the treatments on the level of business oper-

ations, as well as on economic status measures of the firm owners’ household.

Tables 5.2 and 5.3 show results for key economic variables of the business or the

business owner’s household. In addition to estimating effects on the deflated and PPP

adjusted monetary values, we also report results for three dummy variables for high

profits, sales and household income. A value=1 indicates that the value of the corre-

sponding monetary variable lies above the highest threshold of intervals that we used

in the questionnaire to measure the profits/sales/household income for respondents

who were not able to recall the precise value. Table 5.2 displays the findings for prof-

its, sales, and expenditures of the business. Controlling for baseline differences, the

profits at endline are significantly higher for businesses in the Early treatment group

than for those in the Flat treatment, the Late treatment and the control group. Com-

pared to the mean profit level among Flat treatment firms of 10,273 PPP adjusted US

dollars, profits in the Early treatment firms are on average 50% higher, while they are

on average 53% higher than in the Late treatment firms. However, the fact that we

5.5. RESULTS 185

find no significant result for the high profit dummy suggests that the result for profits

may be driven by a small number of successful firms in the Early treatment group.13

Point estimates for the other variables in Table 5.2 are imprecise and with the current

sample, it is hard to draw conclusions about effects on sales and costs.

Table 5.3 shows the results for types of employees and asset value in the business

that the loan was officially intended for, as well as for household asset value and income.

Column 1 of Table 5.3 indicates that, controlling for the baseline number of employees,

the number of paid employees is higher at endline for all treatment groups relative to

the Flat treatment group or to the pure control group. This difference is significant

at least at the 95 percent confidence level for the Early, Late and Subsidy treatments,

and approaching conventional levels of significance for the Flexible treatment group

(p-value 0.108). In terms of magnitudes, the number of paid employees are 36.6%

higher in the early than in the Flat group, it is 33.7% higher in the Late than in

the flat, and 27.7% higher in the Subsidy treatment than among Flat treatment group

clients. None of the differences between the treatments (excluding Flat) are significant.

Meanwhile, the endline number of unpaid employees in the Early treatment group is

36% lower than in the Flat treatment group, and a similar pattern is seen for the

three other treatment groups: Late, Flexible, and Subsidy. Controlling for the baseline

values, firm owner household income is higher at endline in the Early treatment group

than in the Flat, Late, and Subsidy treatment groups. These findings are significant

at least at the 95 percent confidence level, both when measured as a monetary value

(column 5), or as the likelihood to be above the "high income" threshold (column 6).14

To account for the fact that the real value of a repayment voucher depends on the

timing of its use due to inflation, we randomly assigned 50% of the firms within each

treatment group to receive monthly cash transfers, making the present discounted value

13Indeed, in Table A.2 in the Appendix, we show the regression results for the monetary variables,with the sample trimmed for outliers that lie above the 99th percentile of the distribution of thevariable. In this trimmed sample we do not observe significantly higher profits for our Early rebatetreatment firms.

14The finding for household income is robust: also in the trimmed sample in Table A.2 we observesignificantly higher household income in the Early rebate treatment group than in the Flat, Late,and Subsidy treatment groups, and the control group.


of their rebates equivalent to the Subsidy treatment (T5) where no discounting was

necessary (since the cash grant was paid at the beginning of the loan cycle), assuming

a 10% annual discount rate. Tables A.3 and A.4 in the Appendix show the results for

the same outcomes as in Tables 5.2 and 5.3, broken down by whether or not a client

was assigned to this cross-cutting discounting adjustment treatment. For three out of

the four variables where we found significant differences between the treatment groups

in the pooled sample, point estimates are similar across the two sub-groups. For the

number of unpaid employees we observe a difference, with point estimates being more

negative in the non-discounting adjusted group. This can be due to the small sample

size in our preliminary sample.

5.5.3 Additional outcomes

In the previous subsection, we found effects on the types of employment in the firm

and on household income in the absence of any clear effects on profits or sales of the

business. In this subsection, we take a closer look at expenditures connected to em-

ployment and at other income-generating activities of the business owner’s household.

In small businesses such as the ones we study, the distinction between the finances

of the business and those of the business owner’s household is often vague. Previous

studies of microfinance have found direct effects of business loans on spending in the

business owner’s/borrower’s household. One reason for why treated businesses reduce

their unpaid labor and increase the paid labor may be that they switch from employ-

ing unpaid household members to hiring externally employed workers. To investigate

this, in Table 5.4 we show results from regressions using equation 5.1 on three addi-

tional outcomes that can help rationalize our results for employment and household

income. Specifically, we examine the effect of the treatments on wage expenditures

(in the borrowing firm) and additional businesses owned by members of the owner’s

household. Subsidy treatment firms, for whom we saw a significant increase in paid

labor relative to the Flat treatment, also have significantly higher wage costs than all

the other treatment groups, and the control group. Compared to the average wage

5.6. DISCUSSION 187

costs among Flat treatment firms of 818 PPP adjusted US dollars, wage costs in the

Subsidy treatment firms almost doubled, or increased by 781 PPP adjusted US dol-

lars, an effect significant at the 95 percent confidence level. Early treatment firms, for

which we saw the most pronounced reduction in unpaid labor and increase in paid la-

bor, have a positive point estimate for wage costs, but it is not significant. Turning to

the effect on other businesses owned by the firm owners’ households, Early treatment

firms, for whom we saw large positive effects on household income relative to the Flat

reference group, also have significantly more household owned businesses, controlling

for the number of businesses owned at baseline (column 1) and has started more new

household businesses during the loan cycle than owner households in the Flat treat-

ment group (column 3).

To sum up this section, we observe an increase in profits in the Early rebate treat-

ment group, but it appears to be driven by a small number of firms. Even in the

trimmed sample, we observe increases in paid employment at the expense of unpaid

employment, and we also find increases in the household income, along with increases

in wage costs for the Subsidy treatment group and an increase in the number of

household-owned firms, especially among firm owners in the Early treatment group.

In the next section, we discuss potential mechanisms behind these results.

5.6 Discussion

In this section, we discuss the results in relation to the constraints to business growth

that the experiment is designed to test. More specifically, we discuss what the results

presented in the previous section reveal about the relevance of backloaded returns and

indivisible costs, respectively.


5.6.1 Backloadedness of returns vs. uncertain returns

Early (T1) vs. Flexible (T3) rebate treatment. The Flexible repayment voucher

treatment was introduced as a comparison to the Early repayment voucher. Examining

the timing of voucher use for the clients in the Flexible treatment helps us understand

whether skipping payments in a certain part of the loan cycle is particularly desirable to

clients, and comparing the outcomes between treatments we can understand whether

skipping repayments early in the loan cycle is more beneficial. Table 5.5 shows the

timing of voucher use for the Flexible clients. A majority of Flexible treatment clients

chose to use their vouchers within the first 2 months: 71 percent of clients used the first

voucher at the time of the first installment, and 63 percent used the second voucher at

the second installment occasion. It thus appears as if the borrowers preferred skipping

repayments early in the loan cycle, which supports the importance of backloaded

returns. Moreover, while the Flexible treatment does not yield significantly different

profits compared to the Early rebate treatment, it does yield slightly lower profits

and yield significantly lower household income. This also suggests that backloaded

returns may be more important than uncertainty. The fact that firm owners prefer to

use vouchers early could, however, also be a sign of self-control problems. To further

investigate this we will, once the full data is available, make use of survey modules on

time preferences and household consumption.

Early (T1) vs. Late (T2) rebate treatment. As discussed in section 3, treat-

ments T1 and T2 both stipulate repayments for ten consecutive months, but for the

Early rebate treatment (T1) these are shifted two months ahead in time so that repay-

ment starts after 3 months, while for T2, repayments begin after 1 month. This allows

for a comparison with Field et al. (2013). If we observe greater returns to T1-firms

(Early rebates) relative to T2 (Late rebates), this would support the importance of

the backloaded returns channel. We do find that endline household income is higher

for Early treatment firms than for the Late treatment group. Furthermore, when con-

trolling for the baseline number of household owned businesses, the households of firm

owners in the Early treatment group have a higher number of household owned busi-

5.6. DISCUSSION 189

nesses at endline than the Late treatment group. Thus, there is supportive evidence

that backloaded returns matters more than uncertain returns. These results are in line

with those of Field et al. (2013), who also found greater profits and a higher likelihood

of business formation among their grace period clients. Note, however, that since attri-

tion rates are significantly lower in the Late treatment than in both the Flat reference

group and the Early treatment in this preliminary dataset, any results that include

this treatment group should be interpreted with extra caution.

5.6.2 Indivisibility of costs

Subsidy (T5) vs. Flat (T4) treatment. The Subsidy treatment (T5) was designed

to measure the importance of indivisible or bulky investment costs. To properly inves-

tigate the impact of the subsidies, we would like to examine the timing and types of

investments made in the firms over the entire loan cycle which requires access to the

business diary information unavailable at the time of the current draft. However, the

existing data offers some evidence that the subsidies may have enabled the firms to

undertake indivisible investments otherwise out of reach.

More precisely, the Subsidy firms doubled the wage-related expenditures while

slightly increasing the size of their paid workforce, thus substantially increasing the

wages of their paid employees. Interestingly, the wage costs only rise significantly

in the Subsidy group as compared to all the other treatments (an effect statistically

significant at least at the 90 percent confidence level across treatments). This indicates

that the Subsidy firms may have hired more skilled (and expensive) workers on the

margin.15 To the extent that contracting better-qualified employees requires firms to

make more binding and long-term commitments, this scaling up is a significant fixed-

cost constraint that the Subsidy can help alleviate. Once the business diary data

becomes available, we will further investigate if the costlier workers are matched by

investments in bulkier technology that perhaps require a higher skill level.15The smaller increase in household income offers further indirect evidence that the Subsidy treat-

ment induced increased activity within the respondent’s firm rather than an expansion into otheractivities raising household income.


5.7 Conclusion

This paper presents the preliminary results of a randomized control trial designed to

evaluate the effect of loan contract structure on investment outcomes for borrowing

micro, small and medium sized enterprises. After the one year loan cycle we find no

conclusive effects on the overall profits or revenues of treated businesses compared to

the control groups. We do, however, observe significant effects on the allocation of

labor. Moreover, we find effects on the household income from some of the treatments.

In terms of mechanisms we find suggestive evidence that returns seem to be backloaded

rather than uncertain, while we also find support for the indivisible costs channel.

In particular, firms in the Early treatment group increased their profits and house-

hold income relative to firms in the Late and the Flat treatment and the control

group. Relative to the Flat treatment, the Early treatment firms also changed their

allocation of hired labor between paid and unpaid employees: their number of paid

employees at endline is 36.6% higher than in the Flat rebate group, while the endline

number of unpaid employees in the Early treatment group is 36% lower than that

in the Flat treatment group. Wage expenditures for the Early treatment firms did,

however, not change in accordance. Further, the owner households of Early treatment

firms started significantly more new household-owned firms than the Late and Flat

treatment, as well as the control group. Firms in the Flexible treatment predominantly

chose to use their two rebates in the first months of the loan cycle. All these findings

provide support for the backloadedness of returns being a more important constraint

than the uncertainty of returns. Turning to the indivisible fixed cost channel, firms in

the Subsidy treatment increased their number of employees relative to the Flat treat-

ment group and the control group: the number of paid employees is on average 27.7%

higher in the Subsidy treatment than among Flat treatment group clients, and the

Subsidy firms also significantly increased their wage expenditures. To the extent that

this implies that they hired higher quality workers, which can be seen as an indivisible

investment, this finding provides suggestive evidence for the importance of indivisible

5.7. CONCLUSION 191

costs.

To better understand the channels behind the observed effects on employees and

profits, it is crucial to examine data on specific investments carried out by the firms in

our experiment, and the degree of irreversibility of those investments. The bi-monthly

business diaries will enable us to do this. Examining the impact on the borrower’s

household in more detail is also important as benefits of the loan may accrue to

the household. Once the full data is available to us we plan to address these points.

The preliminary findings on the types of labor employed by treated firms, and the

connection to increases in the number of household owned businesses gives a first

indication that some of the effect of the loan may show up in other parts of the

household income and activities, rather than in the business that the loan was officially

intended to benefit.

We want to stress that results presented here are preliminary and should be inter-

preted with caution. Future analysis using the complete experimental data will enable

us to draw more precise conclusions regarding the effects of our treatments.

Given the scarcity of empirical work investigating the interaction of firms’ financial

structure and their production technology, this paper and project provides unique

evidence on the constraints governing firm behavior, complementing the micro finance

literature’s previous emphasis on access to finance. The findings are relevant both

to academia and policymakers. A better understanding of the economic impact of

debt contract design will provide insights to entrepreneurial behavior in developing

as well as developed markets. Also, shedding light on contractual mechanisms that

affect the size and profitability of business investment is important to guide credit

policies aimed at boosting firm growth, especially given the large sums spent by the

development community on credit programs.


References

Aghion, P. and Bolton, P. 1997. A theory of trickle-down growth and development.

The Review of Economic Studies, 64(2), pp.151-172.

Angelucci, M., Karlan, D. and Zinman, J. 2015. Microcredit impacts: Evi-

dence from a randomized microcredit program placement experiment by Compartamos

Banco. American Economic Journal: Applied Economics, 7(1), pp.151-182.

Attanasio, O., Augsburg, B., De Haas, R., Fitzsimons, E. and Harmgart, H.

2015. The impacts of microfinance: Evidence from joint-liability lending in Mongolia.

American Economic Journal: Applied Economics, 7(1), pp.90-122.

Augsburg, B, De Haas, R., Harmgart, H. and C. Meghir 2015. The Impacts

of Microcredit: Evidence from Bosnia and Herzegovina. American Economic Journal:

Applied Economics, 7(1), pp 183-203.

Ayyagari, M., Demirgüç-Kunt, A. and Maksimovic, V. 2011. Small vs. young

firms across the world: contribution to employment, job creation, and growth. World

Bank Policy Research Working Paper, (5631).

Bandiera, O., R. Burgess, N. Das, S. Gulesci, I. Rasul, and M. Sulaiman,

(2016). Labor Markets and Poverty in Village Economies, Mimeo, Bocconi University.

Banerjee, A.V. and Newman, A.F. 1993. Occupational choice and the process of

development. Journal of Political Economy, pp.274-298.

Banerjee, A., Karlan, D. and Zinman, J. 2015.a. Six randomized evaluations

of microcredit: introduction and further steps. American Economic Journal: Applied

Economics, 7(1), pp.1-21.

Banerjee, A., E. Duflo, R. Glennerster, and C. Kinnan 2015.b. The Miracle of

Microfinance? Evidence from a Randomized Evaluation. American Economic Journal:

Applied Economics, 7(1): 22-53

Bruhn, M. and McKenzie, D. 2009. In pursuit of balance: Randomization in

REFERENCES 193

practice in development field experiments. American Economic Journal: Applied Eco-

nomics, 1(4), pp.200-232.

Crépon, B., E. Duflo, F. Devoto, W. Pariente. 2015. Estimating the impact

of microcredit on those who take it up: Evidence from a randomized experiment in

Morocco, American Economic Journal: Applied Economics, Vol. 7, No. 1.

de Mel, S., McKenzie, D. and Woodruff, C. 2008. Returns to capital in mi-

croenterprises: evidence from a field experiment. The Quarterly Journal of Economics,

pp.1329-1372.

Fafchamps, M., McKenzie, D., Quinn, S. and Woodruff, C. 2014. Microen-

terprise growth and the flypaper effect: Evidence from a randomized experiment in

Ghana. Journal of Development Economics, 106, pp.211-226.

Fiala, N., 2013. Stimulating microenterprise growth: Results from a loans, grants and

training experiment in Uganda. Grants and Training Experiment in Uganda (Decem-

ber 4, 2013).

Field, E., Pande, R., Papp, J. and Rigol, N. 2013. Does the classic microfinance

model discourage entrepreneurship among the poor? Experimental evidence from In-

dia. The American Economic Review, 103(6), pp.2196-2226.

Galor, O. and Zeira, J. 1993. Income distribution and macroeconomics. The Review

of Economic Studies, 60(1), pp.35-52.

Kaboski, J.P. and Townsend, R.M. 2011. A Structural Evaluation of a Large-Scale

Quasi-Experimental Microfinance Initiative. Econometrica, 79(5), pp.1357-1406.



Karlan, D. and Valdivia, M., 2011. Teaching entrepreneurship: Impact of business

training on microfinance clients and institutions. Review of Economics and Statistics,

93(2), pp.510-527.


McKenzie, D.J. and Woodruff, C. 2006. Do entry costs provide an empirical basis

for poverty traps? Evidence from Mexican microenterprises. Economic Development

and Cultural Change, 55(1), pp.3-42.

McKenzie, D. 2012. Beyond Baseline and Follow-up: The Case for More T in Ex-

periments. Journal of Development Economics 99: 210-21.

Mixmarket. 2016. Uganda market overview, retrieved on 22 May, 2016,

https://www.themix.org/mixmarket/countries-regions/uganda.

Tarozzi, A., Desai, J. and Johnson, K. 2015. The impacts of microcredit: Evidence

from Ethiopia. American Economic Journal: Applied Economics, 7(1), pp.54-89.

TABLES 195

Tables


Table 5.1: Summary statistics/balance check, full sample

Differences between flat and other treatmentsFlat Early Late Flexible Subs. Contr. All

Mean Coeff. Coeff. Coeff. Coeff. Coeff. N(s.d.) (s.d.) [st.err.] [st.err.] [st.err.] [st.err.]

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

Female owner 0.436 -0.025 -0.013 -0.003 -0.012 -0.035 2339(0.497) [0.035] [0.035] [0.034] [0.034] [0.035]

Years educ. owner 10.038 -0.178 0.048 -0.214 -0.188 0.094 2330(4.042) [0.283] [0.283] [0.281] [0.281] [0.284]

Profit past year 10550.227 -667.209 1025.266 566.522 826.768 973.935 2339(15941.594) [1254.431] [1253.788] [1247.646] [1246.067] [1256.517]

High profits=1 0.196 0.005 0.053∗ 0.027 0.031 0.005 2339(0.398) [0.029] [0.029] [0.029] [0.029] [0.029]

Sales past year 54432.556 6110.512 9979.132 8053.571 1966.017 3332.859 2339(85546.21) [6942.854] [6939.296] [6905.302] [6896.562] [6954.397]

High sales=1 0.173 0.025 0.015 0.013 0.012 0.017 2339(0.379) [0.027] [0.027] [0.027] [0.027] [0.027]

Costs past year 37305.61 15986.543∗∗∗ 4917.341 9320.744 4756.741 5837.885 2339(58449.136) [5853.241] [5850.242] [5821.583] [5814.215] [5862.973]

Wage costs 1094.88 -30.328 187.19 350.965 328.56 595.083∗∗ 2332(3473.682) [291.107] [290.951] [289.518] [289.164] [291.803]

Business capital 15762.831 -622.906 -1585.965 1595.017 1757.667 -1823.75 2327(32901.8) [2359.525] [2362.812] [2346.646] [2345.178] [2368.215]

Age of business (years) 8.429 -0.671 -0.303 -0.556 -0.474 -0.621 2053(7.81) [0.533] [0.533] [0.530] [0.529] [0.534]

#Employees 1.319 -0.142 -0.02 0.06 -0.108 -0.09 2339(1.279) [0.095] [0.095] [0.094] [0.094] [0.095]

# paid empl. 1.14 -0.165 0.023 0.16 -0.072 -0.074 2339(1.449) [0.102] [0.102] [0.102] [0.102] [0.103]

# unpaid empl. 0.309 -0.083∗∗ -0.028 -0.078∗∗ -0.076∗ -0.017 2339(0.666) [0.040] [0.040] [0.040] [0.040] [0.040]

HH asset value 95677.437 1994.468 4301.542 10211.520∗∗ 1878.728 -574.975 2339(124946.48) [4898.733] [4896.223] [4872.237] [4866.070] [4906.877]

HH inc. past year 15721.238 -407.954 -329.863 -2054.335 -1037.384 -911.889 2327(27156.714) [1577.922] [1578.108] [1571.405] [1567.426] [1582.792]

High HH income=1 0.3 -0.02 0.015 -0.021 0.003 -0.031 2327(0.459) [0.030] [0.030] [0.030] [0.030] [0.031]

# HH businesses 0.482 0.062 0.065 -0.026 -0.053 0 2339(0.701) [0.048] [0.048] [0.047] [0.047] [0.048]

F stat joint test 0.97 1.17 0.85 0.88 0.64 0.8P-value joint test 0.510 0.254 0.672 0.642 0.915 0.75

Note: Column 1 shows the descriptive statistics for the flat treatment group (the reference group). The remaining columns show the difference betweeneach additional treatment group and the flat treatment group. The last column shows the number of observations for the full sample (all 5 treatments andthe control group). Coefficients and standard errors are from OLS regressions of each variable on the Early, Late, Flexible and Subsidy treatment, and thecontrol group, controlling for stratification. The two last lines report the F-statistic and p-value from a joint test of the significance of the set of variablesin explaining each treatment dummy. Business capital Value of business assets, including inventory and excluding the value of land and buildings, in PPPadjusted USD. HH asset value Value of household assets, excluding land, in PPP adjusted USD. #HH businesses the number of other businesses ownedby members of the borrower’s household. Robust standard errors are reported in square brackets. All regressions control for stratum fixed effects. * p<0.1,** p<0.05, *** p<0.01.

TABLES 197

Tab

le5.2:

Businessou

tcom

esI

Profitspa

styear

Highprofi

ts=1

Salespa

styear

Highsales=1

Costs

past

year

(1)

(2)

(3)

(4)

(5)

Early

5097

.287

*0.01

5-163

2.810

0.01

018

94.892

[281

5.19

5][0.048

][110

39.531]

[0.044

][938

4.91

9]La

te-609

.306

-0.036

-276

9.57

3-0.015

1159

8.74

1[277

5.76

4][0.048

][108

80.850]

[0.043

][922

8.07

7]Flexible

991.327

-0.025

-104

39.586

-0.033

-168

8.66

1[276

2.80

8][0.047

][108

32.717]

[0.043

][918

7.89

5]Su

bsidy

2132

.773

-0.012

-646

4.64

9-0.018

-588

7.54

2[269

3.05

0][0.046

][105

60.095]

[0.042

][896

7.10

4]Con

trol

-637.817

-0.063

-107

16.653

-0.046

-127

49.150

[286

6.50

8][0.049

][112

38.272]

[0.045

][953

4.44

3]

pearly=

flex

0.147

0.40

90.42

70.33

00.70

4pearly=

late

0.04

40.28

90.91

90.57

20.30

4pearly=

subs.

0.28

50.55

90.65

70.52

00.39

9pearly=

control

0.05

00.11

90.42

80.22

20.13

2pflex=

late

0.56

60.81

20.48

40.67

90.15

3pflex=

subs.

0.67

40.79

00.70

80.71

90.64

2pflex=

control

0.57

20.44

30.98

00.77

90.24

9psubs.=

late

0.31

80.61

30.73

10.94

80.056

psubs.=

control

0.32

80.30

00.70

20.52

90.466

plate

=control

0.99

20.593

0.48

50.49

80.01

2

Meanfla

t10

273.46

0.24

264

529.24

0.18

039

005.08

Stratum

FE

Yes

Yes

Yes

Yes

Yes

AdjustedR

20.13

10.20

70.28

10.22

30.15

1Observation

s73

173

173

173

173

1Note:

Profits/Sa

les/Costs

past

year

areexpressed

intheirOctob

er2014

value,

asPPP-adjusted

US

dolla

rs.Highprofi

ts=1(H

ighsales=1):

Dum

mies=

1ifthevalueof

repo

rted

profi

ts(sales)lie

abovethehigh

estthresholdof

theintervalsthat

weused

tomeasure

theprofi

ts(sales)for

respon

dentswho

wereno

tab

leto

recalltheprecisevalue.

Meanflat

isthemeanvalueof

thevariab

lein

thefla

ttreatm

entgrou

p.Rob

uststan

dard

errors

arerepo

rted

insqua

rebrackets.Allregression

scontrolforstratum

fixed

effects.*p<

0.1,

**p<

0.05,***p<

0.01.

198 CREDIT CONTRACT STRUCTURE AND FIRM GROWTHTable

5.3:Business

outcomes

IIand

householdeconom

icstatus

#paid

empl.

#unpaid

empl.

Business

Household

HH

income

High

HH

capitalasset

valuepast

yearincom

e=1

(1)(2)

(3)(4)

(5)(6)

Early

0.369 ∗∗∗-0.118 ∗

-1231.53011393.741

7045.756 ∗∗∗0.194 ∗∗∗

[0.136][0.071]

[2591.722][9547.490]

[2387.284][0.055]

Late0.339 ∗∗

-0.072-889.771

2313.1161015.147

-0.009[0.133]

[0.070][2572.283]

[9416.095][2354.726]

[0.054]Flexible

0.213-0.113

452.17410168.408

3840.4850.056

[0.133][0.069]

[2544.747][9424.883]

[2350.025][0.054]

Subsidy0.277 ∗∗

-0.08915.109

13063.5031617.632

0.059[0.130]

[0.068][2478.928]

[9144.405][2290.295]

[0.053]Control

0.0850.072

236.0694500.433

4718.861 ∗0.008

[0.138][0.072]

[2656.963][9722.069]

[2430.015][0.056]

pearly

=flex

0.2560.936

0.5180.899

0.1820.013

pearly

=late

0.8240.512

0.8960.346

0.0120.000

pearly

=subs.

0.4920.669

0.6250.859

0.0210.013

pearly

=control

0.0430.009

0.5880.487

0.3470.001

pflex

=late

0.3510.559

0.6030.408

0.2330.235

pflex

=subs.

0.6250.723

0.8610.753

0.3350.958

pflex

=control

0.3530.011

0.9360.564

0.7200.393

psubs.=

late0.644

0.8060.722

0.2480.796

0.207psubs.=

control0.157

0.0230.933

0.3730.197

0.357plate

=control

0.0690.048

0.6770.824

0.1320.768

Mean

flat1.007

0.32813357.69

19034.0712206.79

0.266Stratum

FE

Yes

Yes

Yes

Yes

Yes

Yes

Adjusted

R2

0.5910.154

0.2060.081

0.1840.177

Observations

731731

719731

728728

Note:#

paidem

pl.and

#paid

empl.

aretaken

fromthe

employee

roster.The

majority

ofunpaid

employees

areunpaid

family

workers

ofthe

owner.

Business

capital:Aggregation

ofthe

valueof

listedbusiness

assets(such

astools,m

achinesequipm

entand

furniture)and

thereported

totalvalueof

inventory.Excludes

thevalue

ofland

orbuildings.

Household

assetvalue:

The

valueof

alist

ofhousehold

assets.Excludes

thevalue

ofland

orbuildings.

Business

capital,Household

assetvalue

andHH

incomepast

yearare

expressedin

theirOctober

2014value,

asPPP-adjusted

US

dollars.High

HH

income=1:

Adum

my=

1ifthe

valueof

reportedhousehold

assetslie

abovethe

highestthreshold

ofintervals

thatweused

tomeasure

thehousehold

incomefor

respondentswho

were

notable

torecallthe

precisevalue.M

eanflat

isthe

mean

valueof

thevariable

inthe

flattreatm

entgroup.

Robust

standarderrors

arereported

insquare

brackets.Allregressions

controlfor

stratumfixed

effects.*p<0.1,

**p<0.05,

***p<0.01.

TABLES 199

Table 5.4: Employment details and other household businesses

Wage costs # Other # Newpast year HH businesses HH businesses

(1) (2) (3)

Early 169.697 0.147∗ 0.249∗∗∗[318.661] [0.083] [0.081]

Late -37.485 -0.014 0.131∗[313.995] [0.082] [0.079]

Flexible 270.486 -0.009 0.047[312.843] [0.081] [0.079]

Subsidy 781.552∗∗ 0.101 0.095[305.436] [0.079] [0.077]

Control 29.321 -0.059 0.002[324.692] [0.084] [0.082]

p early = flex 0.753 0.061 0.013p early =late 0.518 0.053 0.147p early =subs. 0.052 0.568 0.052p early =control 0.672 0.016 0.003p flex =late 0.329 0.949 0.291p flex =subs. 0.096 0.17 0.536p flex =control 0.459 0.55 0.584p subs. = late 0.008 0.156 0.642p subs. = control 0.019 0.055 0.251p late = control 0.839 0.595 0.119

Mean flat 818.23 0.422 0.195Stratum FE Yes Yes YesAdjusted R2 0.202 0.255 0.035Observations 729 731 731

Note: Wage costs past year : Wage expenditures in past 12 months in the borrowing firm.#Other HH businesses: Number of other businesses owned by owner’s/borrower’s HH.#NewHH businesses: Number of new businesses started by the household in the past 12 months,measured only at the endline.Mean flat is the mean value of the variable in the flat treatmentgroup. Robust standard errors are reported in square brackets. All regressions control forstratum fixed effects.* p<0.1, ** p<0.05, *** p<0.01.


Table 5.5: Voucher use timing among flexible treatment clients

Month in # Voucher 1 Fraction #Voucher 2 Fractionloan cycle used used used used

1 87 0.71 0 0.002 19 0.15 76 0.633 11 0.09 17 0.144 4 0.03 11 0.095 0 0.00 5 0.046 1 0.01 5 0.047 0 0.00 3 0.038 1 0.01 0 0.009 0 0.00 2 0.0210 0 0.00 0 0.0011 0 0.00 1 0.0112 0 0.00 0 0.00

Total 123 1.00 120 1.00Note: Column 1 shows the month in the loan cycle (1-12). Column 1 reports the number of "voucher1" used in that month, and colum 3 reports the share of "voucher 1" used in the respective month.Columns 4 and 5 does the corresponding thing for "voucher 2". Note that three out of the 123clients in our partial sample that were assigned to the flexible treatment only used one out of theirtwo vouchers.

APPENDIX 201

Appendix


Table A.1: Summary statistics/balance check, partial sample

Differences between flat and other treatmentsFlat Early Late Flexible Subs. Contr. All

Mean Coeff. Coeff. Coeff. Coeff. Coeff. N(s.d.) (s.d.) [st.err.] [st.err.] [st.err.] [st.err.]

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

Female owner 0.485 -0.066 -0.124∗∗ -0.077 -0.128∗∗ -0.092 754(0.502) [0.061] [0.060] [0.059] [0.058] [0.061]

Years educ. owner 10.504 -0.338 -0.761 -0.26 -0.517 -0.269 751(3.997) [0.508] [0.505] [0.496] [0.491] [0.514]

Profit past year 11475.637 -1049.791 -1414.305 -560.479 -354.681 -876.877 754(21155.554) [2023.083] [2010.830] [1974.013] [1944.972] [2047.205]

High profits =1 0.194 -0.001 -0.008 0.065 -0.004 -0.019 754(0.397) [0.050] [0.050] [0.049] [0.048] [0.051]

Sales past year 57920.443 -7157.674 -4882.702 750.596 -2633.959 -4349.727 754(110050.533) [11612.968] [11542.632] [11331.297] [11164.592] [11751.433]

High sales =1 0.142 0.034 0.013 0.045 0.029 0.025 754(0.35) [0.046] [0.046] [0.045] [0.044] [0.047]

Costs past year 36528.647 20010.830∗∗ 3856.71 5952.664 13062.724 7888.914 754(57361.967) [9980.630] [9920.180] [9738.551] [9595.279] [10099.632]

Wage costs year 1405.183 -570.182 228.016 585.596 -68.74 814.982 752(3555.742) [521.198] [518.277] [508.846] [502.404] [527.493]

Business capital 18960.35 -1394.209 -2448.075 -209.893 1685.439 567.108 749(46789.357) [4674.038] [4666.171] [4560.525] [4500.906] [4745.595]

Age of business (years) 8.463 -1.335 -0.388 -0.446 -0.029 -0.894 674(8.016) [0.932] [0.915] [0.901] [0.896] [0.930]

#Employees 1.455 -0.264 0.022 0.167 -0.216 -0.259 754(1.412) [0.192] [0.191] [0.187] [0.184] [0.194]

# paid empl. 1.269 -0.363∗ 0.196 0.375∗ -0.218 -0.112 754(1.537) [0.214] [0.212] [0.208] [0.205] [0.216]

# unpaid empl. 0.351 -0.028 -0.087 -0.151∗∗ -0.141∗ -0.13 754(0.816) [0.079] [0.078] [0.077] [0.076] [0.080]

HH asset value 12741.119 -1298.235 5619.316 21273.015∗∗∗ 8035.263 2270.85 754(29661.88) [7769.903] [7722.843] [7581.445] [7469.908] [7862.546]

HH inc. past year 13487.642 1855.154 1808.075 2048.307 1186.561 427.041 752(17026.426) [2517.816] [2503.706] [2458.145] [2427.026] [2548.224]

High HH income=1 0.263 0.019 0.007 0.072 0.02 -0.027 752(0.442) [0.054] [0.053] [0.052] [0.052] [0.054]

# HH businesses 0.567 0.069 0.131 0.055 -0.057 -0.01 754(0.74) [0.093] [0.093] [0.091] [0.090] [0.094]

F stat joint test 1.14 1.24 0.69 0.93 0.84 0.68P-value joint test 0.30 0.22 0.84 0.55 0.67 0.85

Note: Column 1 shows the descriptive statistics for the flat treatment group (the reference group). The remaining columns show the difference between eachadditional treatment group (or the control) and the flat treatment group. The last column shows the number of observations for all 6 treatments in the partialsample (the clients due until May 25, 2016). Coefficients and standard errors are from OLS regressions of each variable on the Early, Late, Flexible, Subsidy andControl groups, controlling for stratification. Last lines report the F-statistic and p-value from a joint test of the significance of the set of variables in explainingeach treatment dummy. Business capital Value of business assets, including inventory and excluding the value of land and buildings, in PPP adjusted USD.HH asset value Value of household assets, excluding land, in PPP adjusted USD. #HH businesses the number of other businesses owned by members of theborrower’s household. Robust standard errors are reported in square brackets. All regressions control for stratum fixed effects. * p<0.1, ** p<0.05, *** p<0.01.

APPENDIX 203

Tab

leA.2:K

eymon

etaryou

tcom

esrelative

tofla

ttreatm

ent,

trim

med

forou

tliers

Profit

year

Salesyear

Costs

year

Businesscapital

HH

assets

HH

incyear

(1)

(2)

(3)

(4)

(5)

(6)

Early

708.08

0-105

9.41

7-495

8.64

744

1.44

537

64.459

4538

.139∗∗∗

[109

9.52

2][647

5.37

2][682

9.86

2][159

6.765]

[501

9.06

5][160

3.25

9]La

te-909

.790

-587

6.18

612

57.607

249.72

037

22.667

1384

.472

[108

1.14

3][639

2.68

8][670

6.91

3][158

2.645]

[493

0.52

4][157

4.90

1]Flexible

68.935

-886

2.40

5-458

5.47

912

27.821

951.24

028

22.281∗

[107

1.97

0][634

1.04

4][667

9.69

5][156

6.973]

[497

0.73

4][158

1.56

9]Su

bsidy

754.91

0-392

5.17

5-345

5.56

1-376

.452

-499

1.28

310

91.525

[105

1.55

4][619

1.77

8][651

5.21

4][152

9.998]

[482

1.67

5][153

6.21

9]Con

trol

-595

.512

-613

4.68

2-9882.35

910

09.562

1117

.460

1084

.100

[111

5.16

5][657

7.75

2][691

8.65

9][163

9.685]

[509

7.04

4][164

1.42

5]

pearly=

flex

0.56

10.22

80.95

70.62

30.58

20.29

1pearly=

late

0.14

30.46

00.36

80.90

50.99

30.05

1pearly=

subs.

0.96

60.65

20.82

40.60

30.080

0.03

0pearly=

control

0.25

20.44

80.48

80.73

40.61

20.04

0pflex=

late

0.36

50.64

10.38

90.53

70.58

20.36

8pflex=

subs.

0.51

50.42

60.86

40.29

70.22

70.26

6pflex=

control

0.55

20.67

90.44

90.89

40.97

40.29

6psubs.=

late

0.12

00.75

70.47

90.68

90.07

70.85

1psubs.=

control

0.22

10.73

30.34

90.39

30.23

00.99

6plate=control

0.78

10.96

90.114

0.64

80.61

50.85

7

Stratum

FE

Yes

Yes

Yes

Yes

Yes

Yes

AdjustedR

20.24

90.37

00.20

20.373

0.06

80.21

3Observation

s72

171

872

1707

717

716

Note:Profit

year

:Profitsin

thepa

st12

mon

ths.Sa

lesyear

:Sales

inthepa

st12

mon

ths.Businesscapital:Aggregation

ofthevalueof

listed

business

assets

(suchas

tools,

machinesequipm

entan

dfurniture)

andtherepo

rted

totalvalueof

inventory.

Excludesthevalueof

land

orbu

ildings.Hou

seho

ldassets:The

valueof

alistof

householdassets.Excludesthevalueof

land

orbu

ildings.HH

incomeyear

:Hou

seho

ldincomein

thepa

st12

mon

ths.Allmon

etaryvariab

lesareexpressedin

theirOctob

er2014

value,as

PPP-adjustedUSdo

llars.R

obuststan

dard

errors

arerepo

rted

insqua


scontrolforstratum

fixed

effects.*p<

0.1,

**p<

0.05,***p<

0.01.


Table

A.3:B

usinessoutcom

esI,separate

fordiscounting

adjustedtreatm

entclients

Discounting

adjustedtreatm

entNon-D

iscountingadjusted

treatment

profitshigh

saleshigh

costsprofits

highsales

highcosts

yearprofit

yearsales

yearyear

profityear

salesyear

(1)(2)

(3)(4)

(5)(6)

(7)(8)

(9)(10)

Early

480.3000.025

-21792.922-0.046

-15498.2217772.794 ∗

0.00616143.931

0.05412521.394

[2338.820][0.067]

[13642.305][0.063]

[12594.974][4670.535]

[0.068][15590.470]

[0.061][12099.558]

Late-531.845

-0.001-8221.225

-0.02420930.552 ∗

424.675-0.064

7002.1620.016

4447.773[2325.747]

[0.067][13566.278]

[0.062][12450.319]

[4557.927][0.066]

[15162.812][0.059]

[11765.946]Flexible

-660.094-0.047

-20333.559-0.077

-11063.5621713.044

-0.003557.122

0.0235884.033

[2282.571][0.066]

[13315.467][0.061]

[12218.327][4613.906]

[0.067][15389.571]

[0.060][11968.877]

Subsidy1685.277

0.003-18298.955

-0.078-12551.719

1736.383-0.032

5839.0990.044

-1124.610[1954.688]

[0.056][11409.629]

[0.052][10471.840]

[3871.206][0.056]

[12925.125][0.050]

[10039.359]Control

-66.683-0.048

-21406.514 ∗-0.105 ∗

-17290.261-1657.889

-0.0731594.468

0.018-9856.310

[2047.169][0.059]

[11948.404][0.055]

[10963.491][4026.782]

[0.058][13445.306]

[0.052][10432.516]

pearly=

flex0.624

0.2840.914

0.6170.723

0.2040.897

0.3280.620

0.591pearly=

late0.669

0.7010.326

0.7240.004

0.1180.302

0.5600.530

0.507pearly=

sub0.549

0.6990.766

0.5520.785

0.1380.523

0.4480.846

0.196pearly=

control0.792

0.2230.975

0.2960.873

0.0250.196

0.2990.511

0.040pflex=

late0.956

0.4970.372

0.3900.010

0.7830.364

0.6790.900

0.905pflex=

sub0.229

0.3800.858

0.9870.887

0.9950.620

0.6920.693

0.498pflex=

control0.771

0.9850.928

0.6210.569

0.4150.245

0.9400.924

0.142psub=

late0.271

0.9470.391

0.3130.002

0.7400.569

0.9300.584

0.587psub=

control0.296

0.2950.751

0.5600.597

0.3060.393

0.7010.552

0.310plate=

control0.823

0.4360.279

0.1500.001

0.6130.887

0.6930.963

0.180

Mean

flat10196.09

0.23471352.48

0.23446053.19

10350.820.25

577060.125

31956.96Stratum

FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Adjusted

R2

0.2700.219

0.3680.242

0.1610.178

0.2110.357

0.2270.165

Observations

490490

490490

490482

482482

482482

Note:

In[T

1]-[T4]

arandom

50%of

clientsare

assignedto

thediscounting

adjustedtreatm

entversions,

while

for[T

5](the

subsidytreatm

ent)and

theControl

group,there

isno

discountingadjusted

group.Therefore,the

sample

incolum

ns1-5

includesthe

discountingadjusted

treatment-subgroups

oftreatm

entarm

s1-4,plus

allclientsin

T5and

Control.

The

sample

incolum

ns6-10

includesthe

non-discountingadjusted

treatment-subgroups

oftreatm

entarm

s1-4,

plus,again,

allclients

inT5and

Control.

Profits/Sales/C

ostspast

yearare

expressedin

theirOctober

2014value,

asPPP-adjusted

USdollars.

High

profits=1(H

ighsales

=1):

Dum

mies=

1ifthe

valueof

reportedprofits

(sales)lie

abovethe

highestthreshold

ofthe

intervalsthat

weused

tomeasure

theprofits

(sales)for

respondentswho

were

notable

torecall

theprecise

value.Robust

standarderrors

arereported

insquare

brackets.All

regressionscontrol

forstratum

fixedeffects.

*p<0.1,

**p<0.05,

***p<0.01.

APPENDIX 205

Tab

leA.4:B

usinessou

tcom

esII

andho

useholdecon

omic

status,s

eparatefordiscou

ntingad

justed

treatm

entclients

discou

ntingad

justed

treatm

ent

Non

-discoun

ting

adjusted

treatm

ent

#pa

id#

unpa

idBus.

HH

HH

inc

High

#pa

id#

unpa

idBus.

HH

HH

inc

High

empl

empl

capital

assets

year

HH

inc

empl

empl

capital

assets

year

HH

inc

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

Early

0.36

0*0.03

5-206

6.11

415

654.57

010

000.56

7∗∗∗

0.21

3∗∗∗

0.37

1∗-0.255∗∗

609.75

228

9.32

345

28.162

0.19

8∗∗

[0.193

][0.104

][337

4.68

1][140

01.452

][3757.65

0][0.079

][0.210

][0.109

][425

1.55

4][142

39.926

][319

2.32

0][0.078

]La

te0.35

7∗-0.059

1496

.081

6626

.925

-249

.053

-0.038

0.30

6-0.082

-274

5.97

0-436

9.64

630

10.850

0.01

8[0.192

][0.103

][335

5.76

1][139

49.803

][3742.24

1][0.078

][0.204

][0.106

][418

1.12

1][138

58.583

][310

5.90

7][0.075

]Flexible

0.27

80.00

910

71.844

3026.957

5169

.818

0.10

10.13

8-0.188∗

-33.52

014

129.84

624

34.023

0.01

9[0.188

][0.102

][330

7.04

8][137

38.897

][3670.75

7][0.077

][0.208

][0.108

][417

5.33

5][141

27.641

][317

1.43

3][0.077

]Su

bsidy

0.29

8∗0.00

587

2.00

510

492.05

910

02.138

0.04

90.29

3∗-0.162∗

-114

5.25

810

247.00

926

91.818

0.07

7[0.161

][0.087

][282

4.78

0][117

09.037

][3151.76

1][0.066

][0.174

][0.090

][350

9.49

2][118

18.217

][265

0.97

4][0.064

]Con

trol

0.09

20.16

2∗13

52.423

974.57

944

62.740

0.00

50.10

6-0.004

-815

.646

4906

.107

5482

.694∗∗

0.02

9[0.169

][0.091

][297

3.47

7][122

57.436

][3292.37

1][0.069

][0.181

][0.094

][367

0.28

0][122

75.097

][275

2.09

5][0.067

]

pearly=

flex

0.66

60.79

50.35

20.36

80.19

50.15

20.28

00.54

70.88

20.34

30.52

30.02

5pearly=

late

0.98

80.36

60.29

80.52

60.00

70.00

20.757

0.11

50.43

60.74

50.63

70.02

2pearly=

subs.

0.70

60.73

10.31

10.66

80.00

50.01

50.66

80.32

90.63

50.42

30.51

00.07

5pearly=

control

0.11

90.16

80.25

70.23

80.09

60.00

30.161

0.01

10.71

00.71

90.74

00.01

6pflex=

late

0.67

80.50

90.90

00.79

60.14

60.07

60.42

40.33

20.52

40.19

70.85

70.99

4pflex=

subs.

0.899

0.96

50.94

40.52

40.18

30.42

90.39

10.78

10.75

80.75

00.925

0.38

7pflex=

control

0.27

10.08

90.92

50.86

70.82

90.16

00.86

40.05

80.83

40.46

70.28

50.88

4psubs.=

late

0.72

30.47

00.830

0.74

80.69

80.19

90.94

20.38

60.66

10.22

70.90

60.37

4psubs.=

control

0.13

70.03

40.844

0.34

30.19

80.42

90.21

00.04

20.91

30.59

80.22

00.38

8plate=control

0.12

50.017

0.96

20.65

10.15

90.54

60.27

90.41

80.61

20.46

00.37

90.87

6

Meanfla

t1.00

0.30

1159

6.65

1717

9.17

1283

1.56

0.28

11.01

60.35

915

146.69

2088

8.97

1158

2.03

0.25

Stratum

FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

AdjustedR

20.55

70.14

50.29

80.07

50.20

10.17

60.55

30.11

70.19

70.11

70.16

10.17

2Observation

s49

049

048

449

048

848

848

248

247

248

248

048

0Note:

In[T

1]-[T4]

arand

om50%

ofclientsareassign

edto

thediscou

ntingad

justed

treatm

entversions,while

for[T

5](the

subsidytreatm

ent)

andtheCon

trol

grou

p,thereisno

discou

ntingad

justed

grou

p.Therefore,the

samplein

columns

1-6includ

esthediscou

ntingad

justed

treatm

ent-subg

roup

sof

treatm

entarms1-4,

plus

allclie

ntsin

T5an

dCon

trol.T

hesamplein

columns

7-12

includ

estheno

n-discou

ntingad

justed

treatm

ent-subg

roup

sof

treatm

entarms1-4,

plus,again,

allclientsin

T5an

dCon

trol.#paid

empl.an

d#paid

empl.aretakenfrom

theem

ployee

roster.The

majorityof

unpa

idem

ployeesareun

paid

family

workers

oftheow

ner.

Business

capital:Aggregation

ofthevalueof

listedbu

siness

assets

(suchas

tools,

machine

sequipm

entan

dfurniture)

andtherepo

rted

totalvalueof

inventory.

Excludesthevalueof

land

orbu

ildings.Hou

seho

ldassetvalue:

The

valueof

alistof

householdassets.Exclude

sthevalueof

land

orbu

ildings.Businesscapital,Hou

seho

ldassetvaluean

dHH

incomepast

year

areexpressedin

theirOctob

er2014

value,

asPPP-adjustedUSdo

llars.High

HH

income=1:

Adu

mmy=

1ifthevalueof

repo

rted

householdassets

lieab

ovethehigh

estthresholdof

intervalsthat

weused

tomeasure

theho

useholdincomeforrespon

dentswho

wereno

tab

leto

recalltheprecise

value.

Rob

uststan

dard

errors

arerepo

rted

insqua


scontrolforstratum

fixed

effects.*p<

0.1,

**p<

0.05,***p<

0.01.


Table A.5: Attrition rates by treatment

N N N Sharedue clients interviewed attrited attrited

Early 119 116 3 0.03Late 123 123 0 0.00Flexible 129 123 6 0.05Flat 134 128 6 0.04Subsidy 136 134 2 0.01Control 113 107 6 0.05

Total 754 731 23 0.03Note: Due client: A dummy variable =1 if a client’s due date for the endline interviewis before May 25, 2016. Attrited: a dummy=1 if the client’s is Due (before May 25) andshe has not been reached for the endline interview by June 25.

APPENDIX 207

Table A.6: Attrition by treatment

Early −0.016[0.022]

Late −0.048∗∗

[0.022]Flexible 0.000

[0.022]Subsidy −0.032

[0.021]Control 0.011

[0.022]

p early = flex 0.461p early = late 0.166p early = subs. 0.479p early = control 0.233p flex = late 0.031p flex = subs. 0.136p flex = control 0.625p subs. = late 0.473p subs. = control 0.055p late = control 0.011

Stratum FE YesAdjusted R2 -0.005Observations 754

Note: The table shows results from a regression of a dummy for attrition on alltreatments relative to the flat treatment. Robust standard errors are reportedin brackets. Controls for stratum included in regression. * p<0.1, ** p<0.05,*** p<0.01.

Sammanfattning

Denna avhandling består av fyra fristående kapitel som alla kretsar kring teman som

utvecklingspolitik och politisk-ekonomiska aspekter av genomförandet av utvecklings-

program. Kapitel 2, 4 och 5 analyserar utvecklingsinitiativ som är kopplade till eko-

nomisk integration av fattiga medan kapitel 3 behandlar ett utvecklingsprogram med

politiska undertoner.

I utvecklingsländer står icke-statliga organisationer och andra externa aktörer ofta

för en stor del av tillhandahållandet både av offentliga tjänster som ej tillhandahålls av

staten, och av finansiella tjänster som endast i otillräcklig utsträckning erbjuds av ban-

ker och andra formella institutioner (Baland et al., 2011;. Casey et al., 2012; Grossman,

2014). Under de senaste decennierna har ledande icke-statliga organisationer priorite-

rat utvecklingsprojekt som, trots att de införts av externa initiativtagare, involverar

lokalsamhället i beslutsfattandet kring programmen (Mansuri och Rao, 2012). Detta

anses öka projektens legitimitet och långsiktiga hållbarhet. Tidigare utvärderingar av

sådana projekt framhåller flera fördelar med direkt lokalt deltagande i beslut jämfört

med mer centralt beslutsfattande. Än så länge vet vi dock mycket lite om den relativa

effektiviteten hos olika typer av direkt deltagande. Med undantag av Grossman (2014)

har dessa studier inte heller behandlat lokalt ledarskap, trots att en viktig faktor för

effektiviteten i dessa projekt är hur deras lokala ledning är organiserad.

I Kapitel 2, Electoral Rules and Leader Selection: Experimental Evidence

from Ugandan Community Groups (Valregler och val av ledare: experimentell

evidens från ugandiska sparandegrupper), studeras hur utformningen av valregler av-

gör vilka typer av ledare som väljs i sparande-grupper för unga kvinnor i en fattig

209

210 SAMMANFATTNING

del av Uganda. Trots en omfattande teoretisk litteratur om hur valsystem påverkar

policy, vet vi mindre om deras inverkan på ledare. Dessutom är det mycket svårt att

finna sammanhang där exogen variation i valregler förekommer. Vi instruerade ugan-

diska sparande-grupper att använda en av följande två metoder när de valde ledare

för första gången: val genom sluten omröstning eller val i en öppen diskussion med

beslut fattat genom konsensus. Instruktionerna randomiserades mellan grupper, vilket

gör det möjligt för oss att estimera den kausala effekten av valreglerna både på ledares

egenskaper och på mätbara grupputfall till följd av ledarnas genomförda politik. Mer

specifikt undersöker vi här andelen medlemmar som blir kvar i grupperna över tid

samt nivåer och allokering av sparande och lån inom gruppen. Vi finner att grupper

som valt ledare genom sluten omröstning i högre utsträckning väljer ledare som liknar

den genomsnittliga gruppmedlemmen, medan grupper som valt sina ledare i en öppen

diskussion väljer rikare ledare med högre utbildning än den genomsnittliga medlem-

men. Vidare finner vi att avhopp från gruppen är betydligt vanligare i grupper som

använt öppen diskussion, i synnerhet bland de medlemmar som i början var fattigast i

sin sparande-grupp. Efter 3,5 år är grupper som valt ledare genom sluten omröstning

större och deras medlemmar sparar mindre belopp och beviljas mindre lån än medlem-

mar som valt ledare genom öppen diskussion. Våra resultat tyder på att inflytelserika

medlemmar, med mer makt i gruppen initialt, i högre utsträckning påverkar utfallet

i diskussionsförfarandet vilket leder till utfall som är mindre representativa för den

genomsnittliga medlemmens preferenser än vid sluten omröstning. Detta ligger i linje

med resultat från tidigare studier av införandet av slutna val (Baland och Robinson

(2008) och Hinnerich och Pettersson-Lidbom (2014)). Vi finner alltså att valmetoden

som används påverkar både ledartyper och grupputfall, där ett system med sluten

omröstning skapar mer inkluderande grupper medan ett system med öppen diskussion

leder till sämre ekonomisk integration av samhällets fattigaste medlemmar. Med tan-

ke på den avgörande roll som denna typ av grupper spelar för tillhandahållandet av

många offentliga och finansiella tjänster i låginkomstländer har vår studie konsekven-

ser för offentlig service i dessa länder.

211

En utbredd uppfattning är att stormöten samt idéella organisationer och klubbar

främjar socialt kapital genom att utgöra arenor där människor kan mötas, utbyta idéer,

lösa snålskjuts-problem och skapa kollektiva nyttigheter (Grootaert och van Bastelaer,

2002; Guiso, Sapienza och Zingalez, 2008; Knack och Keefer, 1997; Putnam, 2000).

Denna syn förklarar delvis det ökade fokus hos centrala utvecklingsorganisationer på

projekt där byrådsmöten och gräsrotsdeltagande spelar en central roll, som diskuteras

ovan (se Mansuri och Rao (2012) för en översikt). Nyare forskning har dock visat att

föreningsliv och olika sociala forum även kan ha negativa effekter på socialt kapital:

istället för att överbrygga samhälleliga, sociala och etniska klyftor, kan denna typ av

forum istället förstärka dem (Satyanath et al. (2015).

Kapitel 3, Preparing for Genocide: Community Meetings in Rwanda (För-

beredelser för folkmord: bymöten i Rwanda), anknyter till denna litteratur genom att

studera en helt annan typ av utvecklingsprogram, som på grund av sin politiska ka-

raktär fick förödande konsekvenser. Olika former av obligatorisk samhällstjänst på

lokal nivå har använts i Rwanda sedan innan kolonialtiden och liknande institutioner

förekom under den tidiga postkoloniala perioden även i andra Öst- och Centralafri-

kanska länder (Guichaoua, 1991). Under perioden 1973-1994, blev den obligatoriska

samhällstjänsten statlig politik i Rwanda. Varje lördag möttes Rwandiska bybor för

att arbeta med gemensamma projekt som vägar och annan infrastruktur, en praxis

som kallas Umuganda. Detta fenomen motiverades genom utvecklingsargument, men

var också mycket politiserat och enligt kvalitativ forskning av bl.a. Straus (2006) och

Verwimp (2013), användes mötena i anslutning till dessa arbetsdagar regelbundet av

den lokala politiska eliten för att sprida propaganda under åren före folkmordet. Detta

kapitel presenterar de första kvantitativa bevisen för detta (miss)bruk av Umuganda.

Att identifiera den kausala effekten av mötena på senare deltagande i folkmordet är

svårt, av två skäl. För det första har vi inte data över antalet personer som deltog

i Umuganda under perioden innan 1994, eller över antalet möten som ägde rum på

212 SAMMANFATTNING

en viss plats. Även om sådan data fanns tillgänglig skulle våra estimat dessutom ka-

raktäriseras av s.k. omitted variable bias. För att fastställa ett orsakssamband mellan

mötenas intensitet och deltagande i våld under folkmordet utnyttjar vi därför exo-

gena väderförhållanden. Antagandet vi gör är att när det regnar kraftigt blir mötet

antingen inställt eller mindre intensivt. Vi använder daglig regnfallsdata från perioden

1984-1998 samt data från lokala domstolar, särskilt inrättade efter folkmordet, över

mål mot civila för deltagande i folkmordet. Vi finner att en ytterligare regnig lördag i

en viss by under åren före folkmordet resulterade i fem procent lägre civilt deltagande

i folkmordsvåld i samma by. Vi finner inga resultat av regn under andra veckodagar på

deltagande i folkmordet. Dessutom drivs vårt resultat helt av de orter som innan 1994

styrdes av hutu-ledda partier (som stödde folkmordet). På de få platser som styrdes

av pro-tutsiminoriteten är effekterna omvända. Trots vårt specifika geografiska fokus

i denna studie, menar vi att det är av mer allmänt intresse att undersöka eventuellt

negativa effekten av olika typer av stormöten. Medan dessa möten allmänt ses som

något positivt, framhåller vi stöd för att det finns en mörkare sida av dessa möten där

det sociala kapitalet som skapas inte överbryggar samhälleliga, etniska klyftor, utan

istället förstärker dem. Att förstå denna process är ännu viktigare eftersom Umugan-

da, trots sin tidigare användning, formellt återinfördes i Rwanda år 2008, och liknande

institutioner har införts i Burundi och nyligen varit uppe på förslag i Kenya.1

En av de former för utvecklingsstöd som uppmärksammats mest under de senaste

årtiondena är mikrofinansiering. Mikrokrediter och det vidare begreppet mikrofinan-

siering blev känt för allmänheten då Grameen bank och dess grundare Mohammad

Yunus tilldelades Nobels fredspris 2006. Tanken bakom mikrofinansiering är att små

lån kan hjälpa fattiga människor att förbättra sin försörjning genom småskalig kom-

mersiell verksamhet. Som Amendariz de Aghion och Morduch (2005) skriver i sin bok

om mikrofinansiering: Mikrofinans presenterar sig som en ny marknadsbaserad stra-

tegi för fattigdomsbekämpning, fri från de tunga subventioner som fällt stora statliga

1För mer information om den kenyanska fallet, se Daily Nation, (2016).

213

banker. I en värld på jakt efter enkla svar har denna win-win kombination blivit en

riktig vinnare.2 Trots stor entusiasm kring mikrofinansiering som ett verktyg för att

minska fattigdomen har de större utvärderingar av mikrofinansiering som genomförts

de senaste åren funnit att de långsiktiga effekterna på tillväxt och låntagares välfärd är

begränsade (Banerjee et al., 2015). Kapitel 4 och 5 i avhandlingen undersöker möjliga

sätt på vilka mikrofinansiering, genom ändringar i lånekontraktets utformning, bättre

skulle kunna uppfylla löftena om tillväxt. Båda kapitel fokuserar på mikroföretag samt

små och medelstora företag och gäller individuella lån.

Kapitel 4, Selection into Borrowing: Survey Evidence from Uganda (Se-

lektion på kreditmarknaden: evidens från Uganda), rapporterar resultaten från en

undersökning som estimerar efterfrågan på lån hos ett representativt urval av småfö-

retagare i urbana Uganda. Forskning inom kontraktsteori visar att en höjning av priset

på krediter (räntan) kan leda till antingen fördelaktiga (Stiglitz och Weiss, 1981) eller

negativa selektionseffekter (De Meza och Webb, 1987), med avseende på sannolikhe-

ten för projektets framgång, samt att ökningar av storleken på den säkerhet3 som

långivaren kräver leder till fördelaktiga selektionseffekter (Stiglitz och Weiss, 1981;

Wette, 1983). Att förstå selektionseffekter inom mikrofinansiering är av särskilt in-

tresse eftersom denna marknad kännetecknas av kreditransonering, delvis på grund

av asymmetrisk information. Befintliga studier av mikrofinansiering fokuserar på in-

divider eller företag som redan tar lån, och kan därför endast ge begränsad insikt i

hur kontrakts-förändringar skulle påverka kreditefterfrågan och investeringsbeteende

genom förändringar i sammansättningen av låntagare. Jag studerar låneattityder i ett

representativt urval av småföretagare, de flesta utan erfarenhet av att ta lån, som är

aktiva i centrala branscher inom både detaljhandel och tillverkning. Hypotetiska frågor

om efterfrågan på olika typer av lån används för att testa om småföretagare reagerar

2Det ursprungliga citatet lyder: Microfinance presents itself as a new market-based strategy forpoverty reduction, free of the heavy subsidies that brought down large statebanks. In a world in searchof easy answers, this win-win combination has been a true winner itself.

3Med säkerhet menas här en pant som långivaren kräver av låntagaren, vars värde långivaren kangöra anspråk på om återbetalning av lånet uteblir.

214 SAMMANFATTNING

på förändringar i lånevillkoren och huruvida efterfrågan varierar beroende på företa-

gens och entreprenörernas riskattityder och egenskaper. Resultaten visar att kontrakt

med lägre räntor eller lä gre säkerhets-belopp än standard-kontraktet leder till hög-

re efterfrågan blan alla typer av (potentiella) låntagare men effekten är särskilt stor

bland mindre risktagandeföretagare i jämförelse med riskbenägna företagare. Detta

gäller oavsett om risktagandedefinieras utifrån företagets uttalade riskbeteende eller

utifrån riskfaktorer i deras företagsklimat. Resultaten kvarstår även då vi kontrollerar

för förmögenhet. Dessa resultat är starkare bland tillverkningsföretag än bland detalj-

handlare, vilket kan förklaras av skillnader i tillgängliga investeringsalternativ mellan

sektorer. Mindre förmögna företagare blir mer benägna att låna då nivån på säkerheten

sänks. Resultaten tyder på att det finns utrymme för förbättringar av standardiserade

lånevillkor.

Kapitel 5, Credit Contract Structure and Firm Growth: Evidence from a

Randomized Control Trial (Lånekontraktets utformning och företagstillväxt: evi-

dens från en randomiserad kontrollstudie), presenterar de första resultaten från en

pågående randomiserad kontrollstudie bland faktiska låntagare i Uganda. Vi bygger

vidare på tidigare studier som funnit att förändringar av avtalsvillkoren kan förbättra

effektiviteten hos mikrolån (Field et al., 2013;. Karlan och Zinman, 2008), genom ett

experiment utformat att skilja mellan några av de viktigaste hinder för framgångsrikt

investerande som små företag ställs r inför. Vi varierar med vilken frekvens samt i

vilken del (när) i lånecykeln som återbetalning av lånet utkrävs för att kunna sär-

skilja mellan utmaningar såsom (i) fördröjd avkastning på investeringar, (ii) osäker

avkastning på investeringar och (iii) förekomsten av höga fasta kostnader. Företagen

i experimentet fick rabatter, som kunde användas för att täcka 2 av 12 månatliga

avbetalningar i den ettåriga lånecykeln. Resultaten som presenteras i kapitel 5 är pre-

liminära resultat från de första 754 av 2340 företag som avslutat sin lånecykel. Vi finner

att företag/företagare som fick en 2-månaders frist i början av lånecykeln ökade sina

vinster och hushållsinkomster i förhållande till företag/företagare som fick en rabatt

215

senare i lånecykeln, samt till kontrollgruppen. Antalet betalda anställda ökade även i

dessa företag, samtidigt som antalet obetalda sådana inskade, men löneutgifterna öka-

de inte i enlighet med detta. Vidare finner vi att företagsägares hushåll i gruppen som

beviljades en 2-månaders frist i början av lånecykeln startade fler nya verksamheter i

jämförelse med hushållen i gruppen som beviljades en rabatt senare i lånecykeln, samt

kontrollgruppen.

Bland företag som erbjöds en mer flexibel frist, där de kunde hoppa över återbetal-

ningen i två valfria månader, valde de flesta att använda sina rabatter under de första

månaderna av lånecykeln. Dessa resultat ger ett visst stöd för att fördröjd avkastning

utgör ett viktigare hinder än osäker avkastning för företagen i vår studie. Företag som

mottog ett kontant bidrag i början av lånecykeln ökade antalet anställda i förhållande

till kontrollgruppen, och de ökade också sina lönekostnader. Om detta innebär att de i

högre utsträckning anlitade mer välkvalificerad arbetskraft, vilket kan ses som en odel-

bar investering, ger detta resultat stöd för att även odelbara kostnader utgör ett hinder

för framgångsrika investeringar inom ramarna för det standardmässiga lånekontraktet.

216 SAMMANFATTNING

Referenser



Baland, J. M. and J. A. Robinson. 2008 Land and Power: Theory and Evidence


Baland, J.M., R. Somanathan and L. Vandewalle. 2011 Socially disadvantaged



Banerjee, A., Karlan, D. and Zinman, J. 2015. Six randomized evaluations of

microcredit: introduction and further steps. American Economic Journal: Applied Eco-

nomics, 7(1), pp.1-21.



1812.

Daily Nation (hämtad den 5 Mars, 2016). Emuhaya MP Seeks to Introduce Manda-


DeMeza, D. andWebb, D.C. 1987. Too much investment: a problem of asymmetric

information. The quarterly journal of economics, pp.281-292.

Field, E., Pande, R., Papp, J. and Rigol, N. (2013. Does the classic microfinan-

ce model discourage entrepreneurship among the poor? Experimental evidence from

India. The American Economic Review, 103(6), pp.2196-2226.







REFERENSER 217



295-320.



961-993.





Mansuri, G. and V. Rao, 2012. Localizing development: does participation work?.







Stiglitz, J.E. and Weiss, A. 1981. Credit rationing in markets with imperfect in-

formation. The American economic review, 71(3), pp.393-410.





218 SAMMANFATTNING

MONOGRAPH SERIES

1. Michaely, Michael The Theory of Commercial Policy: Trade and Protection,1973

2. Söderström, Hans Tson Studies in the Microdynamics of Production and Pro-ductivity Change, 1974

3. Hamilton, Carl B. Project Analysis in the Rural Sector with Special Reference tothe Evaluation of Labour Cost, 1974

4. Nyberg, Lars and Staffan Viotti A Control Systems Approach to MacroeconomicTheory and Policy in the Open Economy, 1975

5. Myhrman, Johan Monetary Policy in Open Economies, 1975

6. Krauss, Melvyn International Trade and Economic Welfare, 1975

7. Wihlborg, Clas Capital Market Integration and Monetary Policy under DifferentExchange Rate Regimes, 1976

8. Svensson, Lars E.O. On Competitive Markets and Intertemporal Resources Allo-cation, 1976

9. Yeats, Alexander J. Trade Barriers Facing Developing Countries, 1978

10. Calmfors, Lars Prices, Wages and Employment in the Open Economy, 1978

11. Kornai, Janos Economics of Shortage, Vols I and II, 1979

12. Flam, Harry Growth, Allocation and Trade in Sweden. An Empirical Applicationof the Heckscher-Ohlin Theory, 1981

13. Persson, Torsten Studies of Alternative Exchange Rate Systems. An Intertem-poral General Equilibrium Approach, 1982

14. Erzan, Refik Turkey’s Comparative Advantage, Production and Trade Patterns inManufactures. An Application of the Factor Proportions Hypothesis with SomeQualifications, 1983

15. Horn af Rantzien, Henrik Imperfect Competition in Models of Wage Formationand International Trade, 1983

16. Nandakumar, Parameswar Macroeconomic Effects of Supply Side Policies andDisturbances in Open Economies, 1985

17. Sellin, Peter Asset Pricing and Portfolio Choice with International InvestmentBarriers, 1990

219

220 MONOGRAPH SERIES

18. Werner, Ingrid International Capital Markets: Controls, Taxes and ResourcesAllocation, 1990

19. Svedberg, Peter Poverty and Undernutrition in Sub-Saharan Africa: Theory,Evidence, Policy, 1991

20. Nordström, Håkan Studies in Trade Policy and Economic Growth, 1992

21. Hassler, John, Lundvik, Petter, Persson, Torsten and Söderlind, Paul The SwedishBusiness Cycle: Stylized facts over 130 years, 1992

22. Lundvik, Petter Business Cycles and Growth, 1992

23. Söderlind, Paul Essays in Exchange Rates, Business Cycles and Growth, 1993

24. Hassler, John A.A. Effects of Variations in Risk on Demand and Measures ofBusiness Cycle Comovements, 1994

25. Daltung, Sonja Risk, Efficiency, and Regulation of Banks, 1994

26. Lindberg, Hans Exchange Rates: Target Zones, Interventions and Regime Col-lapses, 1994

27. Stennek, Johan Essays on Information-Processing and Competition, 1994

28. Jonsson, Gunnar Institutions and Incentives in Monetary and Fiscal Policy, 1995

29. Dahlquist, Magnus Essays on the Term Structure of Interest Rates and MonetaryPolicy, 1995

30. Svensson, Jakob Political Economy and Macroeconomics: On Foreign Aid andDevelopment, 1996

31. Blix, Mårten Rational Expectations and Regime Shifts in Macroeconometrics,1997

32. Lagerlöf, Nils-Petter Intergenerational Transfers and Altruism, 1997

33. Klein, Paul Papers on the Macroeconomics of Fiscal Policy, 1997

34. Jonsson, Magnus Studies in Business Cycles, 1997

35. Persson, Lars Asset Ownership in Imperfectly Competitive Markets, 1998

36. Persson, Joakim Essays on Economic Growth, 1998

37. Domeij, David Essays on Optimal Taxation and Indeterminacy, 1998

38. Floden, Martin Essays on Dynamic Macroeconomics, 1999

39. Tangerås, Thomas Essays in Economics and Politics: Regulation, Elections andInternational Conflict, 2000

40. Lidbom, Per Pettersson Elections, Party Politics and Economic Policy, 2000

41. Vestin, David Essays on Monetary Policy, 2001

42. Olofsgård, Anders Essays on Interregional and International Political Economics,2001

221

43. Johansson, åsa Essays on Macroeconomic Fluctuations and Nominal Wage Rigid-ity, 2002

44. Groth, Charlotta Topics on Monetary Policy, 2002

45. Gancia, Gino A. Essays on Growth, Trade and Inequality, 2003

46. Harstad, Bård Organizing Cooperation: Bargaining, Voting and Control, 2003

47. Kohlscheen, Emanuel Essays on Debts and Constitutions, 2004

48. Olovsson, Conny Essays on Dynamic Macroeconomics, 2004

49. Stavlöt, Ulrika Essays on Culture and Trade, 2005

50. Herzing, Mathias Essays on Uncertainty and Escape in Trade Agreements, 2005

51. Bonfiglioli, Alessandra Essays on Financial Markets and Macroeconomics, 2005

52. Pienaar, Natalie Economic Applications of Product Quality Regulation in WTOTrade Agreements, 2005

53. Song, Zheng Essays on Dynamic Political Economy, 2005

54. Eisensee, Thomas Essays on Public Finance: Retirement Behavior and DisasterRelief, 2005

55. Favara, Giovanni Credit and Finance in the Macroeconomy, 2006

56. Björkman, Martina Essays on Empirical Development Economics: Education,Health and Gender, 2006

57. Larsson, Anna Real Effects of Monetary Regimes, 2007

58. Prado, Jr., Jose Mauricio Essays on Public Macroeconomic Policy, 2007

59. Tonin, Mirco Essays on Labor Market Structure and Policies, 2007

60. Queijo von Heideken, Virginia Essays on Monetary Policy and Asset Markets,2007

61. Finocchiaro, Daria Essays on Macroeconomics, 2007

62. Waisman, Gisela Essays on Discrimination and Corruption, 2008

63. Holte, Martin Bech Essays on Incentives and Leadership, 2008

64. Damsgaard, Erika Färnstrand Essays on Technology Choice and Spillovers, 2008

65. Fredriksson, Anders Bureaucracy, Informality and Taxation: Essays in Develop-ment Economics and Public Finance, 2009

66. Folke, Olle Parties, Power and Patronage: Papers in Political Economics, 2010

67. Drott, David Yanagizawa Information, Markets and Conflict: Essays on Devel-opment and Political Economics, 2010

68. Meyersson, Erik Religion, Politics and Development: Essays in Development andPolitical Economics, 2010

222 MONOGRAPH SERIES

69. Klingelhöfer, Jan Models of Electoral Competition: Three Essays in PoliticalEconomics, 2010

70. Perrotta, Maria Carmela Aid, Education and Development, 2010

71. Caldara, Dario Essays on Empirical Macroeconomics, 2011

72. Mueller, Andreas Business Cycles, Unemployment and Job Search: Essays inMacroeconomics and Labor Economics, 2011

73. Von Below, David Essays in Climate and Labor Economics,2011

74. Gars, Johan Essays on the Macroeconomics of Climate Change, 2012

75. Spiro, Daniel Some Aspects on Resource and Behavioral Economics, 2012

76. Ge, Jinfeng Essays on Macroeconomics and Political Economy, 2012

77. Li, Yinan Institutions, Political Cycles and Corruption: Essays on DynamicPolitical Economy of Government, 2013

78. Håkanson, Christina Changes in Workplaces and Careers, 2013

79. Qin, Bei Essays on Empirical Development and Political Economics, 2013

80. Jia, Ruixue Essays on the Political Economy of China’s Development, 2013

81. Campa, Pamela Media Influence on Pollution, and Gender Equality, 2013

82. Seim, David Essays on Public, Political and Labor Economics, 2013

83. Shifa, Abdulaziz Essays on Growth, Political Economy and Development, 2013

84. Panetti, Ettore Essays on the Economics of Banks and Markets, 2013

85. Schmitt, Alexander Beyond Pigou: Climate Change Mitigation, Policy Makingand Distortions, 2014

86. Rogall, Thorsten The Economics of Genocide and War, 2015

87. Baltrunaite, Audinga Political Economics of Special Interests and Gender, 2016

88. Harbo Hansen, Niels-Jakob Jobs, Unemployment and Macroeconomic Transmis-sion, 2016

89. Stryjan, Miri Essays on Development Policy and the Political Economy of Con-flict, 2016

essaysondevelopmentpolicyandthe …936631/...the results indicate that contracts with lower interest...

Documents