Essays on Markets and Institutions in Emerging Economies · 2018-10-10 · Under the monopoly manufacturer, late deliveries are common due to outside de- ... violence retain more

Essays on Markets and Institutions in Emerging Economies

by

Tarek Fouad Ghani

A dissertation submitted in partial satisfaction of the

requirements for the degree of

Doctor of Philosophy

in

Business Administration

in the

Graduate Division

of the

University of California, Berkeley

Committee in charge:

Professor Steven Tadelis, Co-chairProfessor Ernesto Dal Bo, Co-chair

Professor Edward Miguel

Spring 2015


Copyright 2015by

Tarek Fouad Ghani

1

Abstract


by

Tarek Fouad Ghani

Doctor of Philosophy in Business Administration

University of California, Berkeley

Professor Steven Tadelis, Co-chair

Professor Ernesto Dal Bo, Co-chair

Market frictions pervade emerging economies and constrain private sector development.In such settings, formal institutions to help address contract enforcement, property rightsand information asymmetries are typically weak or absent. Instead, market participantsmust rely on informal practices and institutions to mitigate uncertainty, instability andopportunism. For example, personalized exchange relationships are useful when contractenforcement is weak, and cash holdings can be attractive when financial institutions areunreliable. In three specific emerging economy settings, I explore how informal practicesand institutions interact with formal market development, and in particular the role thatmarket frictions play in determining outcomes for firms, technologies and employees.

The first chapter of this dissertation explores how changes in formal upstream marketstructure affect the economics of downstream relationships using original data from the iceindustry in Sierra Leone. In this setting, a monopoly ice manufacturer sells through inde-pendent retailers to fishermen buyers. I demonstrate that a shock that increases upstreamcompetition among manufacturers improves the contractual terms offered by retailers tobuyers. Under the monopoly manufacturer, late deliveries are common due to outside de-mand shocks. To help mitigate this uncertainty, retailers prioritize loyal customers whenfaced with shortages, and buyers respond by rarely switching retailers. When manufactur-ers compete, prices fall, quantities increase and services improve with fewer late deliveries.Entry upstream also disrupts collusion among retailers by increasing the value of competingfor buyer relationships. Competing retailers expand trade credit provision as a new basisfor loyalty, and stable buyer relationships reemerge after a period of intense switching. Thefindings suggest that market structure shapes informal contractual institutions, and thatcompetition can reconstitute the nature of relationships.

The second chapter addresses the relationship between violence and financial decisions inAfghanistan. In particular, I investigate how violence affects the tradeoff between informalcash holdings - which are liquid but insecure - and usage of a more secure but less liquid formalfinancial account. Using three separate data sources, I find that individuals experiencing

2

violence retain more cash and are less likely to adopt and use mobile money, a new financialtechnology. First, combining detailed information on the entire universe of mobile moneytransactions in Afghanistan with administrative records for all violent incidents recorded byinternational forces, I find a negative relationship between violence and mobile money use.Second, in the context of a randomized control trial, violence is associated with decreasedmobile money use and greater cash balances. Third, in financial survey data from nineteenof Afghanistan’s 34 provinces, I find that individuals experiencing violence hold more cash.Collectively, the evidence indicates that individuals experiencing violence prefer cash tomobile money. More speculatively, it appears that this is principally because of concernsabout future violence. These results emphasize the difficulty of creating robust financialnetworks in conflict settings.

Finally, in the third chapter, I study how informal behaviors interact with incentives toaffect employees’ decisions to formally save in the context of a large firm in Afghanistan. Ianalyze a mobile phone-based account that allows savings to be automatically deducted fromsalaries. Employees who are automatically enrolled in this defined-contribution account are40 percentage points more likely to contribute than individuals with a default contributionof zero. Analyzing randomly assigned employer matching contributions, I find that the effectof automatic enrollment on participation is approximately equivalent to providing financialincentives equal to a 50 percent match. To understand why default enrollment increasesparticipation, some employees are randomly offered an immediate financial consultation, andothers a financial consultation in one week. Employees are more likely to discuss changingtheir savings contributions in one week, suggesting that defaults raise contributions becauseof the perceived complexity of financial decisions, and because employees procrastinate indeveloping a financial plan for the future.

i

This dissertation is dedicated to my teachers.

ii

Contents

1 Competing for Relationships in Sierra Leone 11.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.4 Theoretical framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.5 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 171.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261.8 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271.9 Chapter Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2 Violence and Financial Decisions in Afghanistan 452.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482.4 Administrative Data Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 512.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522.6 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592.8 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612.9 Chapter Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3 Automatic Payroll Deductions in Afghanistan 843.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853.3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 873.4 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 923.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 943.7 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 953.8 Chapter Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

iii

Bibliography 111

iv

Acknowledgments

I am very grateful for invaluable guidance and support from Steven Tadelis, Ernesto Dal Boand Edward Miguel. I also thank Eli Berman, Joshua Blumenstock, Michael Callen, Fred-erico Finan, Tristan Reed, Jacob Shapiro, Oliver Williamson, Christopher Woodruff, NoamYuchtman and numerous other friends, colleagues and seminar participants for commentsand suggestions. This dissertation was supported by the Paul and Daisy Soros Fellowshipfor New Americans, and fellowships from the UC Berkeley Haas School of Business and theUC Institute on Global Conflict and Cooperation.

Chapter 1 was made possible through the cooperation and support of Niall O’Cathasaigh,Tom Cairnes and the staff at Ice Ice Baby, as well as the excellent research assistance ofAbhay Aneja, Anthony Mansaray, Osman Nabay, James Polit and the Sierra Leone countryteam at Innovations for Poverty Action. Support was provided by the Clausen Center forInternational Business and Policy, the International Growth Center, the Private EnterpriseDevelopment in Low-Income Countries Initiative, and the Weiss Family Fund for Researchin Development Economics.

Chapter 2 was made possible through the cooperation and support of Karim Khoja,Raju Shaulis and staff at Roshan and the Central Asia Development Group, as well as theexcellent research assistance of Gregory Adams, Elizabeth Hastings, Shahim Kabuli, IanKelley and Lucas Koepke. Support was provided by the Center for Effective Global Action,the Institute for Money, Technology, and Financial Inclusion, the Consortium on FinancialSystems and Poverty, the Empirical Studies of Conflict Project, and the Private EnterpriseDevelopment in Low-Income Countries Initiative.

Chapter 3 was made possible through the cooperation and support of Karim Khoja andthe staff at Roshan, as well as the excellent research assistance of Katy Doyle, MohammadIsaqzadeh, Shahim Kabuli, Nasir Mahmoodi, Galen Murray, Maria Qazi and Hugo Gerard.Supportt was provided by the Citi/IPA Financial Capability Research Fund, the Consortiumon Financial Systems and Poverty, the Empirical Studies of Conflict Project, the Institutefor Money, Technology, and Financial Inclusion, and the UC San Diego Policy Design andEvaluation Lab.

1

Chapter 1

Competing for Relationships in SierraLeone

1.1 Abstract

A body of literature suggests that relationships affect contractual and market outcomes,but how does market structure affect the economics of relationships? This paper providesmicroeconometric evidence that upstream market structure affects the value of downstreamrelationships between retailers and buyers. In our setting, a monopoly ice manufacturer sellsthrough independent retailers to fishermen buyers in Sierra Leone. We demonstrate that ashock that increases upstream competition among manufacturers improves the contractualterms offered by retailers to buyers. Under the monopolistic manufacturer, we document thatlate deliveries are common due to outside demand shocks. To help mitigate this uncertainty,retailers prioritize loyal customers when faced with shortages, and buyers respond by rarelyswitching retailers. When manufacturers compete, prices fall, quantities increase and servicesimprove with fewer late deliveries. Entry upstream also disrupts collusion among retailersby increasing the value of competing for buyer relationships. Competing retailers expandtrade credit provision as a new basis for loyalty, and stable buyer relationships reemerge aftera period of intense switching. The findings suggest that market structure shapes informalcontractual institutions, and that competition can reconstitute the nature of relationships.1

1The material in this chapter is based on joint work with Tristan Reed. See Ghani and Reed (2015).

CHAPTER 1. COMPETING FOR RELATIONSHIPS IN SIERRA LEONE 2

1.2 Introduction

Relationships help mitigate transactional hazards, particularly when contracts are incom-plete. Repeated interactions form the basis for informal contractual institutions that allowagents to reward cooperation and punish deviations within the relationship. A theoreticalliterature demonstrates how relationships and informal institutions can determine marketstructure and practices.2 Recent empirical work - mainly focused on emerging economies -has shown that relationships are associated with improved contractual terms such as creditprovision and the prioritization of deliveries (McMillan and Woodruff, 1999b; Macchiavelloand Morjaria, 2012; Antras and Foley, 2014). Yet little is known about how market poweraffects the economics of relationships, including how much value is created within the rela-tionship and which contractual institutions emerge to distribute the potential surplus.

In this paper, we provide microeconometric evidence that upstream market structureaffects the value of downstream relationships. We present a model combining imperfectcompetition by upstream manufacturers with a market-share contest by retailers who providetrade credit to attract downstream buyers. Under monopolistic manufacturing, retailerscollude to maintain the provision of trade credit at low levels and use delivery prioritizationto secure buyer loyalties. Manufacturer entry expands the market size and reduces latedeliveries, leading retailers to compete for buyer relationships. Competing retailers expandtrade credit provision as a new basis for buyer loyalty, utilizing the fact that credit requiresa level of inter-temporal trust that can best be sustained inside a relationship.

We test this theory using a novel dataset on high-volume ice sales to 154 fishermen buyersin Sierra Leone, who have long-lived relationships with the ice retailers selling for a localmonopolistic ice manufacturer. We find that a shock that induces upstream entry by severalice manufacturers leads to improved contractual terms for fishermen buyers in both price andtimeliness. Ice manufacturer entry also leads to increased provision of trade credit by retail-ers, but only where multiple ice retailers compete for buyer relationships. Consistent withthe model, we observe a period of intense switching of buyer relationships after manufacturerentry, followed by the reemergence of stable relationships as credit provision increases.

Low-income countries are characterized by many market frictions that help make relation-ships and informal institutions important. These emerging markets often entail uncertaintyand instability, weak formal contractual enforcement, and insecure property rights (Collierand Gunning, 1999; McMillan and Woodruff, 1999a; Johnson et al., 2002). Low-income coun-tries are also characterized by concentrated market power in upstream industries (Venables,2010).3 In such challenging settings, ongoing relationships help support informal contractual

2For example, Baker et al. (2002) extends seminal work by Williamson (1975, 1979) to show how rela-tionships determine where the firm ends and the market begins. Also, historical economic analysis by Greif(1989, 1993, 2006) explores how informal institutions shaped the development of modern markets.

3While data on market structure in low-income countries is generally lacking, detailed research surveyingthe origins of large firms in Ghana, Ethiopia, Tanzania and Zambia further supports that African industrialsectors often entail a small number of leading companies (Sutton and Kellow, 2010; Sutton and Kpentey,2012; Sutton and Olomi, 2012; Sutton and Langmead, 2013).


institutions to address transactional hazards. Focusing on the effect of market structure oninformal contractual institutions enables us to provide original empirical evidence of howupstream market power limits the set of contractual transactions that can be realized.

Sierra Leone is a compelling environment to examine how markets develop when rela-tionships are important. When it emerged from a 10-year civil war in 2002, the country hadmuch of its infrastructure destroyed and a large internally displaced population, highlightinga challenging business environment (Bellows and Miguel, 2006). Since then, economic activ-ity and firm productivity has steadily recovered as trade and investment flows have grown(Collier and Duponchel, 2013).4 Sierra Leone’s ice industry is a simple product market thatrepresents several major themes from previous work on manufacturing in sub-Saharan Africa,such as high capital costs, supply chain risk and the importance of business relationships(Bigsten et al., 2000; Fafchamps et al., 2000; Fafchamps, 2004). Furthermore, observing theice industry during a period of rapid market expansion provides new insight into the mecha-nisms through which upstream market power affects the value of downstream relationships.

A key challenge to studying informal relationships is that appropriate statistics are rarelyavailable, particularly in developing economy contexts like Sierra Leone. To enable the em-pirical analysis described above, we established partnerships with the incumbent ice manufac-turer, its competitors, and the independent ice retailers in order to collect a transaction-levelpanel dataset of informal contracts with fishermen buyers. The data allows us to track iden-tities of the retailer and fishermen buyer, contractual terms (price, quantity demanded andcredit terms), and contractual outcomes (quantity delivered and timeliness of delivery) overan 18-month period from January 2013 to July 2014. We also managed a team of enumer-ators conducting detailed baseline and biweekly follow-up surveys with fishermen to recordassets and expenditures, ice usage, and fishing trip outcomes. Finally, we conduct interviewswith all of the major upstream manufacturers and retailers, providing a qualitative historythat helps complement our quantitative analysis with additional insight into mechanisms.

This data collection allows us to establish several stylized facts. We document that latedeliveries are common under the monopolistic ice manufacturer - on average delayed by ahalf day - and buyers remain loyal to retailers despite systematic poor performance. Approx-imately 26% of deliveries are late during the first six months of 2013, which is prior to newmanufacturer entry, and survey data confirms that fishermen experience worse fishing out-comes when exposed to lateness. Exploring the sources of lateness, we differentiate betweeninternal factors under the manufacturer’s direct control (e.g. issues with machines, vehiclesor workers) and external factors such as outside demand shocks for ice sales to non-fishermensources, and demonstrate a strong positive relationship between outside demand and latedeliveries to fishermen. We present evidence for two main sources of buyer loyalty by exam-ining one fishing wharf, Goderich, that supports multiple retailers under the monopolisticmanufacturer. We demonstrate that retailers in Goderich respond to delays in manufacturerdeliveries by strategically prioritizing their most loyal buyers. Those designated as “loyal

4While the 2014 Ebola outbreak in the Mano River region countries of Guinea, Liberia and Sierra Leoneposes a large economic and humanitarian challenge, the data from this paper focus on the prior period.


customers” move up between 2-3 places in the prioritization queue. Finally, we documentthat retailers are colluding to maintain low trade credit levels and restrict competition forcustomers, thus limiting buyers’ outside options and further cementing existing relationships.

Next, we show that entry by new ice manufacturers results in improved contractual termsin both price and timeliness for fishermen buyers. We exploit a shock that improves thefinancing terms available to new manufacturers in procuring costly ice production machinery,leading to the establishment of several new ice factories serving fishermen over a half-yearperiod. Midway through our data collection, a Sierra Leonean entrepreneur established a newventure importing industrial ice machines, greatly reducing the procurement and financingcosts of purchasing this equipment. Prior to this entrepreneur’s arrival, ice manufacturershad to contract directly with foreign machine sources, visit their facilities abroad, and pay50% of cost upfront, and 50% on delivery; the ice machine importer offered local salesagreements with financing terms of 50% upfront, 25% on delivery, and 25% six monthsinto production.5 Consistent with our model’s predictions, we document increased pricecompetition and fewer late deliveries corresponding to the arrival of new manufacturers ineach of the fishing wharves. Each additional manufacturer is associated with a 5-6% fallin price, and the overall frequency of late deliveries falls to 1% during the first six monthsof 2014. Our results are robust to restricting attention to year-on-year comparisons and toseparately examining the main effects in each wharf.

Finally, we demonstrates that ice manufacturer entry also helps disrupt collusion amongice retailers, leading to more switching by fishermen buyers across retailers and increasedcredit provision. The size of the retail market expands due to increased capacity and lowerprices, making deviation to steal customers more tempting. Meanwhile, lateness has fallen,removing a key market friction that motivated buyer loyalty. Higher credit provision increasesthe value of retailer relationships to buyers and thus serves two purposes, incentivizingthe buyer to switch retailers while also providing a new rationale for loyalty. We exploittemporal variation in the number of ice manufacturers and cross-sectional variation in thenumber of ice retailers in each fishing wharf to demonstrate that trade credit only increaseswhere both manufacturers and retailers compete in Goderich. In the remaining two wharveswhere retailers are monopolists, credit levels do not change after new manufacturer entrydespite lower prices and reduced lateness, suggesting that increases in Goderich are notsolely due to retailer pass-through.6 In Goderich, we observe a 75 percentage point increasein the number of fishermen who have switched retailers at least once after manufacturerentry. Correspondingly, weekly credit provision levels increase up to 29 percentage points inGoderich after manufacturer entry, while no significant changes are correspondingly observedin the other wharves with monopolistic retailers. To address concerns about autocorrelationwithin a small number of wharves, we also implement the wild cluster bootstrap inferencecorrection recommended by Cameron et al. (2008), Cameron and Miller (2013) and Webb

5A secondary source of increased entry into ice manufacturing involves horizontal expansion by packagewater and ice cube distributors who have large cash reserves allowing them to self-finance machinery costs.

6We document in interviews that when ice retailers are monopolists, they extract rents from increasedcompetition by ice manufacturers while new retailer entrants have difficulty establishing trading relationships.


(2013). While short of conventional levels of significance, the p-values from this conservativeapproach are in a range that suggests our findings are unlikely to be due to random chance.

This paper’s findings have important implications for theory and policy. Recent the-oretical work on relational contracts emphasizes the tradeoff between mitigating holdupand facilitating transactions and reducing the potential scope of trade, with implicationsfor intra-relationship loyalty and inter-relationship market structure (Board, 2011; Barron,2013). Our results emphasize an alternative direction in which changes in upstream mar-ket structure directly affect the value of downstream relationships. Namely, under upstreammonopoly, frequent production shortages provide the basis for delivery relationships betweenretailers and buyers that allow loyal customers to mitigate the uncertainty of late deliveries.And after upstream competition increases the market size, expanded credit relationshipsbetween retailers and buyers emerge as a new basis for customer loyalty. Our work also pro-vides rigorous microeconometric evidence consistent with recent macro-level work by Askeret al. (2013), who argue that demand volatility is a key driver of dispersion across countriesin the marginal revenue product of capital, and Michaillat and Saez (2013), who stress theimportance of trade frictions in leading firms to build long-term relationships. With ourfocus on weak institutional environments we are closely related to Dixit (2007), though ourattention to how informal institutions respond to increasing market development is distinct.

We also integrate and extend two literatures on the costs of monopoly power and theconstraints to private enterprise growth in low-income countries, which have generally re-mained separate. In our context, the cost of monopoly includes more than inflated pricesand restricted output relative to the competitive levels, or the welfare implications of asso-ciated rent-seeking behavior (Tullock, 1967; Posner, 1975). We demonstrate that monopolypower by either manufacturers or retailers is a sufficient condition to worsen contractualterms, in our setting limiting the timeliness of manufacturer deliveries and the provision ofretailer trade credit, respectively. This informal channel for monopoly power to constrainfirm growth is very salient in low-income countries, where previous work has largely em-phasized the importance of property rights, credit constraints and uninsured risk in limitingprivate sector development (Johnson et al., 2002; Bigsten et al., 2003; De Mel et al., 2008;Karlan et al., 2012). Furthermore, we distinguish between cooperative and collusive relation-ships, showing that in our setting it is the collusive relationships between retailers that breakdown with the onset of upstream manufacturer competition, while cooperative relationshipsbetween retailers and buyers prove more resilient.

Our work builds on the foundation provided by the economic literature on relationshipsand exchange. A growing body of empirical work addresses the importance of long-livedrelationships in determining firm-level outcomes such as trade credit provision or prioritiza-tion of deliveries (Fafchamps, 1997, 2004; McMillan and Woodruff, 1999b; Antras and Foley,2014). Our work is complementary, noting that even in a context of long-standing relation-ships, collusion and market power may restrict credit to inefficient levels. Macchiavello andMorjaria (2012) provide evidence from the Kenyan rose export sector that reputational incen-tives play a central role in determining firm delivery decisions following a large supply shockassociated with the 2007 Kenyan election violence. Initially, our setting involves frequent


supply shocks that also highlight the value of delivery relationships, but we demonstrate howinformal institutions evolve as market development reduces the risk of late deliveries.

Finally, two recent papers closely related to our own, Macchiavello and Morjaria (2014)and Casaburi and Reed (2014), provide cross sectional evidence on how competition af-fects credit provision that complement our time-series findings. Studying coffee farmers inRwanda, Macchiavello and Morjaria (2014) find evidence that the presence of more cof-fee washing stations in a given geography are associated with fewer relational contractingpractices such as supply of agricultural inputs and post-harvest credit transactions. Their pa-per emphasizes side-selling concerns under which intra-relationship buyer-seller trust breaksdown due to the presence of outside spot markets, and as a consequence buyer credit provisiondecreases with more buyer entry. In our three-tier supply chain, credit is provided insteadby sellers (retailers), and we find that as manufacturer entry expands the retailer marketsize, competing retailers respond by increasing credit provision to secure buyer relationships.Casaburi and Reed (2014) find evidence that cocoa farmers in Sierra Leone working in anarea with more traders receive more credit provision, which they explain with a model ofpass-through. In our paper, a simple model of pass-through is not consistent with the lackof increased trade credit provision we observe in markets where retailers are monopolists,highlighting the importance of inter-retailer collusive relationships in explaining our results.

The remainder of the paper is structured as follows. The next section provides backgroundon our research setting, and the details of ice manufacturing, retailing and fishing in SierraLeone. Section 3 presents a theoretical framework and derives predictions that we test inthe data. Section 4 describes the data. Section 5 quantifies the effect of increased upstreammanufacturer competition on downstream contractual terms. Section 6 concludes.

1.3 Background

A former British colony located along the coast of West Africa, Sierra Leone is a low-incomecountry with an estimated GDP of $4.93 billion (2013) and a population of approximately6 million. Approximately two-thirds of the labor force is engaged in agriculture, whichcontributes over one-third of the country’s gross domestic product (Johnson et al., 2013).Fishing is a vital sector in Sierra Leone, accounting for approximately 10% of GDP andcomposing approximately 30,000 artisanal fishermen distributed along the country’s coastlineusing over 8,000 boats (Bank, 2006, 2009). Fishing is divided into two sectors: industrialoperations, with larger vessels fishing further out to sea that typically freeze fish on boardfor export, and small-scale (artisanal) fishing operations that serve the domestic market. Allof the fishing activities described in this paper involve artisanal fishermen. Sierra Leone’scoastal waters support shrimp, cephalopods, demersal species (e.g. snappers, groupers),small pelagic species (e.g. herring) and large pelagic species (e.g. tunas, barracudas). Fishingproduction is highest during the country’s dry season from October to June, as the heavyrainy season between July and September poses additional risks to fishermen at sea.


Downstream Market: Freetown Fishermen

Over the past decade, a selection of artisanal fishing firms located in the Freetown Penin-sula have developed a specialized approach to fishing production that makes use of the localavailability of ice. Making use of large boats equipped with wooden iceboxes, the fisher-men are able to conduct multi-day journeys. Without ice to keep their catch fresh, thefishermen must either specialize in lower value fish that can be sold dried, or conduct shortovernight fishing trips to avoid the potential for spoilage. Given the fixed costs of organizingeach fishing trip, ice is a simple technology offering scale economies to the fishermen, butunpredictability around the timeliness of ice deliveries constrains its usage.

Artisanal fishermen typically demand ice deliveries in the morning immediately beforegoing to sea, and will make orders the day prior to departure with their local ice retailer.Because fishermen and retailers do not have access to large freezers or a reliable supply ofelectricity, they cannot reliably store the ice and thus require prompt delivery. If the ice isnot delivered on time, the fishermen lose part or all of the day at sea, paying both wages ofthe fishermen they have retained for the day, and the opportunity cost of their own time andcapital. After fishermen have made specific investments such as equipping their boats for icefishing, they have few options in addressing this risk of late ice deliveries. The monopolisticmanufacturer does not offer refunds for late deliveries, which are difficult to predict.

We study three fishing wharves on the Freetown Peninsula, the locations of which arehighlighted in Figure 1.1. Tombo Wharf (W1) is the largest artisanal landing on the penin-sula, representing approximately 250 large fishing boats, though only about 15% of thesevessels regularly purchase ice. While Tombo represents the largest potential market for icesales to fishermen, it is also located approximately one hour’s drive away from the Freetowncity area where the monopolistic ice manufacturer is located and thus particularly affectedby late deliveries. The remaining two wharves, Aberdeen (W2) and Goderich (W3), are lo-cated within short driving distance of the monopolistic manufacturer’s location in Freetown.Aberdeen supports about 100 large fishing boats, of which approximately 30% regularlypurchase ice. Goderich is the largest market for fishermen sales for the monopolistic icemanufacturer, with approximately 200 large fishing boats, of which almost 50% regularlypurchase ice.

Upstream Market: Ice Manufacturing and Retailing

Like many manufacturing industries in sub-Saharan Africa, the industrial production ofice requires relatively large capital investments and high operating costs - two importantbarriers to entry. In particular, large-scale crushed ice production equipment like those usedby these suppliers are custom-assembled by manufacturers in a few countries such as China,Germany or Italy. Capital investment requirements for a 13-ton daily capacity industrial icemachine typically exceed $100,000, with limited available sources of domestic financing. Bycontrast, the machinery for ice cube production is smaller and cheaper, leading to increasedcompetition here. In the case of the monopolistic ice manufacturer discussed below, payment


terms for its industrial ice machine were 50% upfront and 50% on delivery, and required apersonal visit to their international facilities to agree on final terms. Given the unreliablesupply of public electricity in Sierra Leone, ice suppliers must also invest in large dieselgenerators with expensive daily fuel costs in order to ensure a steady production schedule.

Monopolistic Manufacturer

Ice Ice Baby (“IIB”) is Sierra Leone’s largest ice manufacturing company, selling crushedice to artisanal fisherman and other vendors for use as cold storage, and purified cubedice to restaurants and grocery stores for public consumption. The company produces fromits 29-ton capacity facility based in the western area of the capital Freetown and is whollyowned by ManoCap, a private equity fund manager that operates in Sierra Leone, Liberiaand Ghana. IIB was founded in 2005 by two Sierra Leoneans recently returned from thediaspora, and launched its initial production in fall 2006. ManoCap acquired the business in2008 and expanded its production with capital investments and professional management.

At the start of our data collection in January 2013, IIB served as the monopolisticmanufacturer supplying independent retailers (called “agents”) based at the three fishingwharves mentioned above.7 Retailers source orders, collect money from fishermen, and thenbuy ice directly from the manufacturer with their own working capital. Upon receipt ofpayment, the manufacturer arranges delivery to the retailer’s wharf, where ice is distributedto fishermen. Fishermen can also purchase directly from the factory, but then they mustshow up in person, pay a higher price, and organize their own delivery, so it is uncommonfor them to do so.

However, IIB often faces difficulty delivering ice to all wharves early in the morning whenfishermen typically need it. At the start of our data collection in 2013, delays lasting halfa day or more are common and the manufacturer was unable to satisfy all of its fishermendemand in a timely fashion. In interviews, fishermen reported that late deliveries had beena problem for several years prior to the start of our data collection. Lateness for fishermenderives from two main sources: capacity constraints at the ice factory and the unpredictablenature of outside demand. Capital equipment at the factory must be maintained and me-chanical problems with the ice machines, generators, and delivery vehicles occur frequently.As the public electric grid is unreliable, the factory must rely on expensive diesel genera-tors to power its production process. Likewise, factory labor supply can be unpredictable,with employees showing up late or reporting sick, disrupting the implementation of dailyproduction plans.

In addition to these time-varying sources of capacity constraints, the factory faces a dif-ficult planning problem in addressing outside demand. While fishermen orders are generallyplaced a day in advance and aggregated through the retailers, other sources of demand are

7When IIB entered the market in 2006, it initially competed with two smaller existing suppliers withoutdated equipment to serve the artisanal fishermen market, but they were unable to sustain their operationsand soon exited the market. In addition, in 2012 IIB stopped serving a fourth fishing wharf located in MurrayTown when a government sponsored fishing business was established there with its own ice facilities.


much less reliable and experienced directly by the factory. Restaurants and bars, supermar-kets, parties and other vendors often make large, unpredictable same-day orders and have asubstantially higher willingness to pay for immediate service. In the context of time-varyingcapacity constraints, these outside demand shocks often bind, forcing the factory to priori-tize available inventory between fishermen and the outside market. As the average price perkilogram of ice sold to the outside market is about twice as high as the price to fishermenretailers, factory managers regularly prioritize outside demand orders, leading to lateness forthe fishermen. Because ice is a non-durable good, and the manufacturer controls delivery atthe market, retailers are unable to build inventories to protect against late deliveries.

The manufacturer sets the retail price of ice to fishermen and offers a wholesale price tothe retailers, granting them a fixed commission on each 30 kg bag of ice sold. Interviewsand survey data collection with fishermen confirm that this retail price is understood andfollowed in the market, mostly due to retailer concerns that deviating from the rule will leadto complaints from disadvantaged customers. Also, the manufacturer and retailers reportthat price stability is an important determinant of fishermen long-term demand, and hencehave forgone an auction for ice sales. Whenever the manufacturer is responsible for shortagesor lateness in the supply of ice to retailers, retailers also make rationing allocations amongtheir customers. Retailers indicate that rationing decisions are typically made based on thevalue of the business relationship, the order in which deliveries were made, and whether thatorder was paid for in advance. We present empirical evidence supporting this claim below.

Retailers often provide trade credit to their customers, allowing them to pay for a portionof their expenses after the time of order at 0% interest. In practice, trade credit repaymentstake place over a wide interval of time, from the time of delivery, to the return of a fishing trip,to several weeks later. Retailers report that the common determinants of credit provisioninclude the value of the business relationship, reliability in paying back previous debt, andthe stability of the fisherman in his business. On the demand for trade credit, retailers reportthat fishermen who have recently experienced a bad trip are more likely to request credit,particularly given that ice is the final purchase made before a trip departs.8 Retailer tradecredit provision is partly financed by a monthly line of credit from the manufacturer. If aretailer should exceed the monthly credit line, the manufacturer will withhold payment ofthe retailer’s commission until the outstanding debt has been cleared. Thus, trade credit isnearly costless for retailers to provide up to a certain threshold, but the cost rises afterwardsas it begins to impact the retailer’s working capital. Given the observed demand for creditfrom fishermen, it is clear that retailers find themselves on the costly part of the curve.

Relationships between retailers and their customers play a key role in this market context.IIB recruited its retailers directly from the fishing wharf communities, and fishermen reporthaving known their retailer for an average of 9 years even while only buying ice from himfor about half that time. Fishermen invest in retailer relationships to minimize latenessand rationing, and we present evidence below that retailers reward loyal customers with

8Alternative sources of credit available to fishermen are limited. Local banks will not offer credit linesto the fishermen, and informal lenders (e.g. Susu accounts) charge a weekly interest rate of 2-3%.


prioritized deliveries. Retailers also indicate that trade credit is a useful tool for promotingcustomer loyalty, as the repeated interaction involved in sustaining a lending relationshiphelps deters customer switching across retailers and makes credit default uncommon.

In addition to the vertical relationships between retailers and fishermen, we documentthe importance of horizontal relationships among the three retailers who operate in theGoderich wharf. In interviews, these retailers report the existence of a collusive agreementthat restricted competition for fishermen customers. Specifically, retailers agreed to main-tain exclusive territories over existing customers, so as to avoid making competing offers tocustomers. This agreement was supported by the existence of a single manufacturer, whichlimited potential sources of ice for new retailers and enabled the retailer cartel to handletransitions. In one retailer’s words: “the main reason was to prevent the agents to competeamong themselves with the same wharf. This is why we the agents agreed that IIB cannotget more than 3 agents here at Goderich. Even if an agent wants to quit, he can recommendanother person and we should approve because we want an agent that is dependable.”9

Entry and Competition

Here we describe a shock that set into motion the chain of events explored in this paper.Given interviews, we can note that the shock was unexpected to the participants and unre-lated to underlying fundamentals in any of the fishing wharves. As noted above, the crushedice manufacturing market has high fixed costs associated with the acquisition and financingof required machinery, which served as an important barrier to entry. Two factors played animportant role in overcoming this restriction on entry. First, in late 2012, a Sierra Leoneanentrepreneur with an established business importing refrigerators and air conditioners forthe regional consumer market decided to diversify into importing industrial ice machines.The importer offered local buyers an opportunity to purchase without costly travel, andwith attractive leasing terms that helped smooth the otherwise lumpy investment decision.In practice, this entailed adjusting the standard payment terms from 50% upfront and 50%on delivery, to 50% upfront, 25% on delivery, and 25% six months into production. Theimporter experienced considerable demand for his services, helping to establish six new iceproduction facilities around Freetown between early-2013 and mid-2014, including severalfocused on serving fishermen.

Second, an existing set of firms operating in Sierra Leone’s water packaging and cubedice market, which have lower entry costs, were able to gradually build up sufficient cash re-serves from these business lines to finance the necessary large capital investments. From theperspective of these competitors, horizontal differentiation into the crushed ice market pro-vided an attractive opportunity to diversify their revenue streams while leveraging existinginvestments in land, buildings, managerial capacity and sales networks.

As a consequence of these two factors, four manufacturers with industrial crushed icemachines chose to enter the three fishing wharves already served by IIB. The first two new

9Interview with research associate James Polit, October 2014, Goderich Wharf, Freetown, Sierra Leone.


manufacturers self-sourced and self-financed their expansion into industrial ice production,while the remaining two utilized the importer’s sourcing and financing services. The exacttiming of the start of sales for each factory to artisanal fishermen depended on a variety ofidiosyncratic factors, such as the recruitment of skilled staff, the maintenance and repair ofmachinery, and the limited attention of senior managers or owners. While IIB was awareof the possibility of competitor entry into the artisanal fishing market, the factory ownersand management indicated in interviews that they did not consider new entry to be a likelyoutcome prior to the fact and that they were unaware in advance of the timing of entry.

Figure 1.2 presents the timing of entry by each of the four new manufacturers who com-peted with IIB, the original monopolistic manufacturer, for serving the demand of artisanalfishermen in the three fishing wharves. It also shows the corresponding decisions of IIB tolower price across all three wharves, in particular following the entry of lower-price competi-tors into Goderich and Aberdeen. Manufacturer 2 originally started operations in May 2012as a packaged water distributor based in the near vicinity of the Tombo, and began sellingcubed ice in May 2013 to local bars and hotels. After procuring a crushed ice machine,it launched sales in Tombo beginning in late-August 2013.10 Manufacturer 3 began opera-tions in February 2013 serving cubed ice to the general public from its location in centralFreetown, and then began selling cubed ice to fishermen in both Goderich and Aberdeensimultaneously starting in late-October 2013.11

Manufacturer 4 began operations in May 2012 as a packaged water distributor and addedcubed ice sales in December 2012. While intending to enter the fishermen market earlier,they faced delays in procuring their crushed ice machine and recruiting skilled staff, some ofwho came from IIB’s staff. Manufacturer 4 launched crushed ice sales in Aberdeen nearbyits factory location in December 2013, and then expanded to Goderich in April 2014. Man-ufacturer 5 started both cubed and crushed ice operations in January 2014 from a locationnear the Goderich. After selling crushed ice to the general public for several months, Man-ufacturer 5 began crushed ice sales to fishermen in Goderich in March 2014, with plansto enter Aberdeen and smaller wharves around Freetown in the future. Appendix FiguresA1.1, A1.2, and A1.3 show the entry timing and pricing decisions of all manufacturers brokendown by Tombo, Aberdeen and Goderich, respectively. As we discuss in the results on pricecompetition below, the common presence and coordinated pricing of IIB in all three wharvescontributed to competition spillovers from the entry of manufacturers in outside wharves.For example, IIB maintained the same prices across all three wharves, so prices dropped inTombo in response to additional entry in Goderich and Aberdeen.

10As Figure 1.2 shows, the incumbent manufacturer, IIB, did not respond immediately to the entry ofManufacturer 2 in Tombo. This was partly due to the fact that Tombo is the most distant wharf served byIIB with a maximum of only one delivery per day. In addition, Manufacturer 2 chose not to compete onprice, using its spatial advantage to maintain a high price against the prospect of immediate deliveries.

11For logistical reasons, fishermen customers prefer crushed ice to Manufacturer 2’s cubed ice product,and factory mechanical issues led them to exit Goderich in December 2013 and Aberdeen in February 2014.


1.4 Theoretical framework

In this section, we develop a theoretical model that links competition in upstream manufac-turer and retailer markets to the contractual terms faced by downstream buyers. Based onour empirical setting, manufacturers compete on quantities to determine aggregate supply,while retailers compete on credit provision to secure their share of the buyers in a market.Using a standard Cournot model, we show that more competition among manufacturersresults in lower prices and higher aggregate quantity sold. An implication of the model isthat manufacturer competition also reduces late deliveries. We then focus on two categoriesof inter-temporal relationships: collusive interactions among multiple retailers seeking to re-strict the level of competition for customers, and cooperative interactions between a retailerand a buyer to prioritize deliveries or sustain credit provision in the absence of externalenforcement. We find that upstream manufacturer competition disrupts collusion amongretailers, leading to increased credit provision by retailers. Once a retailer extends morecredit to a customer, the relationship they forge may become more valuable.

We begin by examining the relevant decision problem facing economic actors at threelevels of the supply chain: the manufacturer, the retailer and the downstream buyer. Wemodel the decision problem of N manufacturers (each denoted by Mi), as a Cournot game inwhich the inverse demand, P (Q), is a function of aggregate quantity, Q, and manufacturerMi chooses a level of production, qi taking into account a constant marginal cost, m > 0,and a commission paid to the retailer on each sale, α > 0:

maxqi

πMi= (P (Q)−m− α)qi (1.1)

The α commission reflects the imposition of retail price maintenance by manufacturers,under which retailers do not have discretion in setting the retail price. This special caseamounts to fixing the difference between the retail price faced by buyers and the wholesaleprice faced by retailers to equal α, which we treat as an exogenous parameter.12

We model the decision problem facing two symmetric retailers (each denoted by Rj) whohave a revenue function R that take α as an input and a cost function C that takes as aninput r, the marginal cost of providing credit to each buyer. Both R and C also depend onaggregate quantity, Q, the share of the each order that the first retailer provides on credit,cj ∈ [0, 1], and the symmetric credit decision of the competing retailer, c−j ∈ [0, 1].

maxcj∈[0,1]

πRj= R(α,Q, cj, c−j)− C(r,Q, cj, c−j) (1.2)

Finally, we consider a continuum of downstream buyers (fishermen) with mass 1, whichwe model as firms facing a common production function, Y (Q), who act as price takers for

12In our empirical setting, retailers abide by the retail price and do not offer special discounts to theirbuyers. Interviews suggest two main reasons for this practice: first, both manufacturers and retailers reportbuyers will react negatively to price volatility by reducing future purchases of the good, and second, retailersexpress concern that price discounts offered to some buyers will have to be shared with all customers.Furthermore, most retailers are former fishermen, so α can be interpreted as tracking the reservation earningsunder this option of outside employment.


a market price of their output, pF . For simplicity, output depends solely on the quantity ofthe intermediate good demanded, Q, which they purchase at price P .

maxQ

πB = pFY (Q)− PQ (1.3)

Manufacturer Competition

With manufacturers competing in a Cournot game, we assume a linear inverse demandfunction, P (Q) = A− BQ, for analytic convenience. Given the buyer’s decision problem in(1.3), this is equivalent to assuming a quadratic functional form for the production function.13

Using standard approaches, we derive the following expressions for price and quantity as afunction of exogenous parameters A,B,m, α, and the number of manufacturers, N . Observebelow that price is decreasing in N , while aggregate quantity is increasing in N .

P ? =A−mN − αN

N + 1Q? =

(A−m− α)N

B(N + 1)(1.4)

Next, we discuss the implications of the model for the issue of late deliveries. As discussedin the above background section, under the monopolistic manufacturer idiosyncratic produc-tion problems (e.g. machine breakdowns) and demand shocks from outside markets often ledto shortages in available production capacity, which in turn lead to shortfalls in the availablesupply for the buyers’ market and subsequent late deliveries. We define the probability thatan upstream shock to the manufacturer leads to late deliveries for buyers as p ∈ [0, 1], andnote that a further implication of new manufacturer entry is that p should fall, or dp

dN< 0.14

The micro-foundations for this assumption rely on each new manufacturer increasing thetotal capacity in the industry, and also introducing a new idiosyncratic likelihood of produc-tion problems or outside demand shocks. More formally, we can write p =

∏pMi

, so thatthe total probability of lateness is the product of the independent probabilities that eachfactory is capacity constrained. This directly implies that the available supply will be morelikely to satisfy buyers’ aggregate demand as the number of manufacturers increases.

Retailer Competition

We proceed by solving for retailer decisions on credit provision in the context of a repeatedmarket share game. First considering the per-period game, our main assumption is thata share of buyers, 0 < γ < 1, are completely credit constrained each period and can onlypurchase using the trade credit from the retailer, but that the retailer cannot differentiatethe γ types from the (1−γ) types who will accept any credit offered but purchase whether or

13This can be shown using Y (Q) = aQ− bQ2, on the range Q ∈ [0, ab ] where a, b > 0 are arbitrary scalarvalues. Substituting into Equation (1.3) and solving yields P (Q) = A−BQ, where A = apF and B = 2bpF .

14We assume that lateness is a temporary cost imposed on fishermen that does not affect long-termdemand, and thus does not affect the market equilibrium.


not they receive credit.15 We adopt a perfectly discriminating contest, assuming that buyerspurchase from whichever retailer offers them a higher credit level, though while the (1− γ)types enjoy their full demand, the γ types purchase only as much as they can afford basedon the retailer’s credit provision. Retailer profits are as follows:

πRj(cj) =

α(Q/2)[(1− γ) + γcj]− cjr(Q/2) if cj = c−jαQ[(1− γ) + γcj]− cjrQ if cj > c−j0 if cj < c−j

(1.5)

If both retailers offer the same amount of credit, they split the market evenly, eachreceiving α(Q/2) times the share of the market served. The share of the market serveddepends on the choice of cj, given that (1−γ) are going to buy either way but the remainingγ will only buy up to the share of credit provided, cj ∈ [0, 1]. The total cost of creditprovision is cjr(Q/2), as each retailer ends up only providing credit cj at marginal cost r totheir half of the market. Note that the total cost does not depend on γ as credit goes to allcustomers regardless of type. In the second case, when cj > c−j, retailer Rj claims the entiremarket, where the share of the market served depends on the choice of cj as above. Observethat this also would be the case if a retailer were a monopolist in a given market and thusfaced no competition for customers. Finally, when cj < c−j, retailer Rj loses the market toa competing retailer and receives nothing.

We restrict attention to αγ < r < α, which defines the range of values for which collusionamong retailers is profitable and yet retailers would want to provide credit only to the γ shareof credit constrained types if possible. To see this, first consider a retailer’s decision if hecould target only the γ credit-constrained types. In this scenario, the retailer earns α andpays r for each unit of credit provided, so his participation constraint is given by r < α.Now if the retailer cannot identify the γ credit-constrained types (as is the case here), he willearn instead αγ while still paying r on each unit of credit provided, implying that collusionto restrict credit provision will be profitable if αγ < r.

Next, we use the payoffs from the per-period game as parameters in an inter-temporalcooperation game played by the two retailers. We consider three cases, a “Retail Competi-tion” case to the stage game where both retailers provide full credit, cj = c−j = 1, a “RetailCollusion” case where both retailers provide zero credit, cj = c−j = 0, and a “Defection”case where retailer Rj steals the entire market with cj = ε while c−j = 0. In the final case,we assume that retailer Rj faces a one-time fixed cost of deviating from the collusive equilib-rium, denoted with κ, which can be interpreted as a mobilization cost of stealing customers.Retailer Rj’s profits are:

πCOMPETITIONRj

=(Q/2)(α− r)πCOLLUSIONRj

=(Q/2)(α(1− γ))

πDEFECTIONRj=Qα(1− γ)− κ

(1.6)

15In our empirical setting, credit constraints vary greatly over time due to the stochastic nature of fishingproduction processes, making them difficult for retailers to independently verify. The parameter γ can thusalso be interpreted as the probability that any given buyer is credit constrained in a given period.


Observe that as αγ < r, profits are strictly higher in the collusion case than the compe-tition case, but for arbitrarily small values of κ the defection case will provide the greatestpayoff.16 A standard repeated game approach provides the necessary condition below forthe collusion outcome of zero credit provision by both retailers to be dynamically incentivecompatible, where δ represents the common discount factor of both retailers.

πCOLLUSIONRj

1− δ≥ πDEFECTIONRj

+δπCOMPETITION

Rj

1− δ(1.7)

Solving for δ yields a minimum patience level required to sustain collusion, Φ(Q).

δ ≥ Qα(1− γ)− 2κ

2Qα(1− γ)−Q(α− r)− 2κ≡ Φ(Q) (1.8)

When collusion is not possible because δ is below this threshold, retailers play the competitionequilibrium offering full credit provision. The threshold is a function of aggregate quantity, Q,and other parameters that are exogenous. As it can be easily shown that Φ(Q) is increasing inQ, more patience is required to sustain collusion as the aggregate quantity grows, providinga direct link between manufacturer competition and retailer competition.

Buyer Relationships

Buyers and retailers form valuable relationships in the context of repeated purchases. Ongo-ing relationships help retailers dampen demand volatility through a steady flow of purchases,providing a rationale to incentivize buyer loyalty. We consider two instruments available toretailers for promoting loyalty: delivery prioritization and credit provision.

The expected cost of late deliveries is pω, where p is the above probability of a supplyshock and ω > 0 is the buyer’s cost of delay. Imagine a retailer has the option to fulfill somebut not all of their orders on time after a supply shock, and must choose a rationing rule.Consider two candidate rules: a random prioritization rule in which each order has an equalprobability of being delivered on time, and a loyalty-based rule in which buyers who boughtfrom a given retailer in previous periods are more likely to receive their delivery on time withties among loyal customers decided randomly. As retailers earn a fixed commission for eachpurchase, they choose a rule to maximize their number of buyers. Buyers choose the retaileroffering the lowest likelihood of exposure to late deliveries. Once a buyer has bought froma given retailer in the past, it follows that a loyalty-based rule weakly dominates a randomprioritization rule, and is strictly better whenever there are non-loyal buyers in the pool.17

16More formally, we require that κ < α(1− γ)/2 for the one-time payoff of defection to be attractive.17More formally, consider a sequence of events in which each retailer announces a rationing rule, each

buyer chooses a retailer to place orders, a supply shock is realized, and then retailers fill orders accordingto their rules. Further, assume the number of buyers is distributed equally across retailers. If all buyers areloyal, then a buyer will have a equal probability of being prioritized by returning to a retailer who rewardsloyalty relative to a retailer who uses a random allocation rule. If some buyers are not loyal to that retailer,they will not be prioritized, raising the effective probability of prioritization among the retailer’s loyal buyers.


Credit provision can also provide a potential rationale for buyer loyalty. Credit provisioncan be modeled as a trust game in which retailers provide buyers with the ability to pay inthe future at some risk of default. In a weak institutionalized environment, such cooperationfundamentally relies on the logic of repeated interactions, in which future interactions canbe conditioned on past behavior - namely, the prospect of a breakdown in trade deters therisk of default. As discussed above, buyers face a probability γ of being credit constrainedin each period, and are thus likely to choose a retailer on the basis of who announces themost generous credit policy. Recall that a credit constrained γ-type faces a constrainedproduction choice, Y (cQ)− PcQ. If two retailers, Rj and R−j, make competing generalizedcredit offers cj and c−j, then we know from above that buyers will choose whichever retaileroffers more credit. But if retailers attempt to differentiate by offering higher credit levels totheir customers who have paid them back in the past, this provides a potential new frictionto increase buyer loyalty.

Taking these considerations into account, a reduced form approach that represents thecombined value of these two instruments to promote loyalty is an endogenous switchingcost, η, paid by the buyer whenever changing retailers between successive periods. Forconvenience, we assume that a retailer will prioritize loyal customers following a supplyshock with probability 1, so that switching to a new retailer implies an expected deliverycost of pω. The switching cost of credit provision is given by probability of credit constraints,γ, times the difference in profit outcomes under two competing credit offers, [(pFY (cjQ) −PcjQ) − (pFY (c−jQ) − Pc−jQ)], where Rj is the incumbent retailer and R−j is the newretailer. Note that depending on the relative values of cj and c−j, this second term could bepositive or negative, implying in turn that the entire expression for η could also be positiveor negative, reflecting the potential benefits or costs to switching retailers.

η = pω︸︷︷︸Lateness

+ γ[(pFY (cjQ)− PcjQ)− (pFY (c−jQ)− Pc−jQ)]︸︷︷︸Credit

(1.9)

The above formulation of switching costs allows some brief observations in the context of thediscussion in the previous two sections of late deliveries and credit provision. First, wheneverη > 0, we should not observe switching of buyers across retailers. Under the monopolisticmanufacturer, p is high, implying that late deliveries are common, and retailers collude torestrict credit provision, with cj = c−j = 0. Hence, η will be positive and switching willbe uncommon. But as new manufacturers enter, p will fall, with late deliveries becomingless common, and retailers will find it more difficult to collude on restricting credit provisiongiven the growing market size. During the transition from the no-credit equilibrium tothe high-credit equilibrium, buyer switching should occur whenever η is negative, implyingthat the new retailer’s credit offer exceeds the incumbents offer. But once the high-creditequilibrium is reached, buyer switching should decrease as retailers’ credit offers equilibrate.


Predictions

Combining the above discussions of manufacturer competition, retailer competition andbuyer relationships, we can now develop a set of predictions to guide our empirical analysisof the transition from monopoly to competition in the upstream manufacturing industry.

Proposition 1. Increased competition among manufacturers will result in lower prices,higher aggregate quantities and fewer late deliveries.

Proof. From the equilibrium conditions in Equation (1.4) above, we can directly show thatdPdN

< 0 and dQdN

> 0. In addition, the condition dpdN

< 0 directly states that the frequency oflate deliveries will fall as the number of manufacturers increases.

Proposition 2. As manufacturer competition increases the aggregate quantity to be sold,competing retailers will increase credit provision while monopolistic retailers will not.

Proof. Recall that where retailers are monopolists, they do not face a competitive pressureto increase credit provision. But where multiple retailers serve a market, they can lose buyersto alternative retailers offering higher levels of credit. As the number of manufacturers, N ,increases, we note from above that aggregate quantity, Q, will also increase. But retailercollusion to restrict credit provision becomes more difficult as aggregate quantity increases,dΦ(Q)dQ

> 0. Integrating the manufacturing and retailer stages, we note that aggregate quan-tity, Q, increases in the number of manufacturers, N , implying that the patience thresholdto sustain retailer collusion, Φ(Q), also increases in the number of manufacturers. Thus, asufficiently large increase in manufacturer entry will lead to an increase in credit provisionas retailers transition from the collusion equilibrium to the competition equilibrium.

Implication 1. Increased competition among manufacturers will reduce the incentives forloyalty, temporarily leading to more switching by buyers across retailers.

As discussed above, incentives for loyalty are reorganized following the onset of manufac-turer entry, as the probability of lateness, p, is falling in the arrival of each new manufacturerthat adds additional capacity. During the period of equilibrium transition from retail col-lusion to retail competition, retailers may be offering different levels of credit provision tobuyers. As retailers settle into the high-credit equilibrium, buyer switching should decreaseas the difference between credit offers decreases and retailers begin to condition credit levelson buyer loyalty. By increasing credit provision levels, retailers provide a short-term benefitto credit-constrained buyers and a new rationale to motivate future business interactions.

1.5 Data and Descriptive Statistics

From January 2013 until June 2014, we organized a large-scale original data collection ex-ercise with actors at all levels of the supply chain: upstream ice manufacturers, mid-levelice retailers, and downstream fishermen buyers. With manufacturers, we observe data on


price schedules and quantities produced and sold. Within the incumbent manufacturer, weobserve daily aggregate totals of ice produced and daily aggregate sales data, broken downby retailer or other customer. We recently received similar data from each of the four com-petitor manufacturers, and it will appear in a future draft of this paper. In addition, weconduct interviews with the leadership of each ice factory on strategic decisions related tocompetition and policies on production, delivery, credit, and other industry context.

We also observe transaction level data from the original five independent retailers servingthe incumbent manufacturer on their informal contracts with fishermen customers, includingthe buyer identity, contractual terms (price, quantity demanded, credit), and contractualoutcomes (quantity delivered and timeliness).18 We do observe sales by these retailers thatare sourced from competitor manufacturers. Since the incumbent manufacturer data includesdaily aggregates for each retailer, we independently verify that the individual orders anddeliveries they report are closely correlated to the aggregate supply and payments notedby the factory managers. In addition, we conduct interviews with each of the original fiveretailers on their strategic decisions, and collect rankings of their customers across multipledimensions including buyer loyalty.

Finally, we use the original retailer records to create a sample of all current fishermencustomers of the incumbent manufacturer, and conducted a face-to-face baseline surveyin April 2013. The baseline survey includes information about respondent demographics,fishing practices, experience with the ice retailers, assets, expenditures, and social networkties. We were able to locate and survey all current regular customers at the time of thebaseline survey, and continued to add new fishermen customers to our survey data collectionas they entered the sample. Starting in May 2013 and continuing until July 2014, fishermenreceived brief follow-up biweekly phone surveys that addressed their fishing trips over thepast two weeks, including the use of ice, the selection of retailers, and fishing trip outcomes.An endline survey in July 2014 collected updated assets and expenditure information, aswell as recall data for fishing practices and outcomes over the previous two and a half years.

Table 1.1 summarizes the demographic variables associated with our sample of respon-dents. With an average age of 40 years, and over 15 years of years of fishing experience,about sixty percent of the fishermen respondents owned their own fishing boat and overeighty percent served as boat captain for regular trips.19 Fishermen reported that a typicalmonth during the dry season involved almost 10 trips, and that they had known their iceagent retailer for an average of nine years. In the high frequency data, we observe that theaverage planned trip length is almost 3 days, with one-quarter of trips not involving ice. Theaverage ice purchase was 460 kilograms (∼$75 using the ice prices at the start of the data),though this rises to 615 kilogram (∼$85) conditional on making an ice purchase. Average

18Similar high-quality transaction data is not available for the other retailers who enter following manu-facturer entry, but we document that their aggregate sales are limited and thus unlikely to bias our estimates.

19As our sample was defined as customers of the ice retailer and not by the owners or captains of boats,these are not mutually exclusive groups. This sample of 150 fishermen customers includes 26 owners whoare not captains, 59 captains who are not owners, with the remainder fitting into both categories.


trip gross margins, defined as total trip revenues minus total trip expenses (including laborcosts), were approximately $110, with a large standard deviation of about $150.20

1.6 Results

In the results that follow, we consider the effects of upstream manufacturer entry on multi-ple features of downstream contracts: prices, lateness and credit provision. In each of thesecategories, we demonstrate that a shock that increases upstream competition among manu-facturers improves the contractual terms offered by retailers to buyers. We exploit the panelnature of our data and the plausibly exogenous timing of manufacturer entry to perform em-pirical event study analyses to quantify the magnitude of changes in downstream contracts.Before doing so, we document that under the monopolistic manufacturer late deliveries arecommon, yet relationships between retailers and buyers are stable. When manufacturerscompete, prices fall and services improve with fewer late deliveries. Competition by manu-facturers triggers competition among retailers, leading to increased buyer switching acrossretailers followed by increases in trade credit. Consistent with our theoretical model, theseincreases in trade credit provision only take place where multiple existing retailers serve awharf, as is the case in Goderich. When retailers are monopolists in a wharf, as is the casein Aberdeen and Tombo, we observe no corresponding increase in trade credit.

Lateness and Delivery Relationships Under Monopoly

First, we document that late deliveries are common under the monopolistic ice manufacturer,and buyers remain loyal to retailers despite systematic poor performance. From January toJune 2013, prior to the start of the rainy season, approximately 26% of fishermen orders weredelivered late, often entailing a half-day delay to the planned departure of the fishing trip.Fishermen indicated in focus group discussions that the unpredictability of late deliveriesmade it difficult to plan for future trips, and imposed real costs on their fishing production.In Figure 1.4 we observe the daily ice production totals in kilograms for IIB, the incumbentmanufacturer, and the aggregate share of orders arriving late to all fishing wharves. It isstriking to note the daily variation in the likelihood of late deliveries during Jan-June 2013.

In exploring the sources of lateness, we differentiate between internal factors under themanufacturer’s control (e.g. issues with machines, vehicles or workers) and external factorssuch as outside demand shocks for ice sales to non-fishermen sources. In interviews, fishermenwere most likely to report issues with machines, vehicles or workers as the proximate causeof late deliveries, but retailers were aware that an important component of late deliverieswas the level of demand from outside markets. Figure 1.5 shows the separate time series forIIB’s sales to fishermen and non-fishermen customers, where fishermen sales include retaileraggregate orders and a small number of direct sales to fishermen at the factory location. Weobserve that the two are positively correlated due to seasonal patterns - for example, during

20About 14% of trips had negative profits, reflecting the risky production process in fishing.


the rainy season from July to September, demand drops for both sources - and that bothcontain a high level of daily variation.

To demonstrate how outside demand shocks affect late deliveries to the fishermen, weaggregate our administrative data on factory production and retailer deliveries by wharf, w,and day, d. We then estimate the following regression, which includes time-varying weathercontrols, Xd, and wharf fixed effects, ηw interacted with year by month fixed effects, τm.21

Latenesswd = θ1Log(Non-Fishing Sales)d + ηw ∗ τm +X ′dβ + εwd (1.10)

In the above equation, Latenesswd measures the fraction of fishermen orders which are de-livered late to a fishing wharf, and Log(Non-Fishing Sales)d measures the log of daily factorysales (in kilograms) to non-fishing sources of demand.22 The results from this specification,shown in column (1) of Table 1.2, show a strong positive correlation, which can be interpretedas an 1% increase in outside demand is associated with a 5.2 percentage point increase inthe share of orders delivered late during the production period prior to new manufacturerentry. The magnitude of this relationship increases to 10.8 percentage points in column(2) when restricting attention to the Jan-June 2013 period prior to the 2013 rainy season.In interviews with IIB factory leadership, the underlying mechanism here appears to be apractice of prioritizing the allocation of production to time-sensitive outside demand withhigher profit margins over the repeated high-volume but lower margin fishermen demand.

In Figure 1.6, which focuses on the Goderich Wharf served by multiple retailers, we ob-serve that buyers rarely switch between retailers during the period of common late deliveriesprior to new manufacturer entry. In interviews, fishermen were reluctant to criticize theirretailer for poor performance, often directing responsibility on the manufacturer for late de-liveries. While sensible in the context of monopolistic retailers in Tombo and Aberdeen, thisresponse appears puzzling in the context of Goderich where buyers might attempt to induceadditional retailer effort ensuring timely deliveries from the manufacturer through the threatof switching to an alternative retailer with less lateness.

One factor in explaining this outcome is the importance of delivery relationships in whichretailers prioritized their most loyal customers, namely those who would not purchase froman alternative retailer, on days in which late deliveries were likely. In Table 1.3, we documentthis strategic prioritization in order to reward loyal buyers and strengthen relationships. Werestrict attention to Goderich Wharf in the period between January and October 2013 priorto new manufacture entry, and only consider days on which some orders are delivered latebut others are delivered on time. We find that a retailer’s designation of a fisherman as a“Loyal Customer” is positively correlated with having their order prioritized to arrive on-

21Time-varying weather controls include average daily temperature, rainfall and windspeed on the day ofice order, which are factors that fishermen report are important in their trip planning process.

22While we do not have a direct measure of non-fishing demand, interviews with factory managementconfirm that sales to non-fishing sources are a reasonable proxy for non-fishing demand given a willingnessto prioritize non-fishing orders.


time.23 The magnitude of the prioritization is approximately equivalent to moving up threeplaces in the prioritization queue (i.e. the sequence in which the retailer received orders) incolumn (2) without controlling for retailer identity or time period, or 2.2 places in the queuein column (6) when including fixed effects for retailer and week.

Changes in Market Structure

Figure 1.3 shows the pre- and post-competition market structure of each of the three fishingwharves. The onset of upstream competition resulted in limited successful entry into theretailer market, and existing retailers were able to maintain a dominant market position.Turning to the figure, Tombo and Aberdeen, shown in panels (A) and (B), respectively,each start out with a monopolistic manufacturer and a monopolistic retailer, while Goderichin panel (C) has three retailers. This pre-existing cross-sectional difference in retailer con-centration is due to market size and the proximity to the factory location. Tombo, whichhas more fishermen than Goderich but is further away, only receives one retailer, as doesAberdeen, which is closer to the incumbent manufacturer’s factory location but smaller thanGoderich. The existence of pre-existing differences in retailer concentration in this settingprior to manufacturer entry is ideal for testing the predictions about retailer competition de-veloped in the theoretical section above. Recall that varying retailer concentrations impliesdifferential predictions on the provision of trade credit after new manufacturer entry, wherewe expect credit provision to rise in Goderich due to increased retailer competition, but notin Aberdeen or Tombo.

As of July 2014, we observe in Tombo that both the incumbent Manufacturer 1 andthe competitor Manufacturer 2 are selling through the same original retailer. Citing con-cerns about pricing power from the new upstream manufacturers that had located nearbyto Tombo, Retailer 1 in Tombo negotiated to continue to receive next day deliveries fromthe original monopolistic supplier while taking same-day orders from the new upstream sup-plier. In Aberdeen, despite the entry of two new retailers, only the original retailer fromthe pre-competition period is actively making sales in July 2014. Retailer 2 in Aberdeenaccepted an offer from Manufacturer 4 to receive a higher commission on each sale in ex-change for a debt contract. This resulted in a de facto exclusive territory for Manufacturer 4in Aberdeen, though Retailer 2 reserved the option to order from Manufacturer 1 wheneverManufacturer 4 faces difficulty meeting demand.24 And in Goderich we observe that the twolargest original retailers, 3 and 4, continue to sell for the incumbent Manufacturer 1 whilealso directing orders to the competitor Manufacturer 4.25 The last entrant to the upstream

23Approximately 74% of fishermen customers were designated as loyal by the three retailers in Goderichwharf in a survey completed in September 2013 prior to new manufacturer entry into Goderich wharf.

24Manufacturer 3 and the associated Retailer 6 left the market after several months of unsuccessful effortsto attract customers, and efforts by Manufacturer 1 to introduce a new retailer, 7, to the market were alsolargely unsuccessful. Retailers 6 and 7 reported that it was hard to convince fishermen to leave Retailer 2.

25Manufacturer 3 only briefly served this market, making sales directly to fishermen customers in a spotmarket, before exiting. Retailer 5, who served the market prior to the onset of upstream competition, tookan extended hiatus from the wharf, directing his small number of customers to source from other retailers.


market, Manufacturer 5, conducts sales through its own exclusive agent, Retailer 8.It is noteworthy that the total number of retailers does not change in each market when

comparing June 2013 to June 2014, which is useful for ruling out specific effects due tochanges in retailer concentration. In general, we observe that despite the entry of several newmanufacturers, there is limited successful entry by new retailer intermediaries. In interviews,manufacturers, retailers and fishermen highlighted the role that existing relationships playedas a barrier to new retailer entry. New manufacturers were reluctant to introduce newretailers where incumbent retailers had existing relationships with their buyers and werewilling to sell for more than one factory. Incumbent retailers were then better positioned tooffer reliable access to ice across multiple factory sources, and could use existing relationshipsas a basis to compete for buyers through improved credit provision. And fishermen expressedreluctance to engage new retailers who might later exit the market and could not providecompetitive credit terms.

Effects of Manufacturer Entry on Prices and Lateness

In the first main result of the paper, we find that increased competition among ice man-ufacturers leads to improved market outcomes for fishermen buyers in terms of price andtimeliness. We exploit the timing of manufacturer entry, which we attribute to changes in thecapital stocks and financing terms available to procuring costly ice production machinery,and not to any underlying changes in the fishing wharves served by the incumbent manufac-turer. Consistent with our model of imperfect competition, we document lower prices andfewer late deliveries following the arrival of new manufacturers in each of the fishermen mar-kets. Future paper drafts will include results on the change in aggregate quantity, pendingdigitization of competitor manufacturers’ data.

Table 1.4 documents the effect of manufacturer entry on changes in average prices faced byfishermen purchasing from the original five retailers who served the incumbent manufacturer.We include data on all fishermen purchases recorded by these retailers, including orderssourced from competitor manufacturers, though our results are robust to only selecting IIBorders. We aggregate this transaction data by wharf, w, and week, t, and calculate the logof the mean price paid for ice in each market in each week as our dependent variable. Wethen estimate the following regression, which includes wharf fixed effects, ηw, where the keyexplanatory variable is the number of manufactures operating in a given wharf.

Log(Average Price)wt = θ1Number Manufacturerswt + ηw + εwt (1.11)

In columns (1) and (2) of Table 1.4, we show that each additional manufacturer is asso-ciated with a 5-6% fall in price, depending on the time period used. The common presenceand coordinated pricing of IIB in all three wharves contributed to “competition spillovers”from the entry of manufacturers in outside wharves. In columns (3) and (4), we includecontrol for the number of unique manufacturers selling in outside wharves, demonstratingthat approximately half of the decrease in price can be attributed to competition spillovers


from entry into other wharves.26 Finally, in columns (5) - (8) we test for differential effectsin Aberdeen and Goderich wharves, which unlike Tombo are located in the western coast ofFreetown closer to the bulk of manufacturers. When including a control for the number ofoutside manufacturers, we do find evidence that price competition effects were more intensein this area, with each contributing 2-3% lower prices per new manufacturer. However, thedifference between the estimated coefficients for the interaction terms with Aberdeen andGoderich wharves is not significant, which suggests that the larger market size and largernumber of retailers in Goderich did not lead to differentially greater price competition. InTable A1.1, we find qualitatively similar results when restricting attention to each wharfindividually.

With respect to the effect of manufacturer entry on late deliveries, we note that latedeliveries recorded by the five original retailers in January-June 2014 fell to only 1% of allorders. In columns (3) and (4) of Table 1.2, we confirm the magnitude of this drop asalmost 25 percentage points. We do not observe empirical evidence of differential effects byeach additional manufacturer in a market. Consistent with this, in interviews the retailersreport using their relationships with competing manufacturers to smooth shocks to supplyand demand, often checking the capacity of each manufacturer to supply orders on a givenday and threatening to redirect orders to a competitor if deliveries arrived late. Returningto Figure 1.4, it is striking to note the visual decrease in the propensity for late deliveriesafter manufacturer entry into Goderich and Aberdeen. While there is a major decrease inlateness during the rainy season from July-September when overall production decreases atthe factory, lateness returns again as orders pick up after the rainy season. Late orders thendrop dramatically following manufacturer entry and competition, and do not return to theirprevious levels when comparing January 2014 - July 2014 to the prior year.

Manufacturer Entry and Retailer Competition

In the third main result of the paper, we show that ice manufacturer entry also triggers in-creased competition by ice retailers, leading to sharp increases in the provision of trade creditand more switching by fishermen buyers across retailers. We exploit temporal variation inthe number of ice manufacturers and cross-sectional variation in the number of ice retailers ineach fishing wharf to demonstrate that trade credit only increases where both manufacturersand retailers compete in Goderich. The separate comparison of credit provision by retailersin Goderich to both Aberdeen and Tombo allows us to rule out alternative explanationssuch as a general increase in the availability of credit due to new manufacturers, which isalso discounted in interviews with retailers. Instead, retailers emphasize the breakdown ofa collusive agreement in which both promised not to steal customers, as deviation becamemore tempting when the size of the market expanded and ice became readily available fol-lowing new manufacturer entry. Furthermore, with priority delivery relationships no longer

26As Appendix Figures A1.1, A1.2, and A1.3 demonstrate, this is particularly helpful for understandingthe falling prices in Tombo and Goderich during periods in which local price competition was limited.


present as an incentive for customer loyalty, retailers expand credit provision in the hopesof providing a new friction to reduce the incidence of buyer switching.

Figure 1.6 documents the onset of buyer-retailer switching in Goderich following the entryof Manufacturer 3 in late-October 2013. There exist only a handful of switches documentedprior to new manufacturer entry in Goderich involving approximately five fishermen withoutexpressed retailer loyalties, but more than 200 switches involving about 80% of the 86 fisher-men customers in Goderich afterwards. The cumulative probability of switches rises sharplyafter manufacturer entry, and only begins to level off as the end of data collection approachesin June 2014. The corresponding density of switches shows large spikes of switching in early2014 and a gradual decline toward the end of data collection in mid-2014.27

Figure 1.7 shows the total credit line extended each week by all five original retailersserving the original monopolistic supplier to their fishermen customers.28 Note that theRetailers 3, 4, and 5 in Goderich (shown with dotted lines) provide comparable aggregatelevels of trade credit to Retailer 2 in Aberdeen, while Retailer 1 in Tombo provides theconsistently lowest level of credit. Following manufacturer entry in Goderich, we observea dramatic increase in credit provision by Retailers 3 and 4, corresponding roughly to thetime in which buyer switching began to decline. This is consistent with self-reported retailerstrategies to use increased credit as an incentive for buyer loyalty.

In Table 1.5, we estimate the effect of manufacturer competition on credit provision ineach of the three wharves. The main prediction of our theoretical model is that retailer creditshould only expand in Goderich, where more than one retailer competes for customers, andnot in Tombo and Aberdeen, where retailers serve as monopolists. As above in the results onprice competition, our data includes all orders from fishermen purchasing from the originalfive retailers, including purchases recorded by these retailers from competitor manufacturers.Again as above, our results are robust to restricting attention only to orders sourced fromIIB, providing additional confidence that the results are not driven by increasing availabilityof credit due to new manufacturer entry. After collapsing our data to the fisherman-weeklevel, we estimate the following regression equation, where an observation is the aggregateorders by a fisherman f in wharf w in week t.

Credit (%)fwt = θ1Manufacturers (> 1)wt ∗Goderich (= 1)w+

θ2Manufacturers (> 1)wt ∗ Aberdeen (= 1)w+

θ3Manufacturers (> 1)wt + φf + τm + εfwt

(1.12)

We include fishermen fixed effects, φf and year by month fixed effects, τm to controlfor the effects of time-invariant customer characteristics as well as arbitrary time trends.As the theoretical model is premised on a breakdown of collusive relationships following

27In Appendix Table A1.2, we document the average likelihood of switching retailers in Goderich isapproximately 10% per week following manufacturer entry. However, this estimate is only a lower boundgiven that we do not have data on the identities of fishermen customers purchasing from other retailers whoentered Goderich after the onset of manufacturer competition.

28We convert credit figures to US dollars with the average exchange rate of 4200 Leones during this period.


manufacturer competition, we include a dummy variable Manufacturers (> 1)wt that equals1 whenever more than one manufacturer is present in a market rather than the number ofmanufacturers.29 In columns (1) and (2) of Table 1.5, we observe a positive but statisticallyinsignificant effect of manufacturer competition on credit provision. In columns (3) and(4), we observe that Aberdeen and Goderich are associated with between 12-14 percentagepoints greater trade credit provision than Tombo. And in columns (5) and (6) we testSpecification 1.12 from above, finding a large, positive and significant increase in creditprovision in Goderich following manufacturer competition, but no corresponding significantchange in either Aberdeen or Tombo. The magnitude of this effect is 19 percentage pointswhen using the entire time period of data, or 29 percentage points when restricting to theyear on year comparison between January-June 2013 and January-June 2014. In AppendixTables A1.3 and A1.4, we replicate this finding using both number of orders and amount ofcredit provided as the dependent variable; the credit increase is equivalent in magnitude toabout one-third of an average order or about 18 dollars per fisherman-week.

We find no evidence to support potential concerns that manufacturer entry or otherunobserved factors may be directly affecting retailers credit provision decision outside thecompetition for buyer relationships, for example through expanded credit lines. First, wenote in the data that almost no orders from the original five retailers for manufacturers otherthan IIB are provided on credit. In interviews with retailers and manufacturers, we establishthat each new manufacturer established its own policy on credit lines to retailers, but thatall were more restrictive than IIB and that retailers generally did not find them attractive.Second, we confirm that our results in Table 1.5 are not affected when restricting attentiononly to IIB orders. And third, we find no evidence in our interviews of changing availabilityof alternate, non-manufacturer sources of credit available to retailers in Goderich, addingfurther confidence to the comparisons provided by Aberdeen and Tombo.

Finally, we acknowledge that there may be concerns about our approach to inferenceif there is a high degree of intra-cluster correlation in outcomes within each wharf. Givenour small number of wharves, we implement the wild cluster bootstrap inference correctionrecommended by Cameron et al. (2008), Cameron and Miller (2013) and Webb (2013) tocluster at the wharf level. Specifically, we first impose the null hypothesis when estimatingthe model in columns (5) and (6) of Table 1.5, then randomly select from Webb’s six-pointdistribution of weights to generate new pseudo residuals over 999 iterations, and finallycalculate a counter-factual distribution of Wald statistics under the null hypothesis.30 Bycomparing the original sample Wald statistics to this distribution, we are able to obtaina p-value (but not a standard error) for testing the significance of our main result underthe more strict requirement of clustering at the fishing wharf level. As we show in Figures1.8 and 1.9, the wild cluster bootstrap p-values corresponding to coefficient of interest incolumns (5) and (6) of Table 1.5 are .144 and .188, respectively. While these figures are

29We observe similar results when using a dummy variable for the onset of manufacturer competition ina wharf, suggesting our results are unaffected by the subsequent exit of manufacturers in Aberdeen wharf.

30Following the alternative enumeration approach in Webb (2013) provides nearly identical results.


short of conventional levels of statistical significance, we note that the wild cluster bootstrapis a very conservative correction and that the p-values are in a range that suggests our resultsare unlikely to be due to random chance.

1.7 Conclusion

Ongoing relationships allow suppliers and buyers to overcome transactional hazards. Whileprevious literature has focused on how relationships affect contractual and market outcomes,our analysis highlights the effect of upstream market structure on the value of downstreamrelationships. In our setting of ice sales in Sierra Leone, we find that that a shock thatincreases entry by upstream manufacturers leads not only to lower prices and fewer latedeliveries, but also to increased competition by downstream retailers to secure buyers’ loyalty.The onset of manufacturing competition expands the downstream market size, disrupting acollusive arrangement used by retailers to limit business stealing. During the equilibriumtransition, buyers switch retailers multiple times in search of better credit provision terms,and an informal institution that rewarded loyalty with fewer late deliveries is replaced withan expansion in trade credit relationships. Thus, we observe a transition in the distributionof trade surplus that rewards the “long side of the market” (MacLeod and Malcomson, 1989),as benefits from trade move from the monopolistic manufacturer to the downstream retailersand buyers after the onset of competition. In particular, we observe that while averageprices and lateness decrease everywhere following new manufacturer entry, credit provisiononly increases in the one wharf, Goderich, where multiple existing retailers also compete.

Our findings have broad relevance to those interested in promoting growth and enhanc-ing social welfare in emerging economies. In this work, we demonstrate that the costs ofupstream market power exceed the well-known effects on price and quantity to include fric-tions on the formation of downstream relationships and the provision of services such as tradecredit. We demonstrate sizable improvements in contractual terms following competition,suggesting large externalities from market power that may serve to constrain firm growth.We also find that market power matters at each level of the supply chain, with retailercompetition appearing just as important as manufacturer competition in explaining creditprovision increases. Given the high concentration of market power in upstream industriesin low-income countries, our results suggest that policymakers should pay close attentionto promoting entry and competition. And for the firm strategy audience, we provide directevidence of the institutional sources that make relationships important in thin markets. Ourfindings indicate that managers in emerging economies where contract enforcement is limitedmust account for the value of downstream relationships in forming their business strategies,particularly where future competition for relationships is likely.


1.8 Tables and Figures

Figure 1.1: Map of Freetown Peninsula, Sierra Leone

Notes: Map shows the approximate location of three major fishingwharves served by the ice factories: Tombo (W1), Aberdeen (W2)and Goderich (W3). It also shows the factory locations of the incum-bent manufacturer (M1), and the subsequent competitor manufactur-ers (M2, M3, M4, M5), which are numbered in the order of entry andhighlighted in red text. See paper text for more details.


Figure 1.2: Entry and Price Competition, All Wharves

Notes: Y-axis shows the incumbent manufacturer’s retail price per 30 kilogram bag of ice sold to fishermenin all three wharves (Tombo, Aberdeen and Goderich). Vertical lines mark the first date of ice sales by acompetitor manufacturer in one or more wharves. The dark vertical lines corresponding to price competitionare associated with the simultaneous entry of Manufacturer 3 into Goderich and Aberdeen wharves, and thelater entry of Manufacturer 5 into the Goderich wharf. The light vertical lines are associated with non-pricecompetition entry by Manufacturer 2 into the Tombo wharf and Manufacturer 4 into Aberdeen wharf. Seepaper text for more details.


Figure 1.3: Market Structure Pre- and Post-Competition, All Wharves

MANUFACTURER 1

RETAILER 1

M1 M2

R1

Tombo WharfPre-Competition

(A)Tombo Wharf

Post-Competition

MANUFACTURER 1

RETAILER 2

M1 M4 M3

R6R2R7

Aberdeen WharfPre-Competition

(B)Abderdeen WharfPost-Competition

MANUFACTURER 1

R4R3 R5

M1 M4 M5 M3

R3 R4 R8 R5

Goderich WharfPre-Competition

(C)Goderich WharfPost-Competition

Notes: This figure presents the market structure of manufacturers and retailers in eachof the three wharves prior to the onset of new manufacturer entry and then again inJuly 2014. Both before and after manufacturer entry, Tombo and Aberdeen wharveshave monopolist retailers, while Goderich has multiple competing retailers. Dottedboxes signify manufacturer or retailer exit from market as of July 2014.


Figure 1.4: IIB Production Levels and Lateness

Notes: This figure presents the aggregate production levels and share of orders deliv-ered late to fishermen by the incumbent manufacturer, IIB. The vertical dashed linemarks the first date of ice sales by Manufacturer 3 in Goderich and Aberdeen wharves,corresponding to the start of price competition. See paper text for more details.

Figure 1.5: IIB Sources of Ice Demand (kg)

Notes: This figure presents the aggregate ice sales in kilograms by the incumbentmanufacturer, IIB, to two sources of demand: fishermen and non-fishermen sources.


Figure 1.6: Buyer-Retailer Switching, Goderich Only

Notes: This figure presents the cumulative probability and density of buyer-retailerswitches observed in Goderich wharf during the data collection period. The verticaldashed line marks the first entry by a new manufacturer in Goderich and Aberdeen.

Figure 1.7: Credit Provision, All Wharves

Notes: This figure presents the aggregate credit provision in US dollars provided byretailers in each of the three wharves during the data collection period. The verticaldashed line marks the first entry by a new manufacturer in Goderich and Aberdeen.


Figure 1.8: Wild cluster bootstrap results for Table 1.5 column (5)

Notes: This figure presents the distribution of Wald test statistic values calculatedusing the Wild cluster bootstrap method for Table 1.5 column (5). The vertical linemarks the value of the original sample Wald statistic. See paper text for more details.

Figure 1.9: Wild cluster bootstrap results for Table 1.5 column (6)

Notes: This figure presents the distribution of Wald test statistic values calculatedusing the Wild cluster bootstrap method for Table 1.5 column (6). The vertical linemarks the value of the original sample Wald statistic. See paper text for more details.


Table 1.1: Summary Statistics - Survey Data

Variable Mean Std. Dev. Min Max N

Age 40.28 9.44 24 82 143Head of Household 0.97 0.18 0 1 148Years Fishing Experience 17.29 8.77 1 47 150Own Fishing Boat 0.61 0.49 0 1 150Captain of Fishing Boat 0.83 0.38 0 1 150Number of Fishing Trips Per Month 9.63 10.71 2 86 149Years Known Ice Agent 9.07 7.13 1 40 142Years Bought from Ice Agent 4.44 3.35 1 20 143

Planned Days Per Trip 2.93 1.24 1 7 4565No Ice Purchase for Trip (=1) 0.24 0.43 0 1 4565Ice Purchase (kg) 460.48 353.39 0 1650 4565Trip Gross Margins (dollars) 110.6 152.07 -345 1748 4492

Notes: Demographic and fishing experience data from baseline survey of fishermen above theseparating line, with fishing experience data from high frequency survey below the separatingline. See paper text for more details.

CHAPTER 1. COMPETING FOR RELATIONSHIPS IN SIERRA LEONE 34T

able

1.2:

Outs

ide

Dem

and

and

Lat

enes

s

Shar

eL

ate

Ord

ers

Shar

eL

ate

Ord

ers

(1)

(2)

(3)

(4)

Log

Non

-Fis

hin

gSal

es(k

g)0.

052*

0.10

8*(0

.030

)(0

.057

)M

anufa

cture

rE

ntr

y(=

1)-0

.246

**-0

.241

**(0

.114

)(0

.113

)

Tim

eP

erio

dP

re-E

ntr

yJan

-Jun

2013

All

Jan

-Jun

2013

&20

14M

ean

Dep

Var

0.19

0.26

0.12

0.15

#W

eeks

4426

7852

#O

bse

rvat

ions

615

441

1155

819

R-S

quar

ed0.

450.

400.

480.

48W

eath

erC

ontr

ols

YE

SY

ES

YE

SY

ES

Whar

fF

E-

--

-C

alen

dar

Mon

thF

E-

--

-W

har

fby

Mon

thF

EY

ES

YE

SY

ES

YE

S

Notes:

Dep

end

ent

vari

able

isth

esh

are

of

fish

erm

enord

ers

del

iver

edla

tein

agiv

enw

harf

,an

dan

ob

serv

ati

on

isa

fish

ing

wh

arf-

day

.D

eliv

erie

sd

ata

incl

ud

esp

urc

hase

sfr

om

the

ori

gin

alfi

vere

tail

ers

serv

ing

the

incu

mb

ent

man

ufa

ctu

rer,

and

does

incl

ud

esa

les

by

thes

ere

tail

ers

on

beh

alf

of

oth

erm

anu

fact

ure

r.L

og

Non

-Fis

hin

gS

ales

(kg)

isth

en

atu

ral

loga

rith

mof

the

kil

ogra

mto

tal

of

all

non

-fish

ing

ice

sale

sm

ad

eby

the

incu

mb

ent

sup

pli

eron

that

day

.M

anu

fact

ure

rE

ntr

y(=

1)

isa

du

mm

yva

riab

leth

at

equ

als

one

foll

owin

gth

een

try

ofth

efi

rst

com

pet

itor

man

ufa

ctu

rer

into

that

wh

arf

(see

pap

erte

xt

for

det

ail

s).

Wea

ther

contr

ols

incl

ud

eav

erag

edai

lyte

mp

erat

ure

,h

ours

of

rain

,an

dav

erage

win

dsp

eed

.T

ime

per

iod

cove

rsth

ep

erio

dp

rior

toth

een

try

ofth

efirs

tco

mp

etit

orm

anu

fact

ure

rw

hen

note

das

“P

re-E

ntr

y,”

an

dco

vers

Janu

ary

2013

toJu

ne

2014

wh

enn

oted

as“A

ll.”

Reg

ress

ion

sin

clu

de

wea

ther

contr

ols

,w

harf

fixed

effec

ts,

cale

nd

ar

month

fixed

effec

tsan

dw

har

fby

cale

nd

arm

onth

fixed

effec

tsas

note

d.

Rob

ust

stan

dard

erro

rs,

clu

ster

edat

wee

kle

vel,

inp

aren

thes

es.

***

p<

0.01

,**

p<

0.0

5,

*p<

0.1

.


Table 1.3: Prioritized Deliveries (Goderich Only)

Priority (=1) Priority (=1) Priority (=1)(1) (2) (3) (4) (5) (6)

Loyal Client (=1) 0.31*** 0.25** 0.31*** 0.24** 0.31*** 0.24**(0.09) (0.10) (0.10) (0.10) (0.10) (0.10)

Order Rank (#) -0.08*** -0.10*** -0.11***(0.02) (0.02) (0.02)

Order Size (kg) -0.00 -0.00 0.00(0.00) (0.00) (0.00)

Paid In Full (=1) -0.06 -0.09 -0.07(0.11) (0.12) (0.13)

Wharf Sample Goderich Goderich GoderichTime Period Jan - Jun 2013 Jan - Jun 2013 Jan - Jun 2013Mean Dep Var 0.52 0.52 0.52 0.52 0.52 0.52# Observations 191 191 191 191 191 191R-Squared 0.04 0.13 0.05 0.17 0.10 0.25Agent FE NO NO YES YES YES YESWeek FE NO NO NO NO YES YES

Notes: Dependent variable is a dummy variable for whether a fisherman’s order was prioritizedfor ontime delivery on a day in which late deliveries were made at his wharf. Sample only includesGoderich wharf and is limited to days with late deliveries. Time period covers from January 2013to June 2013, prior to the entry of the first competitor manufacturer into Goderich wharf. LoyalClient (=1) is a dummy variable that equals one if the retailer reported that this fisherman wouldonly buy from him even if other ice supply was available. Order Rank measures the sequence inwhich a fisherman’s order were recorded on a given day for a given retailer. Order Size (kg) isthe quantity of ice demanded by the fishermen. Paid In Full (=1) is a dummy that equals oneif a fisherman paid upfront for his entire order. Regressions include agent fixed effects and weekfixed effects as noted. Robust standard errors, clustered at the retailer-fishermen pair level, inparentheses. *** p<0.01, ** p<0.05, * p<0.1.


able

1.4:

Ret

ail

Pri

ceC

omp

etit

ion

and

Spillo

vers

Log

Pri

ceL

ogP

rice

Log

Pri

ceL

ogP

rice

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

#M

anu

fact

ure

rs-0

.054

***

-0.0

57**

*-0

.031

***

-0.0

28**

*-0

.043

***

-0.0

59**

*-0

.011

**-0

.005

(0.0

04)

(0.0

04)

(0.0

02)

(0.0

01)

(0.0

06)

(0.0

09)

(0.0

05)

(0.0

11)

#O

uts

ide

Man

ufa

ctu

rers

-0.0

29**

*-0

.032

***

-0.0

30**

*-0

.035

***

(0.0

02)

(0.0

03)

(0.0

02)

(0.0

03)

#M

anu

fact

ure

rs*

Ab

erd

een

(=1)

-0.0

19-0

.008

-0.0

26**

-0.0

27*

(0.0

12)

(0.0

18)

(0.0

10)

(0.0

15)

#M

anu

fact

ure

rs*

God

eric

h(=

1)-0

.010

0.00

7-0

.021

***

-0.0

25**

(0.0

08)

(0.0

10)

(0.0

06)

(0.0

10)

Tim

eP

erio

dA

llJan

-Ju

ne

All

Jan

-Ju

ne

All

Jan

-Ju

ne

All

Jan

-Ju

ne

Mea

nD

epV

ar9.

779.

769.

779.

769.

779.

769.

779.

76#

Wee

ks

7851

7851

7851

7851

#O

bse

rvat

ion

s21

914

321

914

321

914

321

914

3R

-Squ

ared

0.53

0.56

0.73

0.75

0.54

0.57

0.74

0.76

Wh

arf

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

enden

tva

riab

leis

the

log

aver

age

pri

cepai

dfo

ra

30k

gbag

ofic

ein

aw

harf

,and

an

obse

rvati

on

isa

whar

f-w

eek.

#M

anufa

cture

rsis

the

count

of

ice

man

ufa

cture

rsse

rvin

ga

whar

f,#

Outs

ide

Man

ufa

cture

rsis

the

count

of

ice

man

ufa

cture

rsact

ive

only

inot

her

whar

ves

,A

ber

dee

n(=

1)is

adum

my

vari

able

that

equal

son

efo

rA

ber

dee

nw

har

f,and

Goder

ich

(=1)

isa

dum

my

vari

able

that

equal

son

efo

rG

oder

ich

whar

f.P

rice

dat

ain

cludes

purc

has

esfr

om

the

orig

inal

five

reta

iler

sse

rvin

gth

ein

cum

ben

tm

anufa

cture

r,and

does

incl

ude

sale

sby

thes

ere

tailer

son

beh

alf

ofot

her

man

ufa

cture

rs.

Tim

ep

erio

dco

vers

Jan

uar

y201

3to

June

2014

when

not

edas

“A

ll,”

and

cove

rsJanuary

2013-

June

2013

and

January

2014-

June

2014

when

note

das

“Jan

-Jun.”

Rob

ust

standar

der

rors

,cl

ust

ered

atth

ew

eek

leve

l,in

par

enth

eses

.***

p<

0.0

1,

**p<

0.05,

*p<

0.1

.


able

1.5:

Ret

aile

rC

redit

Pro

vis

ion

Cre

dit

Pro

vis

ion

Cre

dit

Pro

vis

ion

Cre

dit

Pro

vis

ion

(1)

(2)

(3)

(4)

(5)

(6)

Man

ufa

cture

rs(>

1)0.

050.

07-0

.02

-0.0

5(0

.04)

(0.0

6)(0

.03)

(0.0

5)A

ber

dee

n(=

1)0.

12*

0.13

0.09

0.04

(0.0

7)(0

.09)

(0.0

8)(0

.10)

Goder

ich

(=1)

0.14

*0.

140.

03-0

.03

(0.0

8)(0

.11)

(0.0

9)(0

.12)

Man

ufa

cture

rs(>

1)*

Ab

erdee

n(=

1)-0

.03

-0.0

2(0

.03)

(0.0

4)M

anufa

cture

rs(>

1)*

Goder

ich

(=1)

0.19

***

0.29

***

(0.0

5)(0

.05)

Tim

eP

erio

dA

llJan

-June

All

Jan

-June

All

Jan

-June

Mea

nD

epV

ar0.

110.

120.

110.

120.

110.

12#

Wee

ks

7851

7851

7851

#O

bse

rvat

ions

3977

2935

3977

2935

3977

2935

R-S

quar

ed0.

130.

150.

130.

150.

150.

19C

alen

dar

Mon

thF

EY

ES

YE

SY

ES

YE

SY

ES

YE

SW

har

fF

E-

--

--

-F

isher

men

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

end

ent

vari

able

isth

efr

acti

on

of

fish

erm

an

’sto

tal

pay

men

tto

reta

iler

mad

eon

cred

it,

an

dan

ob

serv

ati

on

isa

fish

erm

an-w

eek.

Man

ufa

ctu

rers

(>1)

isa

du

mm

yva

riab

leth

at

equ

als

on

eif

more

than

on

em

anu

fact

ure

ris

makin

gsa

les

ina

fish

erm

an’s

wharf

,A

ber

dee

n(=

1)

isa

du

mm

yva

riab

leth

at

equ

als

on

efo

rA

ber

dee

nw

harf

,an

dG

od

eric

h(=

1)is

ad

um

my

vari

able

that

equ

als

on

efo

rG

od

eric

hw

harf

.C

red

itd

ata

incl

ud

esp

urc

hase

sfr

om

the

ori

gin

al

five

reta

iler

sse

rvin

gth

ein

cum

ben

tm

anu

fact

ure

r,an

dd

oes

incl

ud

esa

les

by

thes

ere

tail

ers

on

beh

alf

of

oth

erm

anu

fact

ure

rs.

Tim

ep

erio

dco

ver

sJan

uar

y20

13to

Ju

ne

2014

wh

enn

ote

das

“A

ll,”

an

dco

vers

Janu

ary

2013-J

un

e2013

an

dJanu

ary

2014

-Ju

ne

2014

wh

enn

oted

as“J

an-J

un

.”R

egre

ssio

ns

incl

ud

eca

len

dar

month

fixed

effec

ts,

wh

arf

fixed

effec

ts,

an

dfi

sher

men

fixed

effec

tsas

not

ed.

Rob

ust

stan

dard

erro

rs,

clu

ster

edat

wee

kle

vel

,in

pare

nth

eses

.***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1.


1.9 Chapter Appendices

Appendix Tables and Figures

Figure A1.1: Entry and Price Competition, Tombo Wharf

Notes: Y-axis shows the manufacturer’s retail price per 30 kilogram bag of ice sold to fishermen in Tombowharf. Vertical lines mark the first date of ice sales by a competitor manufacturer in one or more wharves.See paper text for more details.


Figure A1.2: Entry and Price Competition, Aberdeen Wharf

Notes: Y-axis shows the manufacturer’s retail price per 30 kilogram bag of ice sold to fishermen in Aberdeenwharf. Vertical lines mark the first date of ice sales by a competitor manufacturer in one or more wharves.See paper text for more details.


Figure A1.3: Entry and Price Competition, Goderich Wharf

Notes: Y-axis shows the manufacturer’s retail price per 30 kilogram bag of ice sold to fishermen in Goderichwharf. Vertical lines mark the first date of ice sales by a competitor manufacturer in one or more wharves.See paper text for more details.


able

A1.

1:R

etai

lP

rice

Com

pet

itio

nan

dSpillo

vers

-W

har

fB

reak

dow

n

Log

Pri

ceL

ogP

rice

Log

Pri

ce(1

)(2

)(3

)(4

)(5

)(6

)

#M

anufa

cture

rs-0

.043

***

-0.0

13**

-0.0

62**

*-0

.020

**-0

.053

***

-0.0

37**

*(0

.006

)(0

.005

)(0

.010

)(0

.008

)(0

.002

)(0

.002

)#

Outs

ide

Man

ufa

cture

rs-0

.028

***

-0.0

53**

*-0

.023

***

(0.0

08)

(0.0

10)

(0.0

02)

Whar

fSam

ple

Tom

bo

Ab

erdee

nG

oder

ich

Tim

eP

erio

dA

llA

llA

llA

llA

llA

llM

ean

Dep

Var

9.77

9.77

9.76

9.76

9.76

9.76

#W

eeks

6666

7575

7878

R-S

quar

ed0.

360.

560.

480.

740.

700.

91

Notes:

Dep

end

ent

vari

able

isth

elo

gav

erage

pri

cep

aid

for

a30kg

bag

of

ice

ina

wh

arf,

an

dan

ob

serv

ati

on

isa

wh

arf-

wee

k.

#M

anu

fact

ure

rsis

the

cou

nt

of

ice

manu

fact

ure

rsse

rvin

ga

wh

arf

,an

d#

Ou

tsid

eM

anu

fact

ure

rsis

the

cou

nt

ofic

em

anu

fact

ure

rsac

tive

ony

inoth

erw

harv

es.

Pri

ced

ata

incl

ud

espu

rch

ase

sfr

om

the

ori

gin

al

five

reta

iler

sse

rvin

gth

ein

cum

ben

tm

anu

fact

ure

r,and

does

incl

ud

esa

les

by

thes

ere

tail

ers

on

beh

alf

of

oth

erm

anu

fact

ure

rs.

Tim

ep

erio

dco

vers

Jan

uar

y20

13to

Ju

ne

2014

wh

enn

ote

das

“A

ll,”

an

dco

vers

Janu

ary

2013-J

une

2013

an

dJanu

ary

2014-

Ju

ne

2014

wh

enn

oted

as“J

an-J

un

.”R

ob

ust

stan

dard

erro

rs,cl

ust

ered

at

the

wee

kle

vel,

inp

are

nth

eses

.***

p<

0.0

1,

**p<

0.05

,*

p<

0.1.


Table A1.2: Buyer-Retailer Switching

Switch Retailer (=1) Switch Retailer (=1)(1) (2) (3) (4)

Manufacturer Entry (=1) 0.03*** 0.05*** 0.10*** 0.10***(0.01) (0.01) (0.04) (0.04)

Wharf Sample All All Goderich GoderichTime Period All Jan-Jun All Jan-JunMean Dep Var 0.02 0.03 0.04 0.04# Fishermen 154 154 86 86# Weeks 78 51 78 51# Observations 10192 6348 5538 3447R-Squared 0.02 0.02 0.04 0.03Calendar Month FE YES YES YES YESFishermen FE YES YES YES YES

Notes: Dependent variable is a dummy variable for whether a fishermen switched to anew retailer for ice purchase relative to last period, and an observation is a fisherman-week. Data includes purchases from the original five retailers serving the incumbentmanufacturer, and does include sales by these retailers on behalf of other manufacturers.Manufacturer Entry (=1) is a dummy variable that equals one following the entry of thefirst competitor manufacturer into that wharf (see paper text for details). Time periodcovers January 2013 to June 2014 when noted as “All,” and covers January 2013-June2013 and January 2014-June 2014 when noted as “Jan-June.” All regressions includecalendar month fixed effects and fishermen fixed effects. Robust standard errors, two-way clustered at fishermen and week level, in parentheses. *** p<0.01, ** p<0.05, *p<0.1.


able

A1.

3:R

etai

ler

Cre

dit

Pro

vis

ion

Cre

dit

Ord

ers

(#)

Cre

dit

Ord

ers

(#)

Cre

dit

Ord

ers

(#)

(1)

(2)

(3)

(4)

(5)

(6)

Man

ufa

cture

rs(>

1)0.

070.

090.

01-0

.01

(0.0

5)(0

.08)

(0.0

4)(0

.06)

Ab

erdee

n(=

1)0.

23**

0.23

0.23

*0.

17(0

.11)

(0.1

5)(0

.12)

(0.1

6)G

oder

ich

(=1)

0.17

0.14

0.08

-0.0

3(0

.11)

(0.1

5)(0

.12)

(0.1

6)M

anufa

cture

rs(>

1)*

Ab

erdee

n(=

1)-0

.09*

-0.0

8(0

.05)

(0.0

6)M

anufa

cture

rs(>

1)*

Goder

ich

(=1)

0.18

***

0.30

***

(0.0

7)(0

.07)

Tim

eP

erio

dA

llJan

-June

All

Jan

-June

All

Jan

-June

Mea

nD

epV

ar0.

190.

190.

190.

190.

190.

19#

Fis

her

men

#W

eeks

#O

bse

rvat

ions

3977

2935

3977

2935

3977

2935

R-S

quar

ed0.

110.

130.

100.

120.

120.

15C

alen

dar

Mon

thF

EY

ES

YE

SY

ES

YE

SY

ES

YE

SW

har

fF

E-

--

--

-F

isher

men

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

end

ent

vari

able

isth

enu

mb

erof

fish

erm

enp

aym

ents

tore

tail

erm

ad

eon

cred

it,

an

dan

ob

serv

ati

on

isa

fish

erm

an-w

eek.

Man

ufa

ctu

rers

/Ret

aile

rs(>

1)

isa

du

mm

yva

riab

leth

at

equ

als

on

eif

more

than

on

em

anu

fact

ure

r/re

tail

eris

mak

ing

sale

sin

afi

sher

man

’sw

har

f.C

red

itd

ata

incl

ud

esp

urc

hase

sfr

om

the

ori

gin

alfi

vere

tail

ers

serv

ing

the

incu

mb

ent

man

ufa

ctu

rer,

and

does

incl

ud

esa

les

by

thes

ere

tail

ers

on

beh

alf

of

oth

erm

anu

fact

ure

rs.

Tim

ep

erio

dco

ver

sJanu

ary

2013

toJu

ne

2014

wh

enn

oted

as“A

ll,”

an

dco

vers

Janu

ary

2013-J

un

e2013

an

dJanu

ary

2014-J

un

e2014

wh

enn

ote

das

“Jan

-Ju

n.”

Reg

ress

ion

sin

clud

eca

len

dar

month

fixed

effec

ts,

wh

arf

fixed

effec

ts,

an

dfi

sher

men

fixed

effec

tsas

note

d.

Rob

ust

stan

dar

der

rors

,tw

o-w

aycl

ust

ered

at

fish

erm

enan

dw

eek

leve

l,in

pare

nth

eses

.***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

.


able

A1.

4:R

etai

ler

Cre

dit

Pro

vis

ion

Tot

alC

redit

($)

Tot

alC

redit

($)

Tot

alC

redit

($)

(1)

(2)

(3)

(4)

(5)

(6)

Man

ufa

cture

rs(>

1)5.

549.

93*

1.89

3.10

(3.3

4)(5

.73)

(2.7

5)(4

.44)

Ab

erdee

n(=

1)19

.29*

*20

.99*

*19

.52*

*16

.58

(8.1

2)(9

.93)

(8.6

1)(1

0.65

)G

oder

ich

(=1)

8.38

7.60

2.84

-1.5

6(5

.51)

(6.6

9)(6

.32)

(8.1

4)M

anufa

cture

rs(>

1)*

Ab

erdee

n(=

1)-6

.18

-3.1

2(4

.02)

(4.7

0)M

anufa

cture

rs(>

1)*

Goder

ich

(=1)

11.9

1***

18.0

9***

(4.2

8)(4

.69)

Tim

eP

erio

dA

llJan

-June

All

Jan

-June

All

Jan

-June

Mea

nD

epV

ar10

.20

11.0

510

.20

11.0

510

.20

11.0

5#

Fis

her

men

#W

eeks

#O

bse

rvat

ions

3977

2935

3977

2935

3977

2935

R-S

quar

ed0.

090.

100.

090.

100.

100.

12C

alen

dar

Mon

thF

EY

ES

YE

SY

ES

YE

SY

ES

YE

SW

har

fF

E-

--

--

-F

isher

men

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

end

ent

vari

able

isth

eto

tal

am

ou

nt

of

fish

erm

enp

aym

ents

tore

tail

erm

ad

eon

cred

itin

US

doll

ars

,an

dan

obse

rvat

ion

isa

fish

erm

an-w

eek.

Man

ufa

ctu

rers

/R

etail

ers

(>1)

isa

du

mm

yva

riab

leth

at

equ

als

on

eif

more

than

on

em

anu

fact

ure

r/re

tail

eris

mak

ing

sale

sin

afi

sher

man

’sw

harf

.C

red

itd

ata

incl

ud

esp

urc

hase

sfr

om

the

ori

gin

al

five

reta

iler

sse

rvin

gth

ein

cum

ben

tm

anufa

ctu

rer,

an

dd

oes

incl

ud

esa

les

by

thes

ere

tail

ers

on

beh

alf

of

oth

erm

anu

fact

ure

rs.

Tim

ep

erio

dco

vers

Jan

uar

y20

13to

Ju

ne

2014

wh

enn

ote

das

“A

ll,”

an

dco

vers

Janu

ary

2013-J

un

e2013

an

dJanu

ary

2014-J

un

e20

14w

hen

not

edas

“Jan

-Ju

n.”

Reg

ress

ion

sin

clu

de

cale

nd

ar

month

fixed

effec

ts,

wh

arf

fixed

effec

ts,

an

dfish

erm

enfi

xed

effec

tsas

not

ed.

Rob

ust

stan

dar

der

rors

,tw

o-w

aycl

ust

ered

at

fish

erm

enan

dw

eek

leve

l,in

pare

nth

eses

.***

p<

0.0

1,

**

p<

0.05

,*

p<

0.1.

45

Chapter 2

Violence and Financial Decisions inAfghanistan

2.1 Abstract

We examine the relationship between violence and financial decisions in Afghanistan. Usingthree separate data sources, we find that individuals experiencing violence retain more cashand are less likely to adopt and use mobile money, a new financial technology. We first com-bine detailed information on the entire universe of mobile money transactions in Afghanistanwith administrative records for all violent incidents recorded by international forces, and finda negative relationship between violence and mobile money use. Second, in the context of arandomized control trial, violence is associated with decreased mobile money use and greatercash balances. Third, in financial survey data from nineteen of Afghanistan’s 34 provinces,we find that individuals experiencing violence hold more cash. Collectively, the evidenceindicates that individuals experiencing violence prefer cash to mobile money. More specu-latively, it appears that this is principally because of concerns about future violence. Thedegree of the relationship between cash holdings and violence is large enough to suggest thatrobust formal financial networks face severe challenges developing in conflict environments.1

1The material in this chapter is based on joint work with Michael Callen and Joshua Blumenstock. SeeBlumenstock et al. (2014a).

CHAPTER 2. VIOLENCE AND FINANCIAL DECISIONS IN AFGHANISTAN 46

2.2 Introduction

Many poor countries are plagued by violence.2 While a substantial literature documentsthe positive relationship between poverty and conflict, economists have only recently begunto explore the micro-economic mechanisms linking violence to economic stagnation and lowlevels of income.3 Conflict destroys capital (Davis and Weinstein, 2002; Miguel and Roland,2011), deters investment (Besley and Mueller, 2012; Singh, 2013), changes economic decision-making (Voors et al., 2012; Callen et al., 2014), and introduces additional uncertainty aboutthe future. Due in part to a lack of reliable data, the impact of violence on financial decisionsis not well understood.

This paper studies the relationship between violence and financial decision-making. Wepursue this using data for the universe of mobile money transactions, precise approximationsof mobile money users physical locations based on their geo-tagged call records, adminis-trative records for all violent incidents recorded by international forces, a cross-section offinancial survey data from nineteen of Afghanistan’s 34 provinces, and monthly panel datafrom an experiment that strongly incentivized mobile money adoption. The combination ofsources allow us to examine the financial responses to violence in two separate large popula-tions and with monthly financial survey data in an experimental sample. Collectively, thesedata provide a rare glimpse into financial behavior in several samples that are both affectedby violence and in the midst of adopting a major new financial technology.

The analysis provides three central findings. First, violence is associated with lowermobile money balances and fewer transactions both for subjects in our experiment and formobile money users in general. Related to this, our experimental intervention involved payingsubjects their entire salary using mobile money, providing a sharp incentive to use mobilemoney. Treatment subjects that experience violence keep roughly half as much of their salaryas mobile balances compared to treatment subjects unaffected by violence. Second, violenceis associated with an increase in cash holdings that is roughly proportional to the reductionin mobile money balances both for subjects in our study and for subjects in the nineteenprovince sample. Last, respondents who think that future violence is more likely exhibitless mobile money usage and higher cash savings in both the nineteen provinces and in theexperimental sample. These relations remain significant, in the latter sample, using onlywithin-subject variation and controlling for any time-invariant individual-specific factors.4

Our non-experimental results are based on an analysis of the complete history of transac-tions made on the M-Paisa mobile money network from its creation in 2008 until 2013.5 We

2Approximately 1.5 billion individuals (roughly 20% of the world population) live in countries affectedby fragility, violence or conflict (World Bank, 2011).

3Blattman and Miguel (2010) and Mueller (2013) provide excellent reviews of the economic causes andconsequences of civil conflict.

4In results available on request, estimates including province by month fixed effects produce similarresults, at least partly controlling for changes in the objective risk of violence experienced by employees inthe same province.

5In this version of the paper, we limit our analysis to the Dec 2010-April 2012 period where we cancurrently match M-Paisa users to geolocations using calling records.


combine this rich administrative dataset with a geocoded database of tens of thousands ofviolent events in Afghanistan. In our population of regular M-Paisa users, we find that theindividuals who are more exposed to violence are less likely to use the mobile money systemas a storage of value or a means of exchange. This finding persists even when controllingfor unobserved heterogeneity at the individual level: the same individual is less likely to usemobile money in the immediate aftermath of violent events.

To better understand why violence impacts the adoption and use of mobile money, weconduct a field experiment in Afghanistan in which we induce random variation to an in-dividual’s propensity to adopt mobile money. In our experiment, employees of a large,Afghan-staffed firm were randomly assigned to receive their monthly salary payments inmobile money or remain in the status quo cash payment system. To ensure consistencyacross treatment and control groups, all employees received new phones and were enrolledfor accounts on the mobile money platform and trained in how to use the new technology.Despite being able to fully cash out their mobile salary deposits, the treatment group showssignificant increases in usage of their mobile account, but both exposure to violence andexpectations of future violence mute these effects. We show that subjects who believe thatfuture violence is more likely hold lower mobile money balances and keep more cash, evenwhen facing identical objective levels of risk.

Moreover, the panel survey data collected in the experiment provides suggestive evidencefor the mechanism by which violence attenuates an individual’s use of mobile money. Namely,we observe that instead of saving money in their mobile accounts, individuals with greaterexpectations of future violence are more likely to retain cash on hand. This finding iscorroborated by a nationwide household survey data from Afghanistan, in which we observe astrong positive correlation between an individual’s subjective expectations of future violenceand the amount he saves in cash relative to other technologies. We are also able to rule outseveral possible alternative explanations, for instance that our effect is driven by increasesin transaction and travel costs or by potential reductions in agent liquidity.6

Individuals experiencing (and expecting) violence in Afghanistan appear to prefer cashto mobile money. This is in line with observations about the limited development of formalbanking services in the country; only 9% of Afghan adults to hold bank accounts and only3% to save money at a financial institutions (Demirguc-Kunt and Klapper, 2012). Thedevelopment of financial systems requires broad participation and long time horizons fromaccount holders. This is likely to be particularly true for mobile money, which provides aprototypical example of a technology subject to network externalities (Mas and Radcliffe,2011). A range of advocates see in mobile money the opportunity to build a new financialsystem that does not require the brick-and-mortar investment of a bank-based financialsystem (Dermish et al., 2011; Mbiti and Weil, 2011; Suri et al., 2012). But while we documentdemand from Afghan firms for paying employee salaries using mobile money, our resultssuggest that individual users will continue to be reluctant to use the financial technology as

6As we document that individuals decrease their mobile money balance following violent events, it isunlikely that the agents are unable to provide withdrawals specifically due to these violent events.


long as violence is part of their daily lives.Our findings complement a growing body of literature attempting to understand the

economic impacts of mobile phones and other information and communications technologiesin developing countries. Beginnining with work by Jensen (2007) and Aker (2010), themass proliferation of mobile phones has been linked to increased efficiency in agriculturalmarkets. More recent work by Jack and Suri (2014) and Blumenstock et al. (2014b) furtherindicates that mobile money can reduce transaction costs in remittances and help enablemore efficient risk sharing. In work closest to our own, Aker et al. (2011) show that mobilemoney payments can reduce inefficiencies for both the payer and payee. Our focus, however,is different. While we find complementary evidence that mobile money salary paymentscreate efficiencies for the employer, we find that the benefits to employees are not uniform.In particular, our analysis of the detailed mobile money transaction records allows us toexamine how different types of individuals, and in particular those exposed to violence, usethe technology differently from the average subscriber.

The remainder of the paper is structured as follows. The next section reviews the settingand provides institutional details. Section 3 provides initial evidence on the relationshipbetween violence and mobile money adoption from two large administrative datasets fromAfghanistan during 2010-2012. Section 4 presents further evidence from the randomizedexperiment conducted in Afghanistan during 2012-2013. Section 5 examines underlyingmechanisms, and Section 6 concludes.

2.3 Background

Violence in Afghanistan

Afghanistan is one of the world’s poorest and most-conflict affected countries. Beginningwith a communist coup in 1978 and the Soviet invasion in 1979, the country has enduredalmost three and a half decades of civil conflict. After US and NATO military forces beganoperations to defeat the Taliban regime in October 2001, the new Afghan government hasworked with international aid donors to make significant progress in increasing primary schoolenrollment, reducing child and maternal mortality, and increasing income per capita. But asthe Taliban insurgency gained strength starting in 2006, the civilian population’s exposureto violence has continued to be a major issue. The United Nations estimates that duringthe six years from 2007 to 2012, over 14,500 civilians lost their lives in the armed conflict,including over 2,750 civilian deaths in 2012 alone. Approximately 80% of civilian casualitiesin 2012 were attributed to the insurgency, including a rise in both targetted killings andthe indiscriminate use of improvised explosive devices (United Nations 2013). As shown inFigure 2.1, recent violence has been particularly concentrated in the south and east of thecountry along the border with Pakistan where the insurgency is based.


Financial Development in Afghanistan

Afghanistan’s number of commercial bank branches per 100,000 adults is approximately2%, which is less than a quarter of the South Asia regional average of 8% (IMF 2011).Bank branches are typically limited to major urban centers, such as provincial capitals, andrarely operate in more remote areas of the country. The 2010 collapse of Kabul Bank, oneof the country’s largest financial institutions and the primary vehicle used to pay severalhundred thousand Afghan goverment salaries each month, further shook confidence in theformal financial system (Filkins, 2011). With only 3% of Afghans saving with a formal bankaccount, most rely on cash holdings and other informal savings vehicles (Demirguc-Kunt andKlapper, 2012). The money exchange network of hawala brokers offers an parallel system fordomestic and international payments, with limited funtionality for long-term savings, butdata on its size and scope in Afghanistan is limited by its informal nature (Maimbo, 2003).

Mobile Money in Afghanistan

Mobile phone ownership in Afghanistan has grown rapidly over the past decade, from ap-proximately 25,000 subscribers in 2002 to over 18 million subscribers in 2012 (World Bank2014). Roshan, the largest Afghan telecommunications operator, developed its M-Paisa mo-bile money platform in late-2008 with the British multinational Vodafone, and now boastsover 1.2 million M-Paisa subscribers, though the number of active users is far smaller.7 TheM-Paisa system was initially focused on micro-loan repayments, but it soon expanded toinclude peer-to-peer transfers and airtime purchases. Starting in 2009, M-Paisa expandedinto the mobile salary payment space as the Government of the Islamic Republic of Afghan-istan began a pilot project to pay Afghan National Police officers through the system, andRoshan began paying its own national employees via M-Paisa. Similar contracts to providemobile cash transfers to beneficiaries of humanitarian assistance soon followed. This periodalso marked a concentrated effort to significantly expand agent coverage outside of Kabul toinclude other major population centers such as Herat, Mazar, Jalalabad, Helmand and Kan-dahar. In early 2012, Roshan’s competitor Etisalat launched its own mobile money service,M-Hawala, and the remaining mobile operators have expressed plans to follow.

As a 2011 market assessment noted, mobile money in Afghanistan faces “the challenge ofdelivering services in a landscape with low levels of trust in formal institutions to consumerswith highly variable degrees of textual, financial and technological literacy” (Chipchase etal., 2011). While M-Paisa enjoys certain clear advantages of cost, time and privacy relative toalternative financial transfer options such as banks, hawala or in-person exchange, potentialusers also cited common concerns about penetration, accessibility and perceived risk as

7Four major mobile operators compete in Afghanistan: Afghan Wireless Communications Company(AWCC), Etislat, Mobile Telephone Network (MTN), and Roshan. In addition, two minor operators are inthe market: Afghan Telecom and Wasel Telecom, with each covering less than 3% of the market. In 2012,Roshan had an estimated subscriber base of over 5.6 million and an estimated market share of 32%, withcoverage in all 34 provincial capitals and 230 of Afghanistan’s 398 districts (Hamdard, 2012).


deterring adoption. However, brand recognition and trust in major mobile operators such asRoshan continues to grow, alongside efforts to expand the coverage of mobile money agentsand increase the number of channels willing to accept mobile money as a means of exchange.One noteworthy feature of mobile money in Afghanistan is that government regulationsrequire mobile operators to maintain regular deposits in local banks equal to the entire valueheld on their mobile money system, creating a significant connection between mobile moneyusers and the existing financial system.

Mobile money adoption in Afghanistan is best understood in the broader context ofthe global adoption of mobile money. Launched in 2007, the most successful and well-known deployment of mobile money in the developing world has been Safaricom’s M-PESAplatform in Kenya, which is used by approximately 17 million Kenyans (over two-thirdsof the population) and carries approximately 25% of the country’s gross national product(Economist 2013). As of late 2013, over two hundred mobile money deployments were activein 80 developing countries, with approximately two-thirds being launched in the past threeyears (GSMA 2014). But despite some notable exceptions such as MTN Uganda, VodacomTanzania, FNB in South Africa, and GCASH and Smart Money in the Phillipines, globalmobile money adoption has struggled to match the impressive growth rate of Safaricom’s M-PESA. In 2012, only six mobile money platforms had more than 1 million active customers- three of which crossed that threshold during that year (GSMA 2013). According to WorldBank figures, approximately 16% of adults in Sub-Saharan Africa report having used a mobilephone to pay bills or send or receive money over the past year, though much of that massis concentrated in the successful East African deployments.8 In Afghanistan, almost 7% ofadults report using a mobile phone to receive money and 3% report sending money by mobilephone (Demirguc-Kunt and Klapper, 2012).

Mobile Salary Payments

Given widespread adoption of mobile phones, mobile money provides a promising alternativeto bank or cash transfers for moving funds across large distances at low cost using a simpleSMS technology.9 In the particular case of mobile salary payments - wage transfers madeby an employer to an employee using mobile money - large firms are able to instaneouslycomplete individual financial transfers to their employees. Individual users are notified ofa transfer into their account by SMS message, and can check their balance and completeother functions using a simple interface that does not require smart-phone technology. Forthe firm, mobile salary payments offer a means to address concerns around physical security,logistics and corruption associated with cash salary payments by effectively outsourcing cashmanagement to the mobile operator’s network of mobile money agents. These agents functionas “human ATMs,” providing deposit and withdrawal services to individual users interestedin converting either their cash holdings into mobile money or vice-versa. Individuals users

8For example, there are now more mobile money accounts than bank accounts in Kenya, Madagascar,Tanzania and Uganda (GSMA 2013).

9Illiterate users can also access the M-Paisa platform using an interactive voice response (IVR) system.


can maintain a balance on their mobile money account, providing them with a storage ofvalue functionality.10 Individual users also can use the mobile money platform as a meansof exchange: to purchase pre-paid airtime directly from their mobile operator, to send andreceive mobile money with other mobile subscribers in the same country (either on the samemobile network or on a competitor’s network), and to receive remittance transfers fromoutside their country through partnerships with firms such as Western Union.11

2.4 Administrative Data Results

Our primary focus is on understanding the effect of violence on an financial decision-makingin Afghanistan. We begin by providing robust evidence that exposure to violence decreasesthe likelihood that an individual will use, and store balance in, his M-Paisa mobile moneyaccount. To do this, we create a novel dataset that combines the complete history of M-Paisatransactions over a 6-year period with administrative records of all violent incidents recordedby international forces in Afghanistan. To join these datasets and determine each M-Paisasubscriber’s exposure to violence over time, we have worked with Afghanistan’s primarymobile phone operator to obtain the complete anonymized and geo-tagged mobile phone callrecords of each M-Paisa user, which allows us to approximately locate each individual useron every day for which we have data.

Using methods described in greater detail in Appendix A, we create a balanced panel ofdata that captures, for each individual i in each time period t, several different measures ofM-Paisa use, which we denote by Yit. The mobile phone records are then used to determineeach individual’s “Center of Gravity”, a weighted centroid of the locations from which heis known to make or receive phone calls, which provides an approximate location COGit

for each individual in each time period. Finally, we measure each individual’s exposure toviolence V iolenceit by assigning each known violent incident vlt at location l at each time tto each individual who is within a fixed radius R of the incident, i.e.

V iolenceit = 1

[∑vlr

1 > 0

],∀vlt s.t. distance(COGit, vlt) < R

Given this balanced panel, we estimate the impact of violence on M-Paisa use with aregression model that includes individual fixed effects πi, district fixed effects ηd and timefixed effects µt.

Yit = βV iolenceit + πi + ηd + µt + εit (2.1)

The results we present below use a specification that attaches each violent incident toany individual within a 10 kilometer radius, i.e. R = 10, but Appendix Table A2.1 shows

10As in the case of Afghanistan, local regulations may restrict the payment of interest on mobile moneyaccounts not linked to a bank account, and also impose maxium balance limits on mobile money accounts.

11While deposits and airtime purchases are costless on Roshan’s M-Paisa platform, other mobile moneytransactions such as withdrawals and peer-to-peer transfers involve a graduated tariff structure. The mobilesalary payments product includes the cost of one withdrawal each month.


that our estimates are robust to a wide range of plausible values for R. We aggregate eventsand transactions at the weekly level, though again our results are robust to different levelsof temporal aggregation. We will further focus attention on specific population of M-Paisausers who we consider most relevant for policy analysis: (i) users who have at least two daysof recorded activity on the M-Paisa platform - allowing us to ignore short term users whoare automatically enrolled or who use the platform very briefly, and (ii) users who receivesalary payments via the platform, as we observe limited evidence of deposits and peer-to-peertransfers in the general population of users.12 These restrictions limit our sample to a totalof 7,784 individual salary users during the period from December 2010 to April 2012.

Results

Using these administrative data, we find a strong negative relationship between violenceexposure and M-Paisa usage. Table 2.1 presents the results from the fixed-effect specificationin Equation 2.1, and is identified based on within-individual changes over time. In otherwords, on average, individuals exposed to violence significantly reduce their M-Paisa balanceduring periods of heightened violence (column 1). More precisely, exposure to violenceis associated with a decrease in a user’s average daily M-Paisa balance of 259 Afghanis(approximately $5 USD), which is 12% of the mean value of the dependent variable.

Columns (2) - (6) of Table 2.1 indicate violence has similar effects on the extensivemargin of M-Paisa use: violence is associated with a reduction in activity in all of themost common M-Paisa transaction types, including deposits, withdrawals, and peer-to-peertransfers. Column (3) shows the coefficient on the violence indicator for withdrawals is 9%of the mean of the dependent variable, while columns (4)-(6) show related effect sizes of 62%on deposits, 9% on airtime purchases, and 22% on peer-to-peer transfers.

The negative correlation between violence and M-Paisa use also exists in the cross section,such that individuals located in violent areas are also less likely to use M-Paisa. These resultsare presented in Appendix Table A2.2, where we estimate variants of Equation 2.1 with andwithout a variety of fixed effects. However, since a large number of omitted variables couldreasonably account for the observed correlation between violence and M-Paisa use, we findthese results less straightforward to interpret.

2.5 Experimental Results

The administrative results provide compelling evidence that exposure to violence is asso-ciated with reduced use of Afghanistan’s mobile money system, even when controlling forunobserved heterogeneity at the individual level. However, a causal interpretation of theseresults is difficult, since we are unable to control for unobserved and time-varying hetero-geneity in which users join the mobile money platform. Moreover, the administrative data

12As shown in Appendix Table A2.3, our estimates are qualitatively similar when we relax the latterassumption to include non-salary users.


alone provides limited insight into the mechanisms driving individual decisions to reduceusage of M-Paisa.

To address these econometric concerns and better understand the impact of violence ona wider range of financial decisions, we conduct a randomized control trial in Afghanistanin which we induce random variation to an individual’s propensity to adopt mobile money.In our experiment, employees of a large, Afghan-staffed firm operating in some of the mostviolent areas of the country were randomly assigned to receive their monthly salary paymentsin mobile money or remain in the status quo cash payment system. We combine detailedadministrative transaction records with monthly survey data on both the treatment andcontrol group to achieve a more detailed understanding of the mechanisms underlying theindividual decisions to reduce usage of M-Paisa.

Research Partner

Headquartered in Singapore, the Central Asia Development Group (CADG) is a privatecontractor that delivers engineering, aviation, agricultural services and development assis-tance to remote and challenging locations. In Afghanistan, CADG’s flagship developmentinitiative has been a USAID-supported Community Development Program (CDP), primarilybased in the conflict-affected southern and eastern provinces of the country. CDP’s primaryobjective is to provide labor-intensive community development projects to reduce the impactof economic vulnerability and increase support for the Government of the Islamic Republicof Afghanistan. The projects undertaken by the communities involved reconstructing munic-ipal infrastructure, irrigation systems and valued public facilities such as schools and clinics.CDP’s main beneficiaries are at-risk populations including unemployed men of combat age,internally displaced persons, those suffering from extreme poverty and other marginalizedsegments of Afghan society. In 2011, a small number of CADG’s CDP staff in Kabul andKandahar entered a pilot of Roshan’s mobile salary payment program on the M-Paisa plat-form. Salaries were authorized directly from CADG’s Singapore headquarters using an onlineinterface and delivered monthly to the participating employees’ mobile phones via SMS no-tification. In mid-2012, the firm decided to scale up its use of mobile salary payments in theCDP program, and agreed to a randomized experiment to study the effects on its employees.

Protocol

In July 2012, CADG’s Community Development Program (CDP) employed approximatelythree hundred seventy-five (375) employees based in eight offices located in the capital Kabuland in the southern and eastern provinces of Afghanistan. The research study was launchedin August 2012 with a randomized experiment involving 341 CDP employees operating inseven provinces: Ghazni, Helmand, Kabul, Kandahar, Khost, Paktia and Paktika (see Fig-ure 2.2).13 Throughout the analysis that follows, we trim the top .5% of outliers in M-Paisa

13Employees in Zabul province could not be included due to a lack of reliable mobile coverage on theRoshan network in their area.


balances, which results in discarding one extreme outlier observation in the treatment groupwith an average M-Paisa balance 10 standard deviations above the mean, leaving a final sam-ple of 340 employees.14 The experimental sample included all CDP employees who workedin office locations with Roshan mobile coverage, and excluded the CDP security staff whowere being transitioned to an alternative payment system under the Afghan Public Protec-tion Force (APPF). Half of the employees in the experiment were randomly assigned to themobile salary system, while the other half were paid by CADGs existing cash-based systemto provide a valid comparison group during the study period. A single treatment arm wasselected to make full use of the employee sample, to ensure compliance with the experimentaldesign, and to isolate the causal effect of mobile salary payments from associated treatmentsinvolving training, distribution of phones and registration for mobile money.

Employees in the control group receive a basket of interventions that closely resemblethose received by the employees in the treatment group. Both sets of employees receivea group training on the use of the M-Paisa mobile money system, including how to send,receive, deposit and withdraw funds, as well as how to purchase mobile airtime using mobilemoney. Both sets of employees are distributed new phones, which are identified as theirnew official work phones, and both sets of employees are given Roshan SIM cards, whichare identified as their personal property. As all phone usage is pre-paid, employees wereencouraged to use these new phones and SIMs for their personal calls as well, and theyare instructed not to remove the Roshan SIMs and replace them with other network SIMs.Finally, both sets of employees are individually registered for the M-Paisa service, which dueto know-your-customer regulations requires the recording of biographical information andcopies of photos and a national ID card. The key difference between treatment and controlgroups is that members of the treatment group had their salary distributed via the M-Paisamobile money service, while members of the control group continued to be paid in cash bytheir employer.

In addition to stratifying treatment within each province, the randomization protocolincluded two further blocking variables: the share of monthly income transferred to a family,and the level of monthly expenditure on phone airtime. In both cases, the variable’s distri-bution was divided into above and below the median, and the stratification was implementedusing that definition. While employees in five provinces are able to withdraw their mobilesalary funds by visiting a mobile money agent (typically a teller at a local bank branch ora local merchant with significant turnover to enable regular liquidity), employees in Pakitaand Paktika received regular in-person visits from an agent to their office in order to addresssecurity concerns specific to those two provinces.15

To address the logistical challenges of travelling within Afghanistan, treatment followeda staggered rollout plan in which Kabul employees received the intervention in July 2012,followed by employees in Paktia and Paktika in August 2012, employees in Ghazni and

14We also consistenly present results trimming the top .5% of outliers in self-reported cash savings inorder to address a handful of extreme values that appear to be enumerator data collection errors.

15Our main results are robust to excluding employees from both of these provinces from the analysis.


Khost in September 2012, and employees in Helmand and Kandahar in October 2012. Be-fore each group received new phones, training and M-Paisa registration (or notification oftheir treatment status), a first wave of face-to-face interviews takes place to collect moredetailed baseline information. Following the in-person baseline, monthly phone surveys wereconducted with employees at all sites. A second wave of face-to-face endline surveys tookplace at each province based on availability.16 We thus create an unbalanced monthly panelof employees in which provincial offices are enrolled in different months, but then experiencea similar monitoring regime in relative time.

Take-up

The randomization assignment protocol was implemented with 100% compliance, meaningall 171 employees assigned to receive mobile salaries were in fact paid by mobile salaries, andthe remaining 169 employees in the control group continued to be paid by cash payments forthe duration of the research study.17 Baseline administrative and survey data summarizedin Table A2.4 indicates balance on employee observables such as age, marital status, numberof children, ethnicity, tenure, salary, and usage of formal banks and hawala system.

Administrative and survey data summarized in Table 2.2 shows monthly M-Paisa ac-count usage, violence exposure and expectations, and other economic survey data. M-Paisaaccount usage data includes monthly average account balance, monthly total transactioncounts, and self-reported travel time and costs to M-Paisa agents. Employees report high-levels of violence exposure in response to the question “Has the neighborhood in whichyou currently live experienced an attack in the current calendar month (previous calendarmonth)?”, with approximately half of our sample answering affirmatively to this question atsome point during study period. We measure violence expectations using the following sur-vey question, which was collected from individual respondents on a monthly basis: “In youropinion, please tell us how likely you think it is that insurgent-related violence will occurin your neighborhood. Is this extremely likely, very likely, somewhat likely, not very likely,or extremely unlikely?” When coded on a likert scale, where 0 is extremely likely and 4 isextremely unlikely, this variable takes on an average value of 1.66 with a standard deviationof 1.13. For our analysis, we define a dummy variable Expects Violenceit that equals one ifrespondent i answered either “extremely likely” or “very likely” in month t.18 Additional

16Paktia and Paktika province offices were closed in December 2012, necessitating endline surveys inNovember 2012. Ghazni province office was closed in January 2013, allowing for an endline survey inDecember 2012. All remaining provinces had their endline face to face survey conducted in February 2013,followed by one additional month of phone surveys prior to the end of the study.

17The randomization pool included additional employees who had their employement terminated afterassignment but before treatment was implemented, so they are excluded from this analysis. We also excludefrom our analysis approximately one dozen CADG employees who had participated in the mobile salariespilot project prior to the research study.

18This violence expectations variable is strongly correlated with our violence exposure variables, partic-ularly Attack Last Month (=1), even when including employee and month fixed effects. We interpret it asa violence forecast based on a combination of updated priors based on recent exposure, private information


monthly survey data reported in this table includes monthly cash savings, expenditures,bank savings and cash transfers to friends and family members.

Results

We begin by demonstrating increased usage of mobile money in the treatment group incolumns (1) - (3) of Table 2.3, with large, positive and statistically effects on mobile moneybalances. We gradually introduce month fixed effects, strata fixed effects and employee fixedeffects to show the robustness of our results to increasingly restrictive sources of variation.We aggregate our transaction data to the monthly level and estimate the following differencein differences specification, where the onset of treatment is defined as the date of the firstmobile salary payment in a given province.

Yit = Treat x Postit + Treati + Postt + γt + ηi + τi + εit (2.2)

In the above specification, i indexes employees and t indexes months. Yit is the outcomevariable of interest, Treati is a dummy variable that equals one for individuals randomlyassigned to receive mobile salary payments, Postt is a dummy variable that equals one afterthe onset of treatment, Treat x Postit is a dummy variable that equals one if both Treatiand Postt equal one, γt is a month fixed effect, ηi is a strata fixed effect and τi is an employeefixed effect.

We next extend this regression framework to a triple-difference by including interactionswith the Expects Violenceit variable. As shown in columns (4) - (6) of Table 2.3, we find anaverage effect of the treatment on mobile savings balances during periods of high violencebeliefs that is consistently negative in sign, large in magnitude and statistically significant.It is satisfying to note that the magnitude and significance of the estimated effects does notvary across these increasingly restrictive specifications, especially when limiting attentiononly to within-employee variation in column (6).19

Figure 2.3 presents a graphical representation of average daily M-Paisa balances in thetreatment and control groups. While mobile money balances are slowly rising in the controlgroup over time, they are not significiantly different than zero during the period of theexperimental study. By contrast, the M-Paisa balances in the treatment group are largeand significantly different than zero, even after allowing for cash withdrawals immediatelyfollowing each pay period. Figure 2.4 presents a corresponding graphical representationof the treatment effect on M-Paisa balances when broken down into violence subgroups,though here the violence groups are fixed over the full period for each individual by takingthe average violence belief across all reported months. Again, we see strong evidence thatviolence expectations drives a faster exit from mobile money in our treatment sample.

and other subjective beliefs.19As Table A2.5 shows, our results in column (4)-(6) are qualitatively similar when separating the violence

expecations variable into each answer, though grouping them improves power.


In Table 2.4, we show corresponding and opposite effects of violence expectations onself-reported cash savings. In columns (1) - (3) we show that there is no direct effect oftreatment on cash savings. In columns (4) - (6) we then pool our treatment and control ob-servations and examine the effect of violence expectations directly on cash savings withoutany treatment interaction. Again, it is noteworthy that the magnitude and statistical signifi-cance of our results to not change dramatically when including fixed effects for month, strataand individual employee. Given the organization of our data in a high-frequency panel, thisrelationship seems covincingly causal. It is noteworthy that the magnitude of the increasein cash savings observed in columns (4) - (6) of Table 2.4 is more than 80% of the decreasein mobile money savings seen demonstrated in the corresponding columns of Table 2.3. InTable 2.5 we show that our results are unique to cash savings; other economic measures suchas bank savings, individual transfers and expenditure show no effect from increased violenceexpectations. In additional results presented in Table A2.6, we find that high violence beliefsare characterized by faster withdrawals immediately following pay day, consistent with thisinterpretation of switching from mobile savings to cash savings as expectations of futureviolence rise.

2.6 Mechanisms

Why do we observe individuals responding to violence by reallocating their financial portfo-lios to cash from mobile money? In examining this question, we consider the precautionarymotive (Keynes, 1936). If current conflict portends a more unstable future, the experienceof violence may cause individuals to update their beliefs. Correspondingly, the ability torespond flexibly to changing circumstances may feel more urgent, creating a preference forliquidity. To consume from mobile money, it must first be converted to cash from an agent.20

By this logic, violence should increase the relative demand for cash.Countervailing against this, mobile money offers security advantages compared with cash.

There are at least three reasons that these may not be enough to compensate for the reduc-tion in liquidity. First, the violence (and corresponding expectations) we measure relate togeneral, and mostly political, instability. We do not observe direct predation from theft orbribery or other forms of violence that are associated with a risk of carrying cash. Second,eruptions of violence in Afghanistan drive tremendous migration, usually to Pakistan andIran.21 Mobile money users tend to be wealthier, especially in our CADG sample, and maybe considering whether to leave Afghanistan after coalition troops withdraw at the end of2014. Mobile money is not convertible outside of Afghanistan. Third, the liquidity of mobilemoney might be a function of levels of violence. Mobile money operators based in insecure

20An exception to this is a small number of locations in Kabul directly accept mobile money as payment.21According the United Nations High Commissioner for Refugees (UNHCR), since 2002, 3.8 million

Afghans, about 12.75 percent of Afghanistan’s total population, have repatriated from Pakistan alone. Thereremain roughly 1.6 million Afghan refugees in Pakistan, with numbers likely to swell in coming years (UnitedNations High Commissioner on Refugees, 2014).


region demand much higher premia to transact mobile money than those in more stableregions. Mobile money operators refuse to operate altogether in highly unstable regions. Anincrease in violence might both increase the effective cost to withdraw mobile money anddecrease the probability that it can with be withdrawn at all.22

Violence and Cash Savings in a Large Household Survey

We test the relationship between violence expectations and cash savings in an entirely sepa-rate sample from Afghanistan, as described by Callen et al. (2014). These data, collected inDecember 2010, reflect 468 different primary sampling units (elections polling centers) acrossnineteen provincial capitals. Enumerators were told to begin at the coordinates of the pollingcenter and survey either 6 or 8 subjects. Surveys were conducted in individuals homes. Enu-merators adhered to the right hand rule random selection method and respondents withinhouses were selected according to a Kish grid (Kish, 1949). Keeping with Afghan custom,men and women were interviewed by field staff of their own gender. Three features of thesedata provide a means of testing whether our results might generalize beyond our experimen-tal sample. First, they afford much greater spatial coverage. Second, they reflect a periodtwo years prior to the mobile salary experiment. Last, they contain nearly identical savingsand violence expectations modules as in the data for the experiment.23

Table 2.6 presents results using the 2010 sample, where all columns include demographiccontrols and province fixed effects. Column (1) reports the relationship between cash savingsand an indicator variable for exposure to violence (defined as a violent attack recorded in theINDURE database in a 1km radius of the polling center within the past 3 years).24 Column(2) reports the relationship between cash savings and an indicator variable for violenceexpectations, where the indicator equals one for an above median value on the ten pointlikert scale. Both violence exposure and violence expectations are associated with highercash savings. Column (3) shows that the relationship between cash savings and individualexpectations of violence is robust to controlling for violence exposure. Column (4) shows thatthe interaction term between exposure and expectations is negative but insignificant whilethe direct effects of both variables remain significant, and column (5) demonstrates thatresults are qualitatively similar when not trimming the top .5% of outliers in cash savingsfrom the sample.

22In additional results presented in Table A2.7, we find no evidence that violent events in a district directlyaffect the operation of the mobile network, but do find evidence that violence decreases the number of agentspresent in a district and conducting transactions by approximately 5%. In further analysis presented inTable A2.8, we find that our main experimental results are robust to including such time-varying confoundsas household shocks, salary problems, salary satisfaction and expectations of future government control.

23The only difference between these modules that the expectations elicitation question in 2010 used a tenpoint likert scale while in 2013 it used a five point scale.

24Reported results are robust to alternative radius specifications, as well as to the exclusion of demographiccontrols and province fixed effects.


What Does our Violence Expectations Variable Measure?

Our violence expectations question asks subjects to directly state their subjective beliefs thata particularly state of the world “insurgent-related violence will occurring in your neighbor-hood” will obtain. A substantial literature discusses the elicitation of future probabilitiesand a large number of studies use Likert scale responses about a future event as a means ofobtaining a proxy for subjective beliefs about future events. Delavande et al. (2011) providea review of efforts to elicit subjective probabilities in developing countries, arguing that apoint estimates of the probability events may afford some advantages over using a Likertscale, but that Likert scale measures provide valid proxies. More relevant to our study, Dela-vande and Kohler (2009) show that individuals’ Likert scale responses about the probabilitythat they have HIV successfully predicts their actual status.

A simple way to describe the objective of the question is to think of a basic two periodmodel where payoffs are state-contingent. Imagine that an individual can consume a fractiona of their salary s and save a fraction (1 − a) at an interest rate of r. They will save untilthe indifference condition u(c0 + as) + δE[u(c1)] = u(c0) + δE[u(c1 + (1 + r)(1 − a)s)] issatisfied. Assume that, in the future period, they will survive with probability p and thatu(0) = 0. Then, the indifference condition simplifies to u(c0 + as)− u(c0) = δp[u(c1 + (1 +r)(1− a)s)− u(c1)]. Using the implicit function theorem, it is straightforward to show that∂x∂p

= δ[u(c1+(1+r)(1−a)s)−u(c1)]u′(c0+a)

> 0. This provides a simple result, which obtains in a range of

models. Ceteris paribus, increasing survival probabilities (or the probability that savings canbe converted into consumption) should increase current savings. We designed this question,using insights from the literature on subjective elicitation, to provide a proxy for p.

In practice, this question could be correlated with a range of confounds including: (i)general optimism; (ii) risk aversion; (iii) discount factors; and (iv) present bias. Table A2.11includes measures of each of these confounds as an additional regressor. The magnitude ofthe coefficient is stable and remains significant, providing additional evidence that the Likertscale measure of violence expectations contains additional information beyond that availablein the set of confounds.

2.7 Conclusion

Our data suggest that conflict substantially reduces the financial involvement of Afghans.Across three separate data sets, we find that violence-affected individuals hold substantiallymore cash. In some cases, these individuals hold twice as much cash as individuals who arenot affected. At the same time, in our experiment, we find that violence is associated witha halving of the amount of mobile salaries kept as mobile money.

Financial networks and mobile money in particular exhibit network externalities. Thevalue of a mobile money account depends on the number of people with whom a client cantransact. Moreover, mobile money agents will not operate unless they achieve a certainvolume of customers. The same is true of bank savings and electronic bank transfers, which


are virtually nonexistent in Afghanistan. The magnitudes we find are large enough to suggestthat violence poses a substantial barrier to the development of formal financial networks.

Subjects in our experiment provided a monthly panel of forecasts of violence. Usingwithin-subject estimates, a one-standard deviation increase in forecasts is associated withholding 20% percent less mobile money and 20% percent more cash. Expectations also appearto have more explanatory power than actual violence exposure. This finding is corroboratedin a separate sample using nationwide household survey data from Afghanistan, in whichwe observe a strong positive correlation between an individual’s subjective perception ofuncertainty and the amount he saves in cash relative to other technologies. Our empiricalanalysis also allows us to rule out several possible alternative explanations, for instance thatour effect is driven by increases in transaction and travel costs or by reductions in agentliquidity coinciding with violent events.

The adoption failure we observe does not appear to be primarily about the effects ofviolence on the general economy, transaction costs or the mobile money system. Rather,it operates at the level of individual decisions. Our work highlights the importance of in-dividual decision-making channels in understanding the economic consequences of violence,and suggests that the preference for cash which attends experience (and expecting) violencecreates an obstacle to the develop of robust formal financial networks.



Figure 2.1: Violent Incidents in Afghanistan (Dec 2010 - April 2012)

Figure 2.2: CADG Provincial Office Locations (2012)


Figure 2.3: Treatment Effect on M-Paisa Balance

Figure 2.4: Treatment Effect on M-Paisa Balance By Violence

CHAPTER 2. VIOLENCE AND FINANCIAL DECISIONS IN AFGHANISTAN 63T

able

2.1:

Adm

inis

trat

ive

Dat

aset

:V

iole

nce

and

M-P

aisa

Use

Dep

enden

tV

ar.

M-P

aisa

Bal

ance

Tra

nsa

ctio

ns

(#)

Wit

hdra

wal

s(#

)D

epos

its

(#)

Air

tim

e(#

)Sen

dM

oney

(#)

(1)

(2)

(3)

(4)

(5)

(6)

Vio

lent

Eve

nt

in10

km

(=1)

-259

.08*

**-0

.030

***

-0.0

07**

*-0

.002

***

-0.0

04**

*-0

.002

***

(35.

45)

(0.0

02)

(0.0

01)

(0.0

00)

(0.0

01)

(0.0

00)

Sam

ple

Sal

ary

Use

rsSal

ary

Use

rsSal

ary

Use

rsSal

ary

Use

rsSal

ary

Use

rsSal

ary

Use

rsM

ean

Dep

Var

2107

.34

0.19

10.

064

0.00

20.

034

0.00

8#

Indiv

idual

s77

8477

8477

8477

8477

8477

84#

Obse

rvat

ions

3149

8631

4986

3149

8631

4986

3149

8631

4986

R-S

quar

ed0.

620.

290.

180.

110.

380.

17W

eek

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Dis

tric

tF

EY

ES

YE

SY

ES

YE

SY

ES

YE

SIn

div

idual

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

end

ent

vari

ab

leis

the

M-P

aisa

mob

ile

mon

eyac

cou

nt

bal

ance

inA

fgh

anis

inco

lum

n(1

),th

enu

mb

erof

M-P

ais

atr

an

sact

ion

sin

colu

mn

(2),

the

nu

mb

erof

wit

hd

raw

als

inco

lum

n(3

),th

enu

mb

erof

dep

osit

sin

colu

mn

(4),

the

nu

mb

erof

air

tim

ep

urc

has

esin

colu

mn

(5)

an

dth

enu

mb

erof

pee

r-to

-pee

rm

obil

em

oney

tran

sfer

sin

colu

mn

(6).

Ob

serv

atio

nis

an

ind

ivid

ual

-wee

k.

Vio

len

ceva

riab

leis

ad

um

my

for

wh

eth

era

vio

lent

atta

ckw

asre

cord

edin

the

IND

UR

Ed

atase

tin

a10k

mra

diu

sof

the

Cen

ter

ofG

ravit

ylo

cati

onof

the

M-P

ais

aacc

ou

nt

use

r.R

obu

stst

and

ard

erro

rs,

clu

ster

edat

ind

ivid

ual

leve

l,in

par

enth

eses

.**

*p<

0.01

,**

p<

0.0

5,*

p<

0.1

.T

rim

min

gto

p1%

and

bott

om1%

ofou

tlie

rsin

M-P

aisa

bal

ance

.


Table 2.2: Summary Statistics: Experimental Sample

Variable Mean Std. Dev. N

Treat (=1) 0.502 0.5 2049

M-Paisa Usage:M-Paisa Balance (Afs) 3152.075 185337 2049Airtime (Afs) 52.143 263.977 2049Transactions (#) 1.515 2.229 2049Deposits (#) 0 0.022 2049Deposits (Afs) 0.244 11.046 2049Withdrawals (#) 0.381 0.533 2049Withdrawals (Afs) 11834.096 24344.253 2049Travel Time to M-Paisa Agent (minutes) 91.435 70.336 1700Travel Cost to M-Paisa Agent (Afs) 71.925 129.593 1691

Violence and Expectations:Attack Last Month (=1) 0.186 0.389 1699Attack This Month (=1) 0.166 0.372 1696Expects Violence (=1) 0.241 0.428 1446

Savings and Expenditure:Cash Savings (Afs) 6360.401 31659.076 1592Expenditure (Afs) 26748.62 49799.938 1711Bank Savings (Afs) 7439.492 84129.152 1629Cash Transfers (Afs) 8374.625 20628.509 1711


able

2.3:

Tre

atm

ent

Eff

ects

by

Vio

lence

Exp

ecta

tion

s

Dep

enden

tV

aria

ble

:M

-Pai

saB

alan

ce(A

fs)

(1)

(2)

(3)

(4)

(5)

(6)

Tre

atx

Pos

t69

64.2

1***

6976

.23*

**76

29.8

5***

7802

.61*

**77

09.4

9***

7169

.27*

**(1

020.

94)

(103

9.50

)(1

081.

11)

(138

8.57

)(1

374.

56)

(142

9.11

)T

reat

xP

ost

xE

xp

ects

Vio

lence

-407

7.51

***

-413

2.47

**-4

488.

86**

(141

8.53

)(1

796.

32)

(222

6.22

)E

xp

ects

Vio

lence

(=1)

29.3

6-1

251.

9047

0.46

(58.

80)

(840

.56)

(395

.99)

Sam

ple

All

All

All

All

All

All

Mea

nD

epV

ar31

14.9

731

14.9

731

14.9

731

53.9

631

53.9

631

53.9

6#

Em

plo

yees

335

335

335

334

334

334

#O

bse

rvat

ions

2018

2018

2018

1418

1418

1418

R-S

quar

ed0.

090.

190.

090.

110.

220.

10M

onth

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Str

ata

FE

NO

YE

S-

NO

YE

S-

Em

plo

yee

FE

NO

NO

YE

SN

ON

OY

ES

Notes:

Dep

end

ent

vari

able

isth

eM

-Pai

sam

ob

ile

mon

eyacc

ou

nt

bala

nce

inA

fgh

an

is,

an

dob

serv

ati

on

isan

emp

loye

e-m

onth

.A

vera

ge

exch

ange

rate

was

appro

xim

atel

y50

Afg

han

isto

the

doll

ar

du

rin

gst

ud

yp

erio

d.

Sta

nd

ard

erro

rscl

ust

ered

at

the

emp

loyee

leve

lin

par

enth

eses

,**

*p<

0.01

,**

p<

0.05

,*

p<

0.1

.T

he

Exp

ects

Vio

len

cesu

bgro

up

sco

rres

pon

dto

resp

on

ses

toth

equ

esti

on

“In

you

rop

inio

n,

ple

ase

tell

us

how

like

lyyo

uth

ink

itis

that

insu

rgen

t-re

late

dvio

len

cew

ill

occ

ur

inyo

ur

nei

ghb

orh

ood

.Is

this

extr

emel

ylikel

y,ve

ryli

kely

,so

mew

hat

likel

y,n

otve

ryli

kely

,or

extr

emel

yu

nli

kely

?”E

xtr

emel

yli

kely

an

dve

ryli

kely

are

cod

edas

Exp

ects

Vio

len

ce.

Reg

ress

ion

sin

clu

de

mon

th,

stra

taan

dem

plo

yee

fixed

effec

tsas

note

d.

Str

ata

incl

ud

ep

rovin

ces,

share

of

inco

me

tran

sfer

edto

fam

ily

(ab

ove/

bel

owm

edia

n),

and

leve

lof

mon

thly

exp

end

itu

res

on

mob

ile

air

tim

e(a

bov

e/b

elow

med

ian

).T

rim

min

gto

p.5

%of

ou

tlie

rsin

cash

savin

gs.


Table 2.4: Effect of Violence on Cash Savings

Dependent Variable: Cash Savings (Afs)

(1) (2) (3) (4) (5) (6)

Treat x Post 3563.33 3511.60 2894.64(2534.31) (2453.51) (2051.37)

Expects Violence (=1) 3744.93** 3068.99** 3524.88**(1472.66) (1496.42) (1484.81)

Sample All All All All All AllMean Dep Var 4545.10 4545.10 4545.10 4773.16 4773.16 4773.16# Employees 335 335 335 333 333 333# Observations 1459 1459 1459 1244 1244 1244R-Squared 0.01 0.11 0.02 0.01 0.10 0.02Month FE YES YES YES YES YES YESStrata FE NO YES - NO YES -Employee FE NO NO YES NO NO YES

Notes: Dependent variable is self-reported cash holdings in Afghanis, and observation is an employee-month.Average exchange rate was approximately 50 Afghanis to the dollar during study period. Standard errors clusteredat the employee level in parentheses, *** p<0.01, ** p<0.05, * p<0.1. The Expects Violence subgroups correspondto responses to the question “In your opinion, please tell us how likely you think it is that insurgent-related violencewill occur in your neighborhood. Is this extremely likely, very likely, somewhat likely, not very likely, or extremelyunlikely?” Extremely likely and very likely are coded as Expects Violence. Regressions include month, strataand employee fixed effects as noted. Strata include provinces, share of income transfered to family (above/belowmedian), and level of monthly expenditures on mobile airtime (above/below median). Trimming top .5% of outliersin cash savings.


able

2.5:

Vio

lence

and

Oth

erE

conom

icR

esp

onse

s

(1)

(2)

(3)

(4)

Cas

hSav

ings

(Afs

)B

ank

Sav

ings

(Afs

)T

ransf

ers

(Afs

)E

xp

endit

ure

(Afs

)

Exp

ects

Vio

lence

(=1)

3384

.82*

*23

66.9

111

36.5

110

78.7

4(1

583.

30)

(201

0.71

)(1

115.

04)

(157

5.79

)

Sam

ple

All

All

All

All

Mea

nD

epV

ar44

97.5

230

83.8

267

84.7

323

400.

59#

Em

plo

yees

315

316

316

316

#O

bse

rvat

ions

1165

1173

1233

1233

R-S

quar

ed0.

020.

010.

020.

07M

onth

FE

YE

SY

ES

YE

SY

ES

Pro

vin

ceF

E-

--

-Str

ata

FE

--

--

Em

plo

yee

FE

YE

SY

ES

YE

SY

ES

Notes:

Dep

end

ent

vari

able

isse

lf-r

epor

ted

cash

hold

ings

inco

lum

n(1

),se

lf-r

eport

edb

an

kd

eposi

tsin

colu

mn

(2),

self

-rep

ort

edtr

ansf

ers

inco

lum

n(3

)an

dse

lf-r

epor

ted

exp

end

itu

res

inco

lum

n(4

).A

lld

epen

den

tva

riab

les

are

inA

fgh

an

isan

dob

serv

ati

on

isan

emp

loye

e-m

onth

.A

vera

geex

chan

gera

tew

as

ap

pro

xim

ate

ly50

Afg

han

isto

the

doll

ar

du

rin

gst

ud

yp

erio

d.

Sta

ndard

erro

rscl

ust

ered

atth

eem

plo

yee

leve

lin

pare

nth

eses

,***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

.T

he

Exp

ects

Vio

len

cesu

bgro

up

sco

rres

pon

dto

resp

onse

sto

the

qu

esti

on“In

you

rop

inio

n,

ple

ase

tell

us

how

like

lyyo

uth

ink

itis

that

insu

rgen

t-re

late

dvio

len

cew

ill

occ

ur

inyo

ur

nei

ghb

orh

ood

.Is

this

extr

emel

yli

kely

,ver

yli

kely

,so

mew

hat

like

ly,

not

very

like

ly,

or

extr

emel

yu

nli

kely

?”E

xtr

emel

yli

kely

and

very

like

lyare

cod

edas

Exp

ects

Vio

len

ce.

Reg

ress

ion

sin

clu

de

month

an

dem

plo

yee

fixed

effec

tsas

not

ed.

Str

ata

incl

ud

ep

rovin

ces,

share

of

inco

me

tran

sfer

red

tofa

mil

y(a

bov

e/b

elow

med

ian

),an

dle

vel

of

month

lyex

pen

dit

ure

son

mob

ile

airt

ime

(ab

ove/

bel

owm

edia

n).

Tri

mm

ing

top

.5%

of

ou

tlie

rsin

cash

savin

gs,

ban

ksa

vin

gs,

tran

sfer

san

dex

pen

dit

ure

s.


Table 2.6: Violence and Cash Savings from a Large Household Survey

Dependent Variable: Cash Savings (Afs)(1) (2) (3) (4) (5)

Attacks (=1) 221.39** 222.24** 246.94** 408.84**(88.39) (88.19) (110.69) (164.36)

Expects Violence (=1) 143.59* 145.20* 165.58* 196.19(86.39) (86.82) (100.00) (119.33)

Attacks x Expects -50.63 -100.48(157.46) (214.59)

Sample Trimmed Trimmed Trimmed Trimmed AllMean Dep Var 903.335 903.335 903.335 903.335 990.422# Clusters 468 468 468 468 468# Observations 3033 3033 3033 3033 3047R-Squared 0.148 0.146 0.149 0.149 0.114Demographic Controls YES YES YES YES YESProvince FE YES YES YES YES YES

Notes: Dependent variable is self-reported cash holdings in Afghanis, and observation is anindividual respondent in a 19 province survey during 2011 (see paper text for more details).Average exchange rate was approximately 50 Afghanis to the dollar during survey period. Robuststandard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1. The Attacks variable recordswhether a polling center had experienced an attack within 1km radius in the previous 3 yearsas recorded in the INDURE dataset (see paper text for more details). The Expects Violencesubgroups correspond to responses to the question “In your opinion, please tell us how likely youthink it is that insurgent-related violence will occur in your neighborhood.” Respondents weregiven a 0-10 point likert scale where 10 represented a certainty of violence forecast; responsesabove the median (corresponding to a 5 or higher on the scale) are coded as Expects Violence.Demographic controls include age, gender, education, employment, and risk attitudes. Trimmingtop .5% of outliers in cash savings in columns (1) - (4).



Data Appendix

Administrative Data

M-Paisa transaction records

The M-Paisa transaction records cover the universe of all transactions conducted on Afghan-istan’s primary mobile money network from its launch in November 2008 until December2013. We observe detailed information on each deposit, withdrawal, purchase, and peer-to-peer transfer. We use these transaction histories to calculate each subscriber’s daily“Cumulative Balance,” a running total of the total daily value stored on each subscriber’saccount.25

Violent incidents in Afghanistan

We integrate violence incident records covering the period January 2011 to December 2013from the International Security Assistance Force, a multilateral military body present insince December 2001, obtained through the International Distributed Uniform ReportingEnvironment (INDURE). In addition to geocodes at 5 decimal digit precision (accurate towithin one meter at the equator), these data provide the time and categorization of the inci-dent. In effect, these data capture all types of violence reported to the International SecurityAssistance Force by military, diplomatic, aid and non-governmental sources, including inci-dents in which the force was not directly engaged. These data identify two types of incidents:enemy attacks (including direct fire, indirect fire, suicide attacks and other kinetic activites)and explosions (including improvised explosive device explosions and mines strikes).

We combine both types of incidents in the empirical analysis, and attach each incident toany individual with a 10 kilometer halo. That is, if an incident is further than 10 kilometerfrom any individual’s location it will not be used in the analysis and if an incident lies within10 kilometer of two individuals, it will be attached to both of them. We define an indicatorvariable for violence exposure that equals one on a given day if an attack occurs in the 10kilometer halo of that subscriber’s location.

Physical locations, extracted from call detail records

Finally, to determine which M-Paisa subscribers are likely to have been affected by eachviolent event, we calculate each subscriber’s “Center of Gravity” for every day on whichthey are active on the mobile phone network. While M-Paisa transactions are not labelled

25Due to data recovery issues, we are missing all transaction records associated with 24 days of M-Paisadata. As cumulative account balances are calculated by aggregating over the entire transaction history, thesemissing data days create the potential for extreme positive and negative balances. We address this potentialsource of bias in our analysis by trimming the top 1% and bottom 1% of users by cumulative balance.


with geographic locations, each time a subscriber sends or receives a phone call or textmessage the network operator logs the cellular tower closest to the subscriber at the momentthe call was initiated. We extract all such tower information for each M-Paisa subscriber and,as is discussed in greater detail in Blumenstock (2012), we use this information to esimatethe center of gravity COGit of individual i at time t as

COGit =1

Nit

Tmax∑s=Tmin

K(t− sh

) · qis

where Nit is the total number of phone calls made by i within a window of time [Tmin, Tmax]symmetric around t, and qis is the (known) location of the tower used at time s. The kernelK(x) is a symmetric function that integrates to one, which specifies the extent to whichadditional weight is placed on calls close in time to t. In our results we use a uniform kernelsuch that K(u) = 1/Ni, however very little changes if a different kernel is specified.




able

A2.

1:A

dm

inis

trat

ive

Dat

aset

:V

iole

nce

and

M-P

aisa

Use

Dep

end

ent

Var

.M

-Pai

saB

alan

ceT

ran

sact

ion

s(#

)

(1)

(2)

(3)

(4)

(5)

(6)

Vio

lent

Eve

nt

in5

km

(=1)

-144

.60*

**-0

.014

***

(34.

94)

(0.0

02)

Vio

lent

Eve

nt

in15

km

(=1)

-120

.72*

**-0

.038

***

(37.

04)

(0.0

02)

Vio

lent

Eve

nt

in20

km

(=1)

-84.

84**

-0.0

45**

*(3

9.48

)(0

.002

)

Sam

ple

Sal

ary

Use

rsS

alar

yU

sers

Sal

ary

Use

rsS

alar

yU

sers

Sal

ary

Use

rsS

alar

yU

sers

Mea

nD

epV

ar21

07.3

421

07.3

421

07.3

40.

191

0.19

10.

191

#In

div

idu

als

7784

7784

7784

7784

7784

7784

#O

bse

rvat

ion

s31

4986

3149

8631

4986

3149

8631

4986

3149

86R

-Squ

ared

0.62

0.62

0.62

0.28

0.29

0.29

Wee

kF

EY

ES

YE

SY

ES

YE

SY

ES

YE

SD

istr

ict

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Ind

ivid

ual

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

enden

tva

riab

leis

the

M-P

aisa

mob

ile

mon

eyac

count

bal

ance

inA

fghanis

inco

lum

ns

(1)-

(3),

and

the

num

ber

of

M-P

ais

atr

ansa

ctio

ns

inco

lum

ns

(4)-

(6).

Obse

rvat

ion

isan

indiv

idual-

wee

k.

Vio

lence

vari

able

isa

dum

my

for

whet

her

avio

lent

atta

ckw

asre

cord

edin

the

IND

UR

Edat

aset

ina

5km

,15

km

or20

km

radiu

sof

the

Cen

ter

ofG

ravit

ylo

cati

onof

the

M-P

ais

aac

count

use

ras

note

dab

ove.

Robust

standar

der

rors

,cl

ust

ered

at

indiv

idual

leve

l,in

pare

nth

eses

.**

*p<

0.01

,**

p<

0.0

5,

*p<

0.1.

Tri

mm

ing

top

1%

and

bot

tom

1%of

outl

iers

inM

-Pai

sabala

nce

.


able

A2.

2:A

dm

inis

trat

ive

Dat

aset

:V

iole

nce

and

M-P

aisa

Use

Dep

end

ent

Var

.M

-Pai

saB

alan

ceT

ran

sact

ion

s(#

)

(1)

(2)

(3)

(4)

(5)

(6)

Vio

lent

Eve

nt

in10

km

(=1)

-923

.65*

**-1

014.

52**

*-2

80.1

5***

0.03

1***

0.04

2***

-0.0

22**

*(1

26.3

2)(1

40.6

1)(5

3.35

)(0

.002

)(0

.003

)(0

.002

)

Sam

ple

Sal

ary

Use

rsS

alar

yU

sers

Sal

ary

Use

rsS

alar

yU

sers

Sal

ary

Use

rsS

alar

yU

sers

Mea

nD

epV

ar21

07.3

421

07.3

421

07.3

40.

191

0.19

10.

191

#In

div

idu

als

7784

7784

7784

7784

7784

7784

#O

bse

rvat

ion

s31

4986

3149

8631

4986

3149

8631

4986

3149

86R

-Squ

ared

0.00

0.01

0.05

0.00

0.05

0.08

Wee

kF

EN

OY

ES

YE

SN

OY

ES

YE

SD

istr

ict

FE

NO

NO

YE

SN

ON

OY

ES

Ind

ivid

ual

FE

NO

NO

NO

NO

NO

NO

Notes:

Dep

enden

tva

riab

leis

the

M-P

aisa

mob

ile

mon

eyac

count

bal

ance

inA

fghanis

inco

lum

ns

(1)-

(3),

and

the

num

ber

of

M-P

ais

atr

ansa

ctio

ns

inco

lum

ns

(4)-

(6).

Obse

rvat

ion

isan

indiv

idual-

wee

k.

Vio

lence

vari

able

isa

dum

my

for

whet

her

avio

lent

atta

ckw

asre

cord

edin

the

IND

UR

Edat

aset

ina

10km

radiu

sof

the

Cen

ter

ofG

ravit

ylo

cati

on

of

the

M-P

aisa

acco

unt

use

r.R

obust

standar

der

rors

,cl

ust

ered

atin

div

idual

level

,in

par

enth

eses

.**

*p<

0.01,

**p<

0.05,

*p<

0.1

.T

rim

min

gto

p1%

and

bott

om

1%of

outl

iers

inM

-Pais

abal

ance

.


able

A2.

3:A

dm

inis

trat

ive

Dat

aset

:V

iole

nce

and

M-P

aisa

Use

Dep

enden

tV

ar.

M-P

aisa

Bal

ance

Tra

nsa

ctio

ns

(#)

Wit

hdra

wal

s(#

)D

epos

its

(#)

Air

tim

e(#

)S

end

Mon

ey(#

)

(1)

(2)

(3)

(4)

(5)

(6)

Vio

lent

Eve

nt

in10

km

(=1)

-156

.51*

**-0

.043

***

-0.0

05**

*-0

.005

***

-0.0

11**

*-0

.001

***

(20.

80)

(0.0

05)

(0.0

00)

(0.0

01)

(0.0

04)

(0.0

00)

Sam

ple

All

Use

rsA

llU

sers

All

Use

rsA

llU

sers

All

Use

rsA

llU

sers

Mea

nD

epV

ar15

23.8

30.

177

0.04

30.

009

0.04

40.

006

#In

div

idual

s14

661

1466

114

661

1466

114

661

1466

1#

Obse

rvat

ions

4773

0447

7304

4773

0447

7304

4773

0447

7304

R-S

quar

ed0.

630.

260.

210.

210.

250.

13W

eek

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Dis

tric

tF

EY

ES

YE

SY

ES

YE

SY

ES

YE

SIn

div

idual

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Notes:

Dep

end

ent

vari

ab

leis

the

M-P

aisa

mob

ile

money

acco

unt

bal

ance

inA

fgh

anis

inco

lum

n(1

),th

enu

mb

erof

M-P

aisa

tran

sact

ion

sin

colu

mn

(2),

the

nu

mb

erof

wit

hd

raw

als

inco

lum

n(3

),th

enu

mb

erof

dep

osit

sin

colu

mn

(4),

the

nu

mb

erof

airt

ime

pu

rch

ases

inco

lum

n(5

)an

dth

enu

mb

erof

pee

r-to

-pee

rm

ob

ile

mon

eytr

ansf

ers

inco

lum

n(6

).O

bse

rvati

on

isan

ind

ivid

ual

-wee

k.

Vio

len

ceva

riab

leis

ad

um

my

for

wh

eth

era

vio

lent

atta

ckw

as

reco

rded

inth

eIN

DU

RE

dat

ase

tin

a10

km

radiu

sof

the

Cen

ter

of

Gra

vit

ylo

cati

onof

the

M-P

aisa

acc

ou

nt

use

r.R

obu

stst

and

ard

erro

rs,

clu

ster

edat

ind

ivid

ual

leve

l,in

par

enth

eses

.***

p<

0.01

,**

p<

0.0

5,

*p<

0.1

.T

rim

min

gto

p1%

an

db

otto

m1%

ofou

tlie

rsin

M-P

aisa

bala

nce

.


Table A2.4: Balance Tests (Treatment = Mobile Salary)

Cash Mobile Difference p-valueAge 35.130 36.205 1.075 0.409

[12.469] [11.474] (1.299) .Married (=1) 0.792 0.848 0.056 0.178

[0.407] [0.360] (0.042) .Number Children 2.822 3.386 0.563 0.108

[3.058] [3.386] (0.350) .Pashtun (=1) 0.762 0.788 0.026 0.578

[0.427] [0.410] (0.046) .Tenure (Months) 12.345 11.582 -0.763 0.475

[9.931] [9.664] (1.066) .Monthly Salary (1000 Afs) 34.037 35.555 1.518 0.666

[26.925] [37.018] (3.514) .Monthly Airtime Bill (Afs) 724.398 736.404 12.007 0.725

[312.042] [309.930] (34.084) .Family Transfer Share (=1) 0.508 0.511 0.003 0.936

[0.326] [0.323] (0.036) .Formally Banked (=1) 0.283 0.268 -0.015 0.756

[0.452] [0.444] (0.049) .Hawala User (=1) 0.219 0.216 -0.003 0.955

[0.415] [0.413] (0.045) .Roshan User (=1) 0.515 0.497 -0.018 0.745

[0.501] [0.501] (0.054) .Wants M-Paisa (=1) 0.310 0.312 0.002 0.965

[0.464] [0.465] (0.050) .Observations 169 171

Standard deviations in brackets and standard errors in parentheses.


Table A2.5: Treatment Effects by Violence Expectations

(1) (2) (3)M-Paisa Balance (Afs)

Treat x Post 8221.40*** 8641.70*** 9047.30***(2072.37) (2268.42) (3062.82)

Treat x Post x Extremely Unlikely 1944.98 770.92 -362.31(3352.61) (3378.89) (4439.53)

Treat x Post x Not Very Likely -2785.03 -3616.39 -7961.54**(2099.56) (2676.47) (3257.88)

Treat x Post x Very Likely -3797.13* -4092.17 -5356.30(2142.44) (2550.08) (3661.28)

Treat x Post x Extremely Likely -7252.35*** -12371.65** -11565.98**(2682.41) (5360.41) (5422.04)

Violence Extremely Unlikely -65.92 -181.92 243.12(81.81) (809.30) (349.16)

Violence Not Very Likely -39.74 -313.90 -1518.20(80.85) (1081.98) (1538.72)

Violence Very Likely -4.76 -1360.30 194.79(70.12) (909.58) (288.01)

Violence Extremely Likely -260.53 -5691.11 -506.58(300.07) (4567.01) (739.97)

Sample All All AllMean Dep Var 3153.96 3153.96 3153.96# Employees 334 334 334# Observations 1418 1418 1418R-Squared 0.11 0.22 0.11Month FE YES YES YESStrata FE NO YES -Employee FE NO NO YES

Dependent variable is the M-Paisa mobile money account balance in Afghanis,and observation is an employee-month. Average exchange rate was approxi-mately 50 Afghanis to the dollar during study period. Standard errors clusteredat the employee level in parentheses, *** p<0.01, ** p<0.05, * p<0.1. TheExpects Violence subgroups correspond to responses to the question “In youropinion, please tell us how likely you think it is that insurgent-related violencewill occur in your neighborhood. Is this extremely likely, very likely, somewhatlikely, not very likely, or extremely unlikely?” Strata include provinces, shareof income transfered to family (above/below median), and level of monthlyexpenditures on mobile airtime (above/below median). Trimming top .5% ofoutliers in cash savings.


able

A2.

6:E

ffec

tof

Vio

lence

onD

ays

toM

-Pai

saW

ithdra

wal

(1)

(2)

(3)

(4)

(5)

Day

sto

M-P

aisa

Wit

hdra

wal

Exp

ects

Vio

lence

(=1)

-1.1

7**

-1.0

5**

-1.1

9***

-1.1

7***

-1.2

9**

(0.4

6)(0

.42)

(0.4

5)(0

.41)

(0.5

6)

Sam

ple

Tre

atx

Pos

tT

reat

xP

ost

Tre

atx

Pos

tT

reat

xP

ost

Tre

atx

Pos

tM

ean

Dep

Var

3.22

3.22

3.22

3.22

3.22

#E

mplo

yees

162

162

162

162

162

#O

bse

rvat

ions

580

580

580

580

580

R-S

quar

ed0.

010.

040.

060.

210.

07M

onth

FE

NO

YE

SY

ES

YE

SY

ES

Pro

vin

ceF

EN

ON

OY

ES

--

Str

ata

FE

NO

NO

NO

YE

S-

Em

plo

yee

FE

NO

NO

NO

NO

YE

S

Dep

enden

tva

riab

leis

the

num

ber

ofday

sb

etw

een

sala

rydep

osit

and

firs

tsu

bse

quen

tw

ithdra

wal

inth

eM

-Pai

sam

obile

mon

eyac

count,

and

obse

rvat

ion

isan

emplo

yee-

mon

th.

Sta

ndar

der

rors

clust

ered

atth

eem

plo

yee

leve

lin

par

enth

eses

,**

*p<

0.01

,**

p<

0.05

,*

p<

0.1.

The

Exp

ects

Vio

lence

subgr

oups

corr

esp

ond

tore

spon

ses

toth

eques

tion

“In

your

opin

ion,ple

ase

tell

us

how

like

lyyo

uth

ink

itis

that

insu

rgen

t-re

late

dvio

lence

will

occ

ur

inyo

ur

nei

ghb

orhood.

Isth

isex

trem

ely

like

ly,

very

like

ly,

som

ewhat

like

ly,

not

very

like

ly,

orex

trem

ely

unlike

ly?”

Extr

emel

ylike

lyan

dve

rylike

lyar

eco

ded

asE

xp

ects

Vio

lence

.R

egre

ssio

ns

incl

ude

mon

th,

pro

vin

ce,

stra

taan

dem

plo

yee

fixed

effec

tsas

not

ed.

Str

ata

incl

ude

pro

vin

ces,

shar

eof

inco

me

tran

sfer

edto

fam

ily

(ab

ove/

bel

owm

edia

n),

and

leve

lof

mon

thly

exp

endit

ure

son

mob

ile

airt

ime

(ab

ove/

bel

owm

edia

n).

Tri

mm

ing

top

.5%

ofou

tlie

rsin

cash

hol

din

gs.


able

A2.

7:A

dm

inis

trat

ive

Dat

aset

:V

iole

nce

and

Net

wor

kO

per

atio

n

Dep

enden

tV

ar.

(1)

(2)

(3)

(4)

(5)

(6)

#T

ower

s#

Cal

lers

#C

alls

#A

gents

#A

gent

Txns

#M

-Pai

saT

xns

Vio

lence

inD

istr

ict

(=1)

-0.0

5-4

11.7

1-1

116.

95-0

.01*

**-0

.12

-0.3

7**

(0.1

4)(8

12.5

5)(3

004.

61)

(0.0

0)(0

.09)

(0.1

8)

Mea

nD

epV

ar6.

8812

139.

2560

278.

500.

200.

863.

05#

Dis

tric

ts39

839

839

839

839

839

8#

Obse

rvat

ions

2905

429

054

2905

429

054

2905

429

054

R-S

quar

ed0.

030.

020.

030.

040.

010.

01W

eek

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

Dis

tric

tF

EY

ES

YE

SY

ES

YE

SY

ES

YE

S

Notes:

Dep

end

ent

vari

able

isth

eav

erag

enu

mb

erof

act

ive

tow

ers

each

day

ina

dis

tric

tin

colu

mn

(1),

the

aver

age

nu

mb

erof

un

iqu

eca

ller

sea

chd

ayin

ad

istr

ict

inco

lum

n(2

),th

eav

erage

num

ber

of

call

sea

chd

ayin

ad

istr

ict

inco

lum

n(3

),th

eav

erage

nu

mb

erof

agen

tsea

chd

ayin

ad

istr

ict

inco

lum

n(4

),th

eav

erage

nu

mb

erof

M-P

ais

aagen

ttr

an

sact

ion

sea

chd

ayin

ad

istr

ict

inco

lum

n(5

),an

dth

eav

erag

enu

mb

erof

M-P

ais

atr

an

sati

on

sea

chd

ayin

ad

istr

ict

inco

lum

n(6

).O

bse

rvati

on

isa

dis

tric

t-w

eek.

Vio

len

ceva

riab

leis

ad

um

my

for

whet

her

avio

lent

att

ack

was

reco

rded

inth

eIN

DU

RE

data

set

ina

dis

tric

tin

that

wee

k.

Rob

ust

stan

dar

der

rors

,cl

ust

ered

atd

istr

ict

leve

l,in

pare

nth

eses

.***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

.


able

A2.

8:T

reat

men

tE

ffec

tsby

Vio

lence

Exp

ecta

tion

s-

Rob

ust

nes

s

Dep

end

ent

Var

.M

-Pai

saB

alan

ce(A

fs)

(1)

(2)

(3)

(4)

Tre

atx

Pos

t65

83.3

1***

5767

.57*

**69

77.7

1***

7735

.32*

**(1

344.

84)

(102

7.98

)(1

535.

72)

(174

8.58

)T

reat

xP

ost

xE

xp

ects

Vio

len

ce-4

587.

16**

-532

8.01

**-4

615.

50**

-513

1.05

**(2

264.

56)

(241

3.49

)(2

274.

12)

(244

0.31

)T

reat

xP

ost

xH

HS

hock

3567

.14

(485

2.61

)T

reat

xP

ost

xS

alar

yP

rob

lem

1028

0.70

(705

0.87

)T

reat

xP

ost

xL

owS

alar

yS

atis

fact

ion

963.

97(3

972.

70)

Tre

atx

Pos

tx

Low

Gov

ern

men

tC

ontr

ol24

7.48

(254

6.57

)

Sam

ple

All

All

All

All

Mea

nD

epV

ar31

53.9

631

48.7

531

37.3

033

18.9

1#

Em

plo

yees

334

334

334

332

#O

bse

rvat

ion

s14

1814

1014

1213

26R

-Squ

ared

0.11

0.15

0.11

0.11

Mon

thF

EY

ES

YE

SY

ES

YE

SS

trat

aF

E-

--

-E

mp

loye

eF

EY

ES

YE

SY

ES

YE

S

Notes:

Dep

enden

tva

riable

isth

eM

-Pais

am

obile

money

acc

ount

bala

nce

inco

lum

ns

(1)-

(4).

Obse

rvat

ion

isan

emplo

yee-

mon

th.

Ave

rage

exch

ange

rate

was

appro

xim

ate

ly50

Afg

hanis

toth

edol

lar

duri

ng

study

per

iod.

Sta

n-

dard

erro

rscl

ust

ered

at

the

emplo

yee

level

inpare

nth

eses

,**

*p<

0.0

1,

**p<

0.0

5,*

p<

0.1

.T

he

Exp

ects

Vio

lence

subgr

oups

corr

esp

ond

tore

sponse

sto

the

ques

tion

“In

your

opin

ion,

ple

ase

tell

us

how

like

lyyou

thin

kit

isth

atin

surg

ent-

rela

ted

vio

lence

will

occ

ur

inyour

nei

ghb

orh

ood.

Isth

isex

trem

ely

like

ly,

very

like

ly,

som

ewhat

like

ly,

not

ver

ylike

ly,or

extr

emel

yunlike

ly?”

Extr

emel

ylike

lyand

very

like

lyare

coded

as

Exp

ects

Vio

lence

.R

egre

ssio

ns

incl

ude

mon

than

dem

plo

yee

fixed

effec

tsas

note

d.

Str

ata

incl

ude

pro

vin

ces,

share

ofin

com

etr

ansf

erre

dto

fam

ily

(ab

ove/

bel

owm

edia

n),

and

leve

lof

month

lyex

pen

dit

ure

son

mob

ile

air

tim

e(a

bov

e/b

elow

med

ian).

Tri

mm

ing

top

.5%

of

outl

iers

inca

shsa

vin

gs.


able

A2.

9:V

iole

nce

and

Tra

nsa

ctio

nC

osts

(1)

(2)

(3)

(4)

(5)

M-P

aisa

Bal

ance

(Afs

)

Tre

atx

Pos

t71

69.2

7***

7357

.99*

**62

62.5

8***

7937

.36*

**80

39.5

8***

(142

9.11

)(1

616.

93)

(216

2.46

)(1

525.

25)

(219

6.03

)T

reat

xP

ost

xE

xp

ects

Vio

len

ce-4

488.

86**

-346

9.45

-369

2.22

(222

6.22

)(3

099.

60)

(308

2.06

)T

reat

xP

ost

xH

igh

Tim

eto

Age

nt

-709

.58

1510

.54

(203

1.63

)(3

068.

30)

Tre

atx

Pos

tx

Exp

ects

Vio

len

cex

Hig

hT

ime

toA

gent

-246

6.12

(472

0.16

)T

reat

xP

ost

xH

igh

Cos

tto

Age

nt

-223

5.34

-221

5.53

(207

4.51

)(3

054.

75)

Tre

atx

Pos

tx

Exp

ects

Vio

len

cex

Hig

hC

ost

toA

gent

-210

3.61

(438

7.28

)

Sam

ple

All

All

All

All

All

Mea

nD

epV

ar31

53.9

629

14.7

131

78.4

629

13.6

031

78.8

4#

Em

plo

yees

334

324

323

321

320

#O

bse

rvat

ion

s14

1816

7014

0716

6113

98R

-Squ

ared

0.10

0.11

0.11

0.11

0.11

Mon

thF

EY

ES

YE

SY

ES

YE

SY

ES

Pro

vin

ceF

E-

--

--

Str

ata

FE

--

--

-E

mp

loye

eF

EY

ES

YE

SY

ES

YE

SY

ES

Dep

end

ent

vari

able

isth

eM

-Pai

sam

obil

em

oney

acco

unt

bal

ance

inA

fgh

anis

,an

dob

serv

atio

nis

anem

plo

yee-

mon

th.

Ave

rage

exch

ange

rate

was

app

roxim

atel

y50

Afg

han

isto

the

dol

lar

du

rin

gst

ud

yp

erio

d.

Sta

nd

ard

erro

rscl

ust

ered

atth

eem

plo

yee

leve

lin

par

enth

eses

,**

*p<

0.01

,**

p<

0.05

,*

p<

0.1.

Th

eE

xp

ects

Vio

len

cesu

bgr

oup

sco

rres

pon

dto

resp

onse

sto

the

qu

esti

on“I

nyo

ur

opin

ion

,p

leas

ete

llu

sh

owli

kely

you

thin

kit

isth

atin

surg

ent-

rela

ted

vio

len

cew

ill

occ

ur

inyo

ur

nei

ghb

orh

ood

.Is

this

extr

emel

yli

kely

,ve

ryli

kely

,so

mew

hat

like

ly,

not

very

like

ly,

orex

trem

ely

un

like

ly?”

Extr

emel

yli

kely

and

very

like

lyar

eco

ded

asE

xp

ects

Vio

len

ce.

Hig

hT

ime

toA

gent

rep

rese

nts

abov

eth

em

edia

nin

rep

orte

dtr

avel

tim

eto

the

nea

rest

M-P

aisa

agen

t.H

igh

Cos

tto

Age

nt

rep

rese

nts

abov

eth

em

edia

nin

rep

orte

dtr

avel

cost

toth

en

eare

stM

-Pai

saag

ent.

Reg

ress

ion

sin

clu

de

mon

th,

pro

vin

ce,

stra

taan

dem

plo

yee

fixed

effec

tsas

not

ed.

Str

ata

incl

ud

ep

rovin

ces,

shar

eof

inco

me

tran

sfer

edto

fam

ily

(ab

ove/

bel

owm

edia

n),

and

leve

lof

mon

thly

exp

end

itu

res

onm

obil

eai

rtim

e(a

bov

e/b

elow

med

ian

).T

rim

min

gto

p.5

%of

outl

iers

inca

shh

old

ings

.


able

A2.

10:

M-P

aisa

Bal

ance

-V

iole

nce

Exp

ecta

tion

san

dV

iole

nce

Exp

osure

(1)

(2)

(3)

(4)

M-P

aisa

Bal

ance

(Afs

)

Tre

atx

Pos

t13

169.

65**

*71

69.2

7***

1098

5.53

***

1223

2.55

***

(252

8.21

)(1

429.

11)

(244

3.62

)(2

745.

10)

Tre

atx

Pos

tx

Att

acks

-196

.61*

**-1

41.5

2**

-187

.29*

**(6

3.91

)(5

4.65

)(6

7.38

)T

reat

xP

ost

xE

xp

ects

Vio

len

ce-4

488.

86**

-399

5.67

*-8

958.

74**

*(2

226.

22)

(216

9.40

)(3

345.

89)

Tre

atx

Pos

tx

Exp

ects

Vio

len

cex

Att

acks

179.

03**

(70.

04)

Att

acks

50.7

8***

11.0

145

.00*

*(1

5.91

)(2

1.92

)(2

2.74

)E

xp

ects

Vio

len

ce(=

1)47

0.46

-251

.19

1930

.06*

*(3

95.9

9)(8

26.2

3)(8

14.7

6)

Sam

ple

All

All

All

All

Mea

nD

epV

ar31

14.9

731

53.9

631

53.9

631

53.9

6#

Em

plo

yees

335

334

334

334

#O

bse

rvat

ion

s20

1814

1814

1814

18R

-Squ

ared

0.10

0.10

0.11

0.11

Mon

thF

EY

ES

YE

SY

ES

YE

SS

trat

aF

E-

--

-E

mp

loye

eF

EY

ES

YE

SY

ES

YE

S

Dep

end

ent

vari

able

isth

eM

-Pai

sam

obil

em

oney

acco

unt

bal

ance

inA

fgh

anis

,an

dob

serv

atio

nis

anem

plo

yee-

mon

th.

Ave

rage

exch

ange

rate

was

app

roxim

atel

y50

Afg

han

isto

the

dol

lar

du

rin

gst

ud

yp

erio

d.

Sta

nd

ard

erro

rscl

ust

ered

atth

eem

plo

yee

leve

lin

par

enth

eses

,**

*p<

0.01

,**

p<

0.05

,*

p<

0.1.

Th

eA

ttac

ks

vari

able

mea

sure

sth

enu

mb

erof

insu

rgen

t-re

late

dat

tack

sin

ap

rovin

cial

dis

tric

tca

pit

alov

erth

em

onth

asre

cord

edin

the

IND

UR

Ed

atas

et(s

eep

aper

text

for

det

ails

).T

he

Exp

ects

Vio

len

cesu

bgr

oup

sco

rres

pon

dto

resp

onse

sto

the

qu

esti

on“I

nyo

ur

opin

ion

,p

leas

ete

llu

sh

owli

kely

you

thin

kit

isth

atin

surg

ent-

rela

ted

vio

len

cew

ill

occ

ur

inyo

ur

nei

ghb

orh

ood

.Is

this

extr

emel

yli

kely

,ve

ryli

kely

,so

mew

hat

like

ly,

not

very

like

ly,

orex

trem

ely

un

like

ly?”

Extr

emel

yli

kely

and

very

like

lyar

eco

ded

asE

xp

ects

Vio

len

ce.

Reg

ress

ion

sin

clu

de

mon

than

dem

plo

yee

fixed

effec

tsas

not

ed.

Tri

mm

ing

top

.5%

ofou

tlie

rsin

cash

savin

gs.


able

A2.

11:

Vio

lence

and

Cas

hSav

ings

from

aL

arge

Hou

sehol

dSurv

ey

Dep

enden

tV

aria

ble

:C

ash

Sav

ings

(Afs

)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Exp

ects

Vio

lence

(=1)

284.

96*

293.

47*

299.

57*

305.

33*

337.

26*

320.

90*

283.

87(1

70.7

8)(1

71.0

0)(1

69.6

3)(1

68.6

7)(1

84.9

7)(1

87.3

3)(1

88.0

2)M

onth

lyD

isco

unt

Fac

tor

-284

5.62

-413

5.38

-408

2.08

(222

0.88

)(3

002.

02)

(304

8.89

)P

rese

nt-

Bia

sP

aram

ente

r-2

463.

3663

7.57

761.

05(2

568.

11)

(362

6.19

)(3

584.

23)

Lad

der

ofL

ife

(0-1

0)30

.23

10.7

714

.30

(40.

52)

(42.

25)

(41.

15)

Fin

anci

alR

isk

Lik

ert

(0-1

0)47

.82

52.8

711

.75

(40.

29)

(45.

80)

(44.

98)

Hol

t-L

aury

Ris

kM

easu

re73

6.22

*69

2.21

*90

.37

(393

.09)

(377

.71)

(421

.16)

Con

stan

t34

24.0

631

48.3

162

0.96

***

660.

31**

*37

1.38

*34

77.6

834

07.3

6(2

093.

28)

(250

3.47

)(1

88.6

3)(1

02.1

0)(2

23.4

1)(2

752.

22)

(274

8.00

)#

Clu

ster

s28

728

728

728

728

628

628

6#

Obse

rvat

ions

1122

1122

1122

1122

972

972

972

R-S

quar

ed0.

351

0.35

00.

349

0.35

10.

378

0.38

50.

406

Dem

ogra

phic

Con

trol

sN

ON

ON

ON

ON

ON

OY

ES

Pol

ling

Cen

ter

FE

YE

SY

ES

YE

SY

ES

YE

SY

ES

YE

S

Notes:

Dep

end

ent

vari

able

isse

lf-r

epor

ted

cash

hold

ings

inA

fgh

an

is,

an

dob

serv

ati

on

isan

ind

ivid

ual

resp

on

den

tin

a19

pro

vin

cesu

rvey

du

rin

g20

11(s

eep

aper

text

for

mor

ed

etail

s).

Ave

rage

exch

an

ge

rate

was

ap

pro

xim

ate

ly50

Afg

han

isto

the

dollar

du

rin

gsu

rvey

per

iod

.R

obu

stst

and

ard

erro

rscl

ust

ered

at

the

poll

ing

cente

rle

vel

inp

are

nth

eses

,***

p<

0.0

1,

**

p<

0.0

5,

*p<

0.1

.T

he

Exp

ects

Vio

len

cesu

bgr

oup

sco

rres

pon

dto

resp

on

ses

toth

equ

esti

on

“In

you

ropin

ion

,p

lease

tell

us

how

like

lyyou

thin

kit

isth

at

insu

rgen

t-re

late

dvio

len

cew

ill

occ

ur

inyou

rn

eighb

orh

ood

.”R

esp

on

den

tsw

ere

giv

ena

0-1

0p

oin

tli

kert

scale

wh

ere

10

rep

rese

nte

da

cert

ainty

ofvio

len

cefo

reca

st;

resp

onse

sab

ove

the

med

ian

(corr

esp

on

din

gto

a5

or

hig

her

on

the

scale

)are

cod

edas

Exp

ects

Vio

len

ce.

Dem

ogra

ph

icco

ntr

ols

incl

ud

eag

e,gen

der

,ed

uca

tion

,em

plo

ym

ent,

an

dri

skatt

itu

des

.T

rim

min

gto

p.5

%of

ou

tlie

rsin

all

colu

mn

s.

83

Chapter 3

Automatic Payroll Deductions inAfghanistan

3.1 Abstract

Automatic payroll deductions consistently represent one of the most effective means of in-creasing savings in developed countries. We design and experimentally evaluate a mobilephone-based account that allows savings to be automatically deducted from salaries inAfghanistan, a country with extremely low levels of formal financial inclusion. We findthat employees who are automatically enrolled in a defined-contribution account are 40percentage points more likely to contribute to the account than individuals with a defaultcontribution of zero. We also randomize employer matching contributions and find thatthe effect of automatic enrollment on participation is approximately equivalent to providingfinancial incentives equal to a 50 percent match. To understand why default enrollment in-creases participation, we randomly offer some employees an immediate financial consultation,and others a financial consultation in one week. Preliminary results indicate that employeesare more likely to discuss changing their savings contributions in one week, suggesting thatdefaults raise contributions because of the perceived complexity of financial decisions, andbecause employees procrastinate in developing a financial plan for the future.1

1The material in this chapter is based on joint work with Michael Callen and Joshua Blumenstock. SeeBlumenstock et al. (2015).

CHAPTER 3. AUTOMATIC PAYROLL DEDUCTIONS IN AFGHANISTAN 84

3.2 Introduction

In wealthy nations, defined-contribution accounts can dramatically increase an individual’scontributions to savings, and have been linked to lasting welfare improvements for par-ticipating employees (Thaler and Benartzi, 2004; Benartzi and Thaler, 2007). By allowingindividuals to automatically contribute a portion of their paycheck to savings, these accountshelp overcome self-control issues and improve long-term planning (Thaler and Shefrin, 1981;Madrian, 2012). These effects are most pronounced among poorer employees, and those withlower levels of financial sophistication (Madrian and Shea, 2001; Choi et al., 2004; Beshearset al., 2010).

In developing countries, dynamic inconsistency and difficulties with self-control have beenidentified as key impediments that prevent the poor from saving (Ashraf et al., 2006; Karlanet al., 2010; Dupas and Robinson, 2013). Yet, most savings products designed for thesecontexts require active effort on the part of the saver. Indeed, one of the main reasonswe observe low levels of savings and of formal financial participation is that the degree ofrequired effort is prohibitive. Given the large impact of automatic contributions relative tocontributions requiring an active decision (Chetty et al., 2014), an automatic contributionaccount suited to a developing country context offers considerable potential.

In this paper, we present experimental evidence of the impact of automatic payroll de-duction on savings in Afghanistan, a country where roughly 4 percent of the populationsaves in a formal financial institution (Demirguc Kunt et al., 2015). For this study, all 967full-time employees of a large Afghan telecom firm were given access to a new mobile phone-based savings account, called “M-Pasandaz.” Built into this account is an automatic payrolldeduction feature which, when activated, allows employees to have up to 10 percent of theirregular paycheck automatically deposited to the M-Pasandaz account. Half of employeeswere randomly assigned to a treatment in which this feature was by default activated andset to automatically deposit 5 percent of the employee’s monthly paycheck to M-Pasandaz.The remaining half of employees had their default contribution rate set to zero. Employeeswere further given a randomized level of matching incentives, whereby one third of employeesreceived a 50% bonus for all monthly contributions to M-Pasandaz; one third received a 25%match on all contributions; and the final third received no matching incentives.

We find that employees who are automatically enrolled in the defined-contribution ac-count are 40 percentage points more likely to contribute to the account than individuals witha default contribution of zero (Table 3.2). This effect is observed at all levels of matchingincentives, and persists despite the minimal transaction costs associated with choosing anon-default contribution rate. While about half of employees did in fact change to a non-default rate (Figure 3.2), the impact of the initial assignment is striking. Two months afterthe program launched and after almost all switching had ceased, 47% of employees with nomatching incentives but a default 5% contribution were still contributing at the default 5%level. Similarly, 45% of employees given the strongest matching incentives and a default 0%contribution were still “leaving money on the table” by remaining at the default 0% level.

Our research design also makes it possible to estimate the elasticity of demand for the


M-Pasandaz account relative to financial incentives, and to compare the impact of theseincentives to the impact of default assignment. Relative to a baseline of no matching incen-tives, we find that a 50% match on employee contributions increases employee participationby 47 percentage points, and the 25% match increases participation by 25 percentage points(Table A3.1). This effect is strongest among the group who are assigned a default contribu-tion of 0%, but is only slightly attenuated in the group assigned a default contribution of5%. Comparing the impact of the two orthogonal treatments, we note that for the employ-ees in our study, default enrollment increases participation by roughly the same amount asproviding matching incentives of 50%.

We next turn our attention to understanding why the default contribution rate has sucha large impact on savings. While a large set of candidate explanations exist (cf. Beshearset al., 2009; Madrian, 2012), our experimental setting makes it possible to rule out severalcommon hypotheses. For instance, follow-up surveys with employees allow us to reject thepossibility that employees were unaware of their enrollment status, or that there was anyconfusion about how to change the assigned contribution level. Similarly, because of theway in which the product was first introduced and the nature of the public lottery thatwas used to announce treatment assignments, it seems unlikely that employees perceivedtheir assignment to be a sign of employer endorsement, as is common in situations whenall employees are assigned a uniform plan and rate (Madrian and Shea, 2001; Beshears etal., 2009). Finally, we use three different sources of experimental and non-experimentalvariation to rule out the possibility that the default effects are driven simply by inattentionon the part of the employee. First, all employees are notified each month of their paycheckamount; employees are quite sensitive to this paycheck (which usually provides a paymentthat is a round number), and it thus provides a tangible reminder of their enrollment status.Second, we sent text messages to employees to remind them of their current rate and provideinformation on how to switch, and we find only a small impact of these reminders on switchingbehavior. Third, because we were worried that our follow-up surveys might impact employeebehavior (cf. Zwane et al., 2011), we restricted our post-treatment surveys to a randomlyselected half of the total population. These surveys, which might have served to remindemployees of the M-Pasandaz product and their current status, had very little impact onrate switches by employees.

Instead, we find preliminary evidence that the default effect is driven by dynamic incon-sistency and the perceived complexity of making a financial decision. A core hypothesis inthe literature on savings defaults is that individuals fail to take full advantage of employer-sponsored savings plans because they procrastinate over the task of developing a financialplan and switching their automatically deducted savings allocation. We therefore perform anexperimental test of dynamic inconsistency over the actual task of changing a savings contri-bution. We ask subjects whether they are interested in receiving a call from the M-Pasandazresearch team and explain that “during the call, you will be able to ask questions, andunderstand how much savings you would have for different levels of monthly contribution.You will also have the opportunity change the level of your contribution if you would like.”Subjects are randomly assigned to either having the option of receiving the call immediately,


or receiving the call in one week’s time.2 Our objective is to perform a between-subject testof dynamic inconsistency where the non-monetary choice is the task of actually changinga contribution to a salary-linked savings account. We find that 70% of individuals are in-terested in receiving a free consultation today, while 76% of subjects accept the offer of aconsultation in a week’s time (p-value = 0.12). Because the offer is randomly assigned, weinterpret the difference in take-up as the differential degree of interest in undertaking thetask of changing the contribution amount. In other words, individuals prefer to put off thetask when they must do it immediately, but are less likely to put it off if they must do it inone week.

Our results have implications for policymakers and other stakeholders seeking to pro-mote financial inclusion in developing countries. Over the decade between 2001 and 2011,the share of the developing world’s workforce in the middle class ($4-13/day) or abovenearly doubled from 23% to 42%, dramatically expanding the number of regular wage earn-ers who might benefit from automatic payroll deductions to promote savings (ILO, 2013).Policymakers interested in increasing formal savings could achieve scalable impacts throughmandating minimum default contribution rates to savings accounts in the context of civilservant salaries or private pension plans. And while penetration by formal banks is oftenlimited in developing countries, the mobile phone-based product evaluated in this paper ispotentially applicable to a wide range of country contexts experiencing rapid growth in bothmobile network coverage and mobile financial systems.

The findings we present are thus consistent with the behavioral-economics view of poverty,which argues that poor exhibit the same biases and inconsistencies as their developed countrycounterparts (Bertrand et al., 2004). Taking this approach to its logical conclusion, we showthat one of the most effective means of increasing savings in a developed country context- automatic payroll deductions - also has broad relevance to the emerging middle class indeveloping countries. Furthermore, our findings suggest that behavioral mechanisms, andspecifically dynamic inconsistency, are a key reasons that defaults are so effective.

3.3 Background

Afghanistan

While Afghanistan remains one of the poorest countries in the world, it has experienceda sustained period of growth over the past decade from increased private investment, con-tracts and aid flows from international partners. In 2013, the World Bank estimated percapita income at $690 - about the global average for low-income countries - with almosttwo thirds of the population living above the national poverty line. While the Afghan laborforce is generally characterized by agricultural and small-scale trading activities, a small butgrowing middle class has recently emerged, particularly in major urban areas like Kabul.

2Our protocol is therefore in the spirit of Augenblick et al (forthcoming) in that we attempt to elicit ameasure of dynamic inconsistency using non-monetary choices.


In particular, the Afghan telecom sector has been a major source of new job creation since2002, directly and indirectly employing approximately 100,000 individuals, of which about20% are considered skilled workers (Altai, 2014).

Financial inclusion remains severely limited in Afghanistan, with only 10% of adultsreporting a formal financial account and 4% reporting any formal savings over the pastyear (Demigurc-Kunt et al 2015). The supply of banks is limited, with approximately 2.5bank branches per 100,000 adults, less than one-third of the South Asia average (World Bank2015). Afghan banks offer short-term savings accounts with a floating interest rate and long-term “fixed deposit” accounts with a fixed interest and term, though the reported rates oftenfall below an annual inflation rate of between 5-10%. A demand for savings exists though,with about 25% of Afghans reporting any savings in the previous year - primarily informallythrough cash or in-kind holdings - with the most common reasons including retirement,school fees, and saving for a farm of business (Demirguc-Kunt et al., 2015).3

Mobile Savings

The M-Pasandaz product studied in this paper is a mobile phone-based savings account,which is based on the underlying financial platform of mobile money. Mobile money usesa simple SMS technology to enable the exchange and storage of value over a basic mobilephone interface, complemented by a real-world network of agents providing “cash-in” depositand “cash-out” withdrawal services. As mobile phone penetration rates surge in developingcountries, mobile money has emerged as a leading financial instrument for the poor withmore than a 100 million active users using over 250 mobile money services in 89 countries(GSMA, 2015). Without the overhead costs involved in traditional brick-and-mortar bankinginstitutions, mobile money services offer minimal transaction fees for basic financial services.

While the principal driver of mobile money adoption has been sending purchasing poweracross distance at low transaction costs, mobile savings is becoming increasingly popular.As mobile operators typically lack banking licenses, they generally do not pay interest onmobile money account balances and must enforce maximum balance limits determined byregulators. Despite this fact, mobile money accounts are consistently used for cash storagein a survey of 47 mobile money services, over 50% of accounts reported positive balancesand over 40% reported balances above $10 (GSMA, 2015). And by partnering with banks,mobile operators have begun to offer mobile money-linked bank accounts offering interestpayments and other financial services. In December 2014, 26 mobile savings products existedin 22 countries, totaling almost 10 million dedicated mobile savings accounts (GSMA, 2015).

Blumenstock et al. (2014a) document frequent usage of mobile balances for cash storagein Afghanistan, but at the time of this study no mobile money-linked bank accounts wereoffered in the country. As described in detail below, the M-Pasandaz product was created as aspecial account on Roshan’s M-Paisa mobile money platform that allowed for employer directdeposits and employee withdrawls but no additional transactions. M-Pasandaz deposits do

3See Chipchase et al. (2013) for a recent ethnographic study of savings practices in Afghanistan.


not earn interest, but employer-provided matching contributions are available to workerswho satisfy terms of a commitment contract to avoid withdrawals during a specified period.

Automatic Payroll Deduction

Automatic payroll deductions are widely used to promote savings in developed countries(Beshears et al., 2009). The Social Security system in the United States is based on manda-tory contributions by current private sector workers and firms to pay the benefits of retiredindividuals, where worker contributions are automatically deducted from payrolls. Definedcontribution savings plans, such as the well-known “401(k)” accounts named after the sectionof U.S. tax code that regulates them, follow a similar structure in setting aside automaticdeductions from an employee’s payroll and additional employer contributions into an tax-advantaged savings account that matures at the age of retirement. There are also examples ofautomatic payroll deductions for savings in developing countires, such as publicly-mandatedpension (or “provident”) funds for private sector workers in India, Malaysia and elsewhere,which require fixed employee contributions from automatic payroll deductions and employermatching contributions. While Afghanistan does not currently mandate pension plans forprivate sector employers, several of the larger employers including telecoms and interna-tional NGOs voluntarily offered such programs. During the M-Pasandaz product design, ourresearch team documented several private pension and savings schemes currently active inAfghanistan with employee contribution rates between 5-10% of monthly salaries, employermatches of up to 100% of deposits, and vesting periods ranging from monthly to annual.

3.4 Experiment

Sample Population

Founded in 2003, Roshan Telecom is currently the largest mobile network operator in Afghan-istan. Roshan employs approximately 1,100 full-time staff, of which over 90% are Afghannationals and about 15% are women. All Afghan national employees of Roshan are currentlypaid their monthly wages via a mobile salary payment system program based on Roshan’sM-Paisa mobile money network. Our study population consists of 967 full-time Afghan na-tional employees of Roshan as of the June 2014 baseline survey who are still employed at thetime of product launch in January 2015. In all results below, we exclude a pilot group of 18employees involved in the design and implementation of the M-Pasandaz product, leaving uswith an experimental sample of 949 employees. Employees hold job titles such as Manager,Engineer, Security Guard and Janitor and are located in six major regional offices: Kabul,Kandahar, Mazar, Herat, Ghazni and Kunduz. In our sample, the mean monthly wage isapproximately $590 and the median monthly wage is $410, compared to an average monthlyincome per capita in Afghanistan of $57.50 (World Bank 2013).


Product Specifications

Working with the research team, Roshan created the M-Pasandaz mobile savings account tohelp encourage its employees to save.4 The M-Pasandaz account is automatically activatedand linked to the M-Paisa mobile money profile of each Roshan employee. Deposits intoM-Pasandaz can only be made by the Roshan HR department at the time of monthly wagedisbursements, with a minimum automatic contribution rate of 0% (e.g. no participation)and a maximum automatic contribution rate of 10% of the monthly wage. Employees canchange their automatic contribution rate at any time by calling the HR department, butchanges must be made by the 15th of each month - approximately one week prior to theregular salary disbursement - in order to take effect in the same month. Employees receivean SMS confirmation at the time of the automatic deposit into their M-Pasandaz accounts,and can check their account balance at any time.

Withdrawals from the M-Pasandaz account can be made at any time using a “TransferFunds” option on the application’s menu that allows the employee to move an unlimitedamount of their current M-Pasandaz balance into their regular M-Paisa account, from whichit can either be withdrawn into cash or used on the M-Paisa platform. However, employeesare informed that any withdrawals prior to the end of the 6-month commitment savingsperiod will forfeit their rights to any incentive payments, as described below.

Primary Experimental Treatments

As part of the research design, employees are randomly assigned into one of three M-Pasandaz“plans” that are distinguished only by their level of employer matching incentives: White (0%match), Blue (25% match) and Red (50% match). While employees are free to modify theircontribution rate, they cannot change their plan assignment. Employees are informed duringtraining that incentive matches will only be paid on the total balance in the M-Pasandazproduct at the end of the six-month period following the first automatic contribution. Anywithdrawals from the M-Pasandaz account prior to the end of the six-month period forfeitsrights to an incentive payment. Employees are also assigned to a default contribution rateof either 0% or 5%, which is cross-randomized with the incentive plans. Table 3.1 reportsbalance tests on a range of observable characteristics across all six resulting combinationsof the primary treatments. Employees are individually informed during training of theirincentive plan and default contribution rate, as well as the simple procedure for changingtheir automatic contribution rate by calling a Human Resources representative.

The primary experimental treatments of incentives and defaults are implemented simul-taneously and run for months 1 and 2 (January and February 2015) without any additionalinterventions. In interpreting the main effects of these treatments below, we will restrictattention to employee contribution behavior observed up until February 28, 2015.

4“Pasandaz” means savings in Dari, the most common language spoken in Afghanistan.


Secondary Experimental Treatments

We execute three additional experimental treatments. First, given concerns that our follow-up surveys might impact employee behavior (cf. Zwane et al., 2011), we restrict our follow-upsurveys beginning in month 3 (March 2015) to a randomly selected half of the total sample.

Separately, in month 4 (April 2015), we launch two interventions to better understandwhy the defaults are working by raising the salience and reduce the complexity of changingcontribution rates to the M-Pasandaz account. The first intervention entails sending infor-mation via SMS reminders to a random subset of all employees in English, Dari and Pashto.The message content identifies the employee’s current M-Pasandaz contribution rate and thephone number of the Human Resources representative for changing their rate.5 The secondintervention involves calling employees to offer customized consultations to answer questionsabout the M-Pasandaz product, estimate their payouts under different contribution rates,and allow them to change their contribution rate.6 Subjects are randomly assigned to eitherhaving the option of receiving the call immediately, or receiving the call in one week’s time,allowing us to have a between subject test for dynamic inconsistency in response rates.

Data Collection and Rollout

Data collection began with a face-to-face baseline survey of all employees in May and June2014 to collect information on demographics, past experience with the M-Paisa mobile salarypayment technology, and benchmark savings levels and practices. Roshan provided adminis-trative records on salary payments, M-Paisa mobile money account usage, and calling historydata used to identify social network ties among employees. Employees attended mandatorygroup trainings on the M-Pasandaz product during December 2014, and were individualinformed at that time of their default contribution rate and plan assignments. An openenrollment period to change initial contribution rates lasted from December 30, 2014 untilJanuary 15, 2015, and the first automatic contributions were made on January 21, 2015. Ineach month, employees had the option to change their contribution rate up until the 15thof that month, and could withdraw from the M-Pasandaz account at any time. We observeall contribution changes and all deposits and withdrawals on the M-Pasandaz system. Thesixth and final automatic contribution is scheduled for June 2015, with incentive paymentsdue in July 2015. Endline face-to-face surveys will take place in August 2015.

5Here is a sample message text in English: ”M-Pasandaz Reminder: Next payday, 5% of your salary willbe deposited in your M-Pasandaz account. If you want to change your contribution, call 079999-3708.”

6See Online Appendix B for a full consultation script.


3.5 Results

Effect of Automatic Enrollment

In Table 3.2, we estimate the impact of automatic enrollment on contribution behaviorin the M-Pasandaz savings account.7 As noted, the figures reflect contribution rates as ofFebruary 28, 2015 in order to isolate behavior prior to the onset of the secondary interventionsdescribed above. We find large default effects: employees who are automatically enrolledin a 5% contribution rate are 40 percentage points more likely to contribute to the accountthan employees with a default contribution rate of 0% (column 4). This participation effectis observed at all levels of matching incentives despite minimal transaction costs in choosinga non-default contribution rate (columns 1-3). On average, automatic enrollment is resultsin a 1.77 percentage point increase in contribution rate (column 8), though the effect appearsto be driven primarily by large increases in the 0% match and 25% match groups, with nosignificant increase under the 50% match group (columns 5-7). Figure 3.1 graphs the maincoefficients from Table 3.2, highlighting the large and significant differences in contributionbehavior between those defaulted out of the M-Pasandaz account with an initial contributionrate of 0% and those defaulted in with a rate of 5%.

Contribution Switching Behavior

Table 3.3 breaks down aggregate switching behavior of contribution rates through April15, 2015, the latest data available at the time of this draft. Approximately one-third ofemployees (n=327) changed their contribution rates during the period of open enrollmentbetween December trainings and the January 15, 2015 initial contribution deadline. Anadditional 2.32% of employees (n=22) changed their contribution immediately followingthe first automatic contributions to M-Pasandaz on January 21, 2015, implying a very lowshare of individuals who appear to have not understood their default contribution status.Figure 3.2 plots the frequency of contribution rate switches over time, and shows a sharpdecline in switches following the January payday.

Table 3.3 also summarizes the contribution switches associated with the secondary ex-perimental treatments described above. Less than half a percent (n=2) of the employeesrandomly selected to receive phone surveys changed their contributions immediately after-wards. The SMS treatment had a small effect, with 2.14% (n=5) of the employees randomlyselected to receive text message reminders changing their contributions immediately after-wards. Together, we interpret the phone survey and SMS reminder results as evidence ofvery modest “top of the mind” effects.8 By contrast, the consultation treatment resulted in a

7While data collection is currently ongoing, preliminary analysis of survey data suggest that M-Pasandazcontributions represent new savings that has been primarily reallocated from cash holdings and a possiblereduction in informal loans provided to friends and family. Initial results from comparing the self-reportedsavings behavior from May/June 2014 baseline survey and the March 2015 phone survey are presentedFigure A3.1, Table A3.2 and Table A3.3.

8Table A3.5 presents the regression estimates associated with the phone and SMS treatments.


major number of switches, with 11.34% (n=54) of those selected to be offered consultationsselecting to change their contribution rates immediately afterwards.9

Figure 3.3 plots M-Pasandaz contribution rates as of April 15, 2015, as broken out be-tween panels by the three matching incentive levels (0%, 25% and 50%) and within panelby the two default contribution rates (0% and 5%). This figure provides helpful context inassessing the over 50% of employees who remained at their default contribution levels fourmonths after the launch of M-Pasandaz. Over 90% of those defaulted out of M-Pasandazremained unenrolled in the 0% matching treatment, though this figure dropped to abouttwo-thirds in the 25% matching treatment and only 40% in the 50% matching treatment.By contrast, the number of those defaulted into a 5% contribution rate who remained at thatrate by month four stayed approximately constant at one-third across the three incentivematches.

Consultation Offers & Dynamic Inconsistency

We perform an experimental test of dynamic inconsistency over the task of changing a savingscontribution. Half of all employees are randomly selected to be offered a free consultationon their participation in the M-Pasandaz program which includes answering questions, re-viewing personalized scenarios of estimated savings at different monthly contribution levels,and providing an opportunity to change their contribution. Subjects are also randomly as-signed to either haf the option of receiving the call immediately, or receiving the call in oneweek’s time. This research design allows us to perform a between-subject test of dynamicinconsistency, where the non-monetary choice is the task of actually changing a contribution.

Table 3.4 presents the preliminary results from this experiment. We find that 70% ofindividuals are interested in receiving a consultation today, while 76% of subjects acceptthe offer of a consultation in a week’s time (column 1, p-value = 0.12). Because the offeris randomly assigned, we interepret the difference in take-up as the evidence that employeesare differentially more interested in postponing the task if asked to perform it today.10 Inshort, individuals prefer to put off the task of changing their contribution level when theymust do it immediately, but are less likely to put it off if they can do it one week later.

While we observe high interest in consultation offers among the subsample who had notswitched their default contribution levels, we note that their acceptance rates do not differbased on random assignment to the “today” treatment (Table 3.4, column 2). Instead, thedynamic inconsistency results appear to be driven by previous switchers, where those inthe today group are 17 percentage points less likely to accept a consultation offer for the

9Table A3.4 breaks down switches by default contribution and plan assignment. Consultation switchesare evenly balanced across all three plans and included more employees who were defaulted into the account.

10In results available on request, we control for the level of busyness employees report on either that sameday or next week (according to treatment status) and find no change in employee acceptances, suggestingour results are not driven by differential busyness on the same day. We also observe that employees in the“next week” group are more likely to be reached to complete a consultation conditional on agreeing to onethan those in the “today” group, suggesting the results are not driven by biased acceptance rates.


same day (column 3). Further splitting this subgroup into those who had increased theircontributions in the past (e.g. changed from 0% or 5% default to 10%) and those whodecreased their contributions (e.g. changed from a 5% default to 0%), we find that the latterexhibit the largest tendency to procrastinate over the consultation. All but two of the 71employees in this subgroup had exited the M-Pasandaz account with a contribution rateof 0% as of February 28, and the wedge between the “today” and “next week” treatmentsgrows to 30 percentage points here (column 5). Finally, in column six we test if these effectsare potentially driven by employees in the 0% matching plan, but instead find that effectsare largest in the subgroup with a 50% match. In further analysis available on request, weconfirm that the wedge against accepting a consultation today is greatest among those whohave lower salaries and have self-reported existing loans as a rationale for not participatingin the M-Pasandaz account. Together, this evidence suggests that the procrastination inaccepting consultation offers is driven by employees who face high matching incentives buthave chosen not to participate in the past due to perceived financial constraints.

3.6 Conclusion

We design and experimentally evaluate a mobile phone-based savings account that allowssavings to be automatically deducted from salaries in Afghanistan, a country with extremelylow levels of formal financial inclusion. We find evidence of large default effects: employ-ees who are automatically enrolled with a 5% contribution rate into the account are 40percentage points more likely to contribute to the account than individuals with a defaultcontribution rate of zero. We also study the effects of randomly assigned matching contri-butions, and find that the effect of automatic enrollment on participation is approximatelyequivalent to providing financial incentives equal to a 50 percent match. To understand whydefault enrollment increases participation, we randomly offer some employees an immediatefinancial consultation, and others a financial consultation in one week. Our analysis indi-cates that employees are more likely to discuss changing the savings contributions in oneweek, suggesting that defaults raise contributions due to perceived complexity of financialdecisions, and because employees procrastinate in developing a financial plan for the future.

A growing literature has documented the importance of automatic payroll deductionsin increasing savings in developed countries. Our experimental results suggest that thisapproach could have major implications for policymakers and other stakeholders seekingto promote financial inclusion in developing countries. Over the decade between 2001 and2011, the share of the developing world’s workforce in the middle class ($4-13/day) or abovenearly doubled from 23% to 42% (ILO, 2013). As a consequence, the number of regularwage earners who might benefit from automatic payroll deductions in poor countries hasbeen rapidly expanding. Our analysis suggest that policymakers interested in increasingformal savings among this emerging middle class could achieve scalable impacts throughmandating minimum default contribution rates to savings accounts in the context of civilservant salaries or private pension plans.



Figure 3.1: Contribution Behavior (as of Feb 28, 2015)

0.2

.4.6

.81

Parti

cipa

tes (

=1)

0 10 20 30 40 50Match Rate (%)

Participates (=1)

02

46

8C

ontri

butio

n R

ate

(% o

f Sal

ary)

0 10 20 30 40 50Match Rate (%)

Contribution Rate (%)

Default Out Default In +/- 2 se

Notes: Participates (=1) is a binary variable that equals one if the contribution rate is greater than zero.Variables reflect contribution rate values observed as of February 28, 2015, following the first two paydaysbut prior to the rollout of phone surveys or secondary interventions.


Table 3.1: Summary Statistics

Default Out Default In

0% 25% 50% 0% 25% 50% P-ValueAll Match Match Match Match Match Match of F-Test

Gender (Male = 1) 0.85 0.85 0.87 0.85 0.84 0.81 0.88 0.590.36 0.36 0.33 0.36 0.37 0.39 0.33

Head of Household (=1) 0.52 0.43 0.56 0.58 0.51 0.52 0.54 0.110.50 0.50 0.50 0.50 0.50 0.50 0.50

Married (=1) 0.64 0.66 0.64 0.62 0.66 0.64 0.65 0.980.48 0.47 0.48 0.49 0.48 0.48 0.48

Age 30.82 30.30 30.87 31.30 30.87 29.98 31.61 0.609.27 7.51 9.76 10.48 9.31 7.63 10.54

Hazara (=1) 0.46 0.42 0.43 0.44 0.49 0.50 0.48 0.730.50 0.50 0.50 0.50 0.50 0.50 0.50

Monthly Salary (1000 Afs) 32.43 30.41 31.20 33.86 34.39 31.72 33.04 0.8430.79 25.01 24.12 38.68 34.84 26.25 33.27

Monthly Savings (1000 Afs) 21.09 13.05 45.47 12.01 17.10 12.47 26.79 0.33127.55 28.84 256.84 26.39 35.96 25.34 168.68

Has Bank Account (=1) 0.41 0.42 0.39 0.38 0.41 0.44 0.40 0.880.49 0.49 0.49 0.49 0.49 0.50 0.49

Delayed a Bill Payment (=1) 0.41 0.43 0.36 0.47 0.41 0.37 0.42 0.400.49 0.50 0.48 0.50 0.49 0.48 0.50

Withdraws Entire Salary (=1) 0.41 0.37 0.42 0.42 0.41 0.44 0.40 0.820.49 0.48 0.49 0.50 0.49 0.50 0.49

Interested in M-Pasandaz (=1) 0.85 0.85 0.87 0.84 0.83 0.89 0.84 0.630.35 0.36 0.33 0.37 0.38 0.31 0.37

Expects Violence (=1) 0.57 0.56 0.57 0.56 0.57 0.56 0.58 1.000.50 0.50 0.50 0.50 0.50 0.50 0.49

Experienced Violence (=1) 0.45 0.51 0.47 0.50 0.40 0.39 0.42 0.260.50 0.50 0.50 0.50 0.49 0.49 0.50

Observations 949 161 158 159 158 158 155


Figure 3.2: Calls to change contribution rate, over time (as of April 15, 2015)

Notes: Dots indicate the number of individuals calling in, on a given day, to change their contribution rate.


Figure 3.3: M-Pasandaz Contribution Rates (as of April 15, 2015)

Notes: Distribution of current M-Pasandaz contribution levels, as a percentage of monthly salary. Individualswere randomized into either a default 0% contribution (peach bars, N=478) or a default 5% contribution(green bars, N=471). Individuals were further randomized into three different incentive rates: White (0%match, N=319), Blue (25% match, N=316) and Red (50% match, N=314). Semi-transparent bars indicatethe original assigned contribution rate, solid bars indicate current contribution rate.

CHAPTER 3. AUTOMATIC PAYROLL DEDUCTIONS IN AFGHANISTAN 98T

able

3.2:

The

Eff

ect

ofA

uto

mat

icE

nro

llm

ent

Dep

end

ent

Var

iab

le:

Par

tici

pat

es(=

1)C

ontr

ibu

tion

(Per

cent

ofS

alar

y)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Def

ault

In(=

1)0.

47**

*0.

44**

*0.

30**

*0.

40**

*2.

39**

*2.

24**

*0.

671.

77**

*(0

.04)

(0.0

5)(0

.05)

(0.0

3)(0

.21)

(0.4

5)(0

.48)

(0.2

6)C

onst

ant

0.01

0.27

***

0.56

***

0.28

***

0.03

2.56

***

5.44

***

2.67

***

(0.0

1)(0

.04)

(0.0

4)(0

.02)

(0.0

3)(0

.34)

(0.3

9)(0

.20)

Sam

ple

0%M

atch

25%

Mat

ch50

%M

atch

Com

ple

te0%

Mat

ch25

%M

atch

50%

Mat

chC

omp

lete

#O

bse

rvat

ion

s31

931

631

494

931

931

631

494

9R

-Squ

ared

0.30

20.

191

0.10

70.

161

0.29

10.

072

0.00

60.

046

Notes:

Part

icip

ates

(=1)

isa

bin

ary

vari

able

that

equal

sone

ifth

eco

ntr

ibuti

onra

teis

grea

ter

than

zero

.V

ari

able

sre

flec

tco

ntr

ibuti

onra

teva

lues

obse

rved

asof

Feb

ruary

28,

2015

,fo

llow

ing

the

firs

ttw

opay

day

sbut

pri

orto

the

rollout

ofphone

surv

eys

or

seco

ndary

inte

rven

tions.

*p<

0.1,

**p<

0.0

5,

***p<

0.01.

Rob

ust

standard

erro

rsre

port

edin

pare

nth

eses

.


Table 3.3: Switches Breakdown (as of April 15, 2015)

Frequency Total Percentage

Changed In Open Enrollment 327 949 34.46

Changed After 1st Payday 22 949 2.32

Changed After Other Payday 1 949 0.11

Changed After Survey 2 470 0.43

Changed After SMS 5 234 2.14

Changed After Consultation 54 476 11.34

Changed More Than Once 5 949 0.53

Never Changed Contribution 496 949 52.27

Notes: Total in column 2 reports number of participants that were treated.Payday, Survey, SMS and Consultation switches are recorded if correspondingto the day of the intervention or the day immediately afterwards.


able

3.4:

Con

sult

atio

nO

ffer

Res

ult

s

Dep

end

ent

Var

iab

le:

Acc

epte

dC

onsu

ltat

ion

Acc

epte

dC

onsu

ltat

ion

Acc

epte

dC

onsu

ltat

ion

Off

er(=

1)O

ffer

(=1)

Off

er(=

1)(1

)(2

)(3

)(4

)(5

)(6

)

Con

sult

Tod

ay(=

1)-0

.07

0.01

-0.1

7***

-0.0

9-0

.30*

*(0

.04)

(0.0

6)(0

.06)

(0.0

6)(0

.11)

Con

sult

Tod

ay*

50%

Mat

ch-0

.54*

(0.3

1)C

onsu

ltT

od

ay*

25%

Mat

ch-0

.21

(0.2

3)C

onsu

ltT

od

ay*

0%M

atch

-0.2

3(0

.16)

Mat

chR

ate

=50

%0.

38**

(0.1

7)M

atch

Rat

e=

25%

0.17

(0.2

1)C

onst

ant

0.76

***

0.73

***

0.82

***

0.95

***

0.63

***

0.50

***

(0.0

3)(0

.04)

(0.0

4)(0

.03)

(0.0

8)(0

.12)

Sam

ple

Com

ple

teN

ever

Sw

itch

edS

wit

ched

Sw

itch

edU

pS

wit

ched

Dow

nS

wit

ched

Dow

n#

Ob

serv

atio

ns

444

262

182

111

7171

R-S

qu

ared

0.00

50.

000

0.03

70.

022

0.08

70.

146

Notes:

Acc

epte

dC

onsu

ltat

ion

Off

er(=

1)

isa

bin

ary

vari

able

that

equals

one

ifth

eem

plo

yee

agr

eed

topart

icip

atio

nin

afinan

cial

consu

ltat

ion

regar

din

gth

eir

part

icip

atio

nin

the

M-P

asan

daz

pro

gra

m(s

eepap

erte

xt

for

det

ails)

.C

onsu

ltT

oday

(=1)

isa

bin

ary

vari

able

that

equal

sone

ifth

eem

plo

yee

was

random

lyas

sign

edto

rece

ive

aco

nsu

ltat

ion

onth

esa

me

day

as

the

consu

ltati

onoff

erw

asm

ade,

and

equals

zero

ifth

eco

nsu

ltati

onw

asas

sign

edto

take

pla

ceon

ew

eek

late

r.(N

ever

)Sw

itch

edis

the

sam

ple

of

emplo

yees

who

had

(not)

changed

thei

rdef

ault

contr

ibuti

onra

tes

asof

Feb

ruar

y28

,201

5.Sw

itch

edU

p(D

own)

isth

esa

mple

of

emplo

yees

who

had

incr

ease

d(d

ecre

ased

)th

eir

contr

ibuti

onra

tes

asof

Feb

ruary

28,

2015

.*p<

0.1,

**p<

0.05,

***p<

0.0

1.

Rob

ust

standard

erro

rsre

port

edin

pare

nth

eses

.



Consultation Scripts Appendix

Offering Consultation Script

Treatment A [Consultation Today]

Hello XXX. I am calling on behalf of the M-Pasandaz research team. As you know, Roshanrecently began to offer a new savings account to all Roshan employees, called M-Pasandaz.The reason for this call is that the M-Pasandaz research team would like to provide em-ployees with the opportunity to learn more about M-Pasandaz, and determine how to useM-Pasandaz in the way that is best for you.

If you would like, someone from the M-Pasandaz research team will give you a phonecall later today, to help you determine if it makes sense for you to make contributions toM-Pasandaz. Each employee is in a different financial situation, so the M-Pasandaz researchteam would like to give you information to help you decide if and how to use M-Pasandaz.During the call, you will be able to ask questions, and understand how much savings youwould have for different levels of monthly contribution. During the call, you will also havethe opportunity change the level of your contribution if you would like.

Would you be interested in us giving you a call? If so, a representative will give you aphone call later today.

Can I also ask, on a scale of 1 to 10, where 1 is not at all busy and 10 is extremely busy,how busy are you later today? [Record Response]

Treatment B [Consultation in One Week]

Hello XXX. I am calling on behalf of the M-Pasandaz research team. As you know, Roshanrecently began to offer a new savings account to all Roshan employees, called M-Pasandaz.The reason for this call is that the M-Pasandaz research team would like to provide em-ployees with the opportunity to learn more about M-Pasandaz, and determine how to useM-Pasandaz in the way that is best for you.

If you would like, someone from the M-Pasandaz research team will give you a phonecall in one week, to help you determine if it makes sense for you to make contributions toM-Pasandaz. Each employee is in a different financial situation, so Human Resources wouldlike to give you information to help you decide if and how to use M-Pasandaz. During thecall, you will be able to ask questions, and understand how much savings you would have fordifferent levels of monthly contribution. During the call, you will also have the opportunitychange the level of your contribution if you would like.

Would you be interested in us giving you a call? If so, a representative will give you aphone call in one week.


Can I also ask, on a scale of 1 to 10, where 1 is not at all busy and 10 is extremely busy,how busy are you in one week? [Record Response]

M-Pazandaz Consultation Script

Hello XXX. I am calling on behalf of the M-Pasandaz research team department. I am callingbecause you recently requested that a representative call you to provide you with additionalinformation about M-Pasandaz, and determine how to use M-Pasandaz in the way that isbest for you. This consultation will last roughly 5-10 minutes. Are you able to speak to menow? [RECORD RESPONSE]

Thank you for taking the time to speak with me. As you know, M-Pasandaz is a newbenefit that is being offered to Roshan employees. In this call, you will have the opportunityto ask questions about M-Pasandaz. I will provide information about how much savings youwould have for different levels of monthly contribution. At the end of the call, you will alsohave the opportunity to change the level of your contribution if you would like.

First of all, would you like me to give you a brief overview of the M-Pasandaz account?[YES/NO]

If YES: M-Pasandaz is a new benefit for all Roshan employees that was designed to helpincrease your savings. It is a mobile savings account that is linked to your M-Paisa account. Aportion of your monthly salary - up to a maximum of 10% - can be automatically depositedinto your M-Pasandaz account each month. Participating in the M-Pasandaz account isvoluntary and you may receive benefits from Roshan to encourage you to save for the future.You can access the money in your M-Pasandaz account at any time, but if you contributeand dont make any withdrawals for 6 months, you may be eligible for a bonus from Roshanas a reward for savings.

To begin, we would like to ask if there are any questions we might answer about M-Pasandaz. [YES/NO]

Now, since every person has a different situation, I would like to explain several differentscenarios, to help you understand how different levels of M-Pasandaz contributions wouldwork for you. According to our records, you are in the [WHITE/BLUE/RED] plan, and youcurrently have a monthly contribution rate of [XX%]. Were you aware that this was yourplan and contribution rate? [YES/NO]

According to our records, you have a monthly salary of XXX. Since you are in the[WHITE/BLUE/RED] plan, you are eligible to receive a matching contribution Roshan of[0/25/50] percent for all money that you save in your M-Pasandaz account. Our records alsoshow that you [HAVE/HAVE NOT] made a withdrawal from your M-Pasandaz account,meaning that you [ARE NOT/ARE] still eligible to receive your matching contribution.Therefore, if you continue to contribute at your current rate and make no withdrawals, atthe end of the trial period in July, you would have a total value of MMM in your M-Pasandazaccount. This reflects both your contribution and the contribution of Roshan to the accounton your behalf. Would you like me to repeat this information for you? [YES/NO]


Thank you. Of course, you are always free to change your monthly contribution rate.If you like, I can explain to you exactly what would happen if you decided to change yourmatch to a different amount. Would you like me assist you by explaining what would happenif you changed your contribution rate to a different amount? [YES/NO]

If YES: What scenario would you like me to explain? The contribution rate can beanywhere between 0% and 10% of your monthly salary. [RECORD ANSWER]

Do you have any additional questions about how M-Pasandaz works, or can I provideany additional information that can help you determine how to use M-Pasandaz in the waythat is best for you? [YES/NO]

Thank you. Now, I would like to offer you the opportunity to change your contribu-tion rate. If you wish, you can tell me your preferred rate, and I will change it for you.Alternatively, you always have the opportunity to call HR at a later date and change thecontribution. Would you like me to change your contribution rate? [YES/NO]

If YES: What would you like your new rate to be: [RECORD RESPONSE]Thank you very much for your time. Goodbye.



Figure A3.1: Savings Behavior (as of March 2015)

050

0010

000

1500

0A

fgha

nis

1 2Survey Wave

Default Out

050

0010

000

1500

0A

fgha

nis

1 2Survey Wave

Default In

bank_savings cash_savingsloans_given mpaisa_savingsmpaz_savings

Notes: Average self-reported values for all categories of savings, excluding top .5% of outliers in total savings.Survey wave 1 corresponds to baseline conducted in May and June 2014, and survey wave 2 corresponds tofollow-up survey conducted in March 2015.


able

A3.

1:T

he

Eff

ect

ofE

mplo

yer

Mat

ches

Dep

enden

tV

aria

ble

:P

arti

cipat

es(=

1)C

ontr

ibuti

on(P

erce

nt

ofSal

ary)

(1)

(2)

(3)

(4)

(5)

(6)

Mat

chR

ate

=25

%0.

26**

*0.

23**

*0.

25**

*2.

53**

*2.

39**

*2.

47**

*(0

.04)

(0.0

5)(0

.04)

(0.3

5)(0

.36)

(0.2

7)M

atch

Rat

e=

50%

0.55

***

0.38

***

0.47

***

5.41

***

3.69

***

4.56

***

(0.0

4)(0

.05)

(0.0

4)(0

.39)

(0.3

4)(0

.27)

Con

stan

t0.

010.

47**

*0.

24**

*0.

032.

42**

*1.

21**

*(0

.01)

(0.0

4)(0

.02)

(0.0

3)(0

.21)

(0.1

2)Sam

ple

Def

ault

Out

Def

ault

InC

omple

teD

efau

ltO

ut

Def

ault

InC

omple

te#

Obse

rvat

ions

478

471

949

478

471

949

R-S

quar

ed0.

257

0.11

30.

147

0.25

80.

179

0.20

6

Notes:

Par

tici

pat

es(=

1)is

ab

inar

yva

riab

leth

at

equ

als

on

eif

the

contr

ibu

tion

rate

isgre

ate

rth

an

zero

.V

ari

ab

les

refl

ect

contr

ibu

tion

rate

valu

esob

serv

edas

ofF

ebru

ary

28,

2015,

foll

owin

gth

efi

rst

two

pay

day

sb

ut

pri

or

toth

ero

llout

of

ph

on

esu

rvey

sor

seco

nd

ary

inte

rven

tion

s.*p<

0.1,

**p<

0.05,

***p<

0.01.

Rob

ust

stan

dard

erro

rsre

port

edin

pare

nth

eses

.


able

A3.

2:N

ewSav

ings

Dep

enden

tV

aria

ble

:T

otal

Sav

ings

(Afs

.)T

otal

Sav

ings

(Afs

.)M

-Pas

andaz

Sav

ings

(Afs

.)T

otal

Sav

ings

(Afs

.)O

LS

OL

SO

LS

IV(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)

M-P

asan

daz

Sav

ings

(Afs

.)2.

88*

2.44

0.49

1.00

(1.6

2)(1

.49)

(1.6

1)(1

.32)

Def

ault

In*

Pos

t60

76.4

960

56.2

796

2.96

***

962.

96**

*(4

279.

78)

(380

4.81

)(3

69.0

7)(3

69.0

5)D

efau

ltIn

(=1)

-198

5.03

0.00

***

(239

7.92

)(0

.00)

Pos

t(=

1)-3

752.

28-2

989.

6994

5.36

***

945.

36**

*(2

673.

39)

(235

2.61

)(1

94.4

9)(1

94.4

8)C

onst

ant

1013

4.25

***

1042

7.97

***

1334

0.97

***

1200

5.44

***

-0.0

0***

0.00

1171

4.79

***

(113

4.74

)(9

86.5

5)(1

926.

41)

(900

.62)

(0.0

0)(8

5.09

)(1

169.

95)

#O

bse

rvat

ions

876

876

890

890

876

876

876

820

R-S

quar

ed0.

059

0.06

20.

003

0.00

60.

087

0.14

10.

018

0.04

0E

mplo

yee

FE

No

Yes

No

Yes

No

Yes

No

Yes

Notes:

Sel

f-re

por

ted

dat

afr

om

bas

elin

eco

nd

uct

edin

May

an

dJu

ne

2014

and

foll

ow-u

psu

rvey

con

du

cted

inM

arc

h20

15.

Tri

mm

ing

top

.5%

ofou

tlie

rsin

tota

lsa

vin

gs.

*p<

0.1

,**p<

0.05

,**

*p<

0.01

.R

obu

stst

and

ard

erro

rsre

por

ted

inp

aren

thes

es.


able

A3.

3:Sav

ings

Subst

ituti

on

Dep

enden

tV

aria

ble

:C

ash

onH

and

(Afs

.)L

oans

Giv

en(A

fs.)

Food

Exp

endit

ure

Gen

eral

Exp

endit

ure

sL

ast

Wee

k(A

fs.)

Las

tM

onth

(Afs

.)(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)

M-P

asan

daz

Sav

ings

(Afs

.)-0

.16*

**-0

.37*

*-0

.05

-0.1

1*0.

090.

020.

61-0

.45

(0.0

4)(0

.18)

(0.0

5)(0

.06)

(0.0

6)(0

.06)

(0.6

9)(0

.68)

Con

stan

t28

17.7

1***

2959

.31*

**12

20.7

9***

1254

.67*

**40

29.6

8***

4074

.55*

**19

337.

68**

*20

048.

66**

*(2

61.7

0)(1

22.7

1)(2

00.2

2)(3

6.28

)(1

60.0

3)(3

8.18

)(1

186.

09)

(458

.59)

#O

bse

rvat

ions

875

875

875

875

854

854

873

873

R-S

quar

ed0.

003

0.02

40.

001

0.00

30.

002

0.00

00.

002

0.00

2E

mplo

yee

FE

No

Yes

No

Yes

No

Yes

No

Yes

Notes:

Sel

f-re

por

ted

dat

afr

om

bas

elin

eco

nd

uct

edin

May

an

dJu

ne

2014

and

foll

ow-u

psu

rvey

con

du

cted

inM

arch

2015

.T

rim

min

gto

p.5

%of

ou

tlie

rsin

each

dep

end

ent

vari

able

.*p<

0.1,

**p<

0.05

,**

*p<

0.0

1.R

obu

stst

an

dar

der

rors

rep

orte

din

par

enth

eses

.


able

A3.

4:Sw

itch

esB

reak

dow

nby

Def

ault

and

Pla

nSta

tus

(as

ofA

pri

l15

,20

15)

Def

ault

Out

Def

ault

In

Fre

quen

cyT

otal

Per

centa

ge0%

25%

50%

0%25

%50

%

Chan

ged

InO

pen

Enro

llm

ent

327

949

34.4

60

3268

7680

71

Chan

ged

Aft

er1s

tP

ayday

2294

92.

320

74

36

2

Chan

ged

Aft

erO

ther

Pay

day

194

90.

110

10

00

0

Chan

ged

Aft

erSurv

ey2

470

0.43

00

01

10

Chan

ged

Aft

erSM

S5

234

2.14

00

12

20

Chan

ged

Aft

erC

onsu

ltat

ion

5447

611

.34

107

68

1112

Chan

ged

Mor

eT

han

Once

594

90.

530

01

13

0

Nev

erC

han

ged

Con

trib

uti

on49

694

952

.27

151

109

6458

5361

Notes:

Tot

alin

colu

mn

2re

por

tsnum

ber

of

part

icip

ants

that

wer

etr

eate

d.

Pay

day

,S

urv

ey,

SM

San

dC

on

sult

ati

on

swit

ches

are

reco

rded

ifco

rres

pon

din

gto

the

day

of

the

inte

rven

tion

or

the

day

imm

edia

tely

aft

erw

ard

s


Tab

leA

3.5:

“Top

ofth

eM

ind”

Tre

atm

ents

Dep

enden

tV

aria

ble

:C

han

ged

Aft

erP

hon

eSurv

ey(=

1)C

han

ged

Aft

erSM

SR

emin

der

(=1)

(1)

(2)

(3)

(4)

(5)

(6)

Phon

eSurv

ey(=

1)0.

004

0.00

00.

006

(0.0

03)

(0.0

00)

(0.0

06)

Def

ault

*P

hon

eSurv

ey0.

009

(0.0

06)

25%

Mat

ch*

Phon

eSurv

ey0.

000

(0.0

09)

50%

Mat

ch*

Phon

eSurv

ey-0

.006

(0.0

06)

SM

SR

emin

der

(=1)

0.02

1**

0.00

80.

026

(0.0

09)

(0.0

08)

(0.0

18)

Def

ault

*SM

SR

emin

der

0.02

6(0

.019

)25

%M

atch

*SM

SR

emin

der

-0.0

01(0

.025

)50

%M

atch

*SM

SR

emin

der

-0.0

13(0

.022

)D

efau

ltIn

(=1)

0.00

0-0

.000

(0.0

00)

(0.0

00)

Mat

chR

ate

=25

%0.

000

-0.0

00(0

.000

)(0

.000

)M

atch

Rat

e=

50%

0.00

0-0

.000

***

(0.0

00)

(0.0

00)

Con

stan

t-0

.000

**-0

.000

-0.0

000.

000*

**0.

000

0.00

0***

(0.0

00)

(0.0

00)

(0.0

00)

(0.0

00)

(0.0

00)

(0.0

00)

#O

bse

rvat

ions

949

949

949

473

473

473

R-S

quar

ed0.

002

0.00

60.

004

0.01

10.

019

0.01

3

Notes:

Ch

ange

dA

fter

Ph

one

Su

rvey

isa

bin

ary

vari

able

that

equ

als

one

ifan

emp

loye

ech

an

ged

thei

rco

ntr

ibu

tion

rate

eith

eron

the

day

they

rece

ived

ap

hon

esu

rvey

orth

ed

ayim

med

iate

lyfo

llow

ing.

Ch

ange

dA

fter

SM

SR

emin

der

isa

bin

ary

vari

able

that

equ

als

on

eif

anem

plo

yee

chan

ged

thei

rco

ntr

ibu

tion

rate

eith

eron

the

day

they

rece

ived

asm

sre

min

der

orth

ed

ayim

med

iate

lyfo

llow

ing.

Ph

one

Su

rvey

isa

bin

ary

vari

able

ifth

eem

plo

yee

was

ran

dom

lyas

sign

edto

rece

ive

ap

hon

esu

rvey

.S

MS

Rem

ind

eris

ab

inary

vari

able

ifth

eem

plo

yee

was

ran

dom

lyass

ign

edto

rece

ive

ansm

sre

min

der

.*p<

0.1,

**p<

0.05

,**

*p<

0.01

.R

obu

stst

an

dar

der

rors

rep

ort

edin

par

enth

eses

.

110

Bibliography

Aker, Jenny C, “Information from markets near and far: Mobile phones and agriculturalmarkets in Niger,” American Economic Journal: Applied Economics, 2010, 2 (3), 46–59.

, Rachid Boumnijel, Amanda McClelland, and Niall Tierney, “Zap It to Me: TheShort-Term Impacts of a Mobile Cash Transfer Program,” Center for Global Developmentworking paper, 2011, 268.

Altai, “ICT Economic Impact Assessment,” Technical Report 2014.

Antras, Pol and C Fritz Foley, “Poultry in Motion: A Study of International TradeFinance Practices,” Journal of Political Economy, forthcoming, 2014.

Ashraf, N., D. Karlan, and W. Yin, “Tying Odysseus to the mast: Evidence from acommitment savings product in the Philippines,” The Quarterly Journal of Economics,2006, 121 (2), 635–672.

Asker, John, Allan Collard-Wexler, and Jan De Loecker, “Dynamic Inputs andResource (Mis) Allocation,” Journal of Political Economy, forthcoming, 2013.

Baker, George, Robert Gibbons, and Kevin J Murphy, “Relational Contracts andthe Theory of the Firm,” Quarterly Journal of Economics, 2002, pp. 39–84.

Bank, World, “Sierra Leone: Adding Value through Trade for Poverty Reduction,” Tech-nical Report 2006.

, “Sierra Leone: Second Poverty Reduction Strategy Paper,” Technical Report 2009.

Barron, Daniel Vincent, “Essays in Cooperation and Repeated Games.” PhD disserta-tion, Massachusetts Institute of Technology 2013.

Bellows, John and Edward Miguel, “War and institutions: New evidence from SierraLeone,” American Economic Review, 2006, pp. 394–399.

Benartzi, Shlomo and Richard H. Thaler, “Heuristics and Biases in Retirement SavingsBehavior,” The Journal of Economic Perspectives, July 2007, 21 (3), 81–104.

BIBLIOGRAPHY 111

Bertrand, Marianne, Sendhil Mullainathan, and Eldar Shafir, “A Behavioral-Economics View of Poverty,” The American Economic Review, May 2004, 94 (2), 419–423.

Beshears, John, James Choi, David Laibson, and Brigitte Madrian, “The limita-tions of defaults,” Technical Report, National Bureau of Economic Research 2010.

, James J. Choi, David Laibson, and Brigitte C. Madrian, “The importance ofdefault options for retirement saving outcomes: Evidence from the United States,” in“Social security policy in a changing environment,” University of Chicago Press, 2009,pp. 167–195.

Besley, Timothy and Hannes Mueller, “Estimating the Peace Dividend: Evidence fromHouse Prices in Northern Ireland,” American Economic Review, 2012, 102 (2), 810–833.

Bigsten, Arne, Paul Collier, Stefan Dercon, Marcel Fafchamps, Bernard Gau-thier, Jan Willem Gunning, Abena Oduro, Remco Oostendorp, Cathy Patillo,and Mans Soderbom, “Contract Flexibility and Dispute Resolution in African Manu-facturing,” The Journal of Development Studies, 2000, 36 (4), 1–37.

, , , , , , , , , and , “Credit Constraints in Manufacturing Enterprisesin Africa,” Journal of African Economies, 2003, 12 (1), 104–125.

Blattman, Christopher and Edward Miguel, “Civil war,” Journal of Economic Liter-ature, 2010, pp. 3–57.

Blumenstock, Joshua Evan, “Inferring Patterns of Internal Migration from Mobile PhoneCall Records: Evidence from Rwanda,” Information Technology for Development, 2012,18 (2), 107–125.

Blumenstock, Joshua, Michael Callen, and Tarek Ghani, “Violence and FinancialDecisions: Evidence from Mobile Money in Afghanistan,” 2014.

, , and , “Mobile-izing Savings: Experimental Evidence on the Impact of AutomaticContributions in Afghanistan,” 2015.

, Nathan Eagle, and Marcel Fafchamps, “Risk and Reciprocity on the Mobile PhoneNetwork: Evidence from Rwanda,” 2014.

Board, Simon, “Relational Contracts and the Value of Loyalty,” The American EconomicReview, 2011, pp. 3349–3367.

Callen, Michael, Mohammad Isaqzadeh, James D. Long, and Charles Sprenger,“Violence and Risk Preference: Experimental Evidence from Afghanistan,” AmericanEconomic Review, 2014, 104 (1), 123–48.

Cameron, A Colin and Douglas L Miller, “A Practitioner’s Guide to Cluster-RobustInference,” Forthcoming in Journal of Human Resources, 2013.

BIBLIOGRAPHY 112

, Jonah B Gelbach, and Douglas L Miller, “Bootstrap-based improvements for in-ference with clustered errors,” The Review of Economics and Statistics, 2008, 90 (3),414–427.

Casaburi, Lorenzo and Tristan Reed, “Interlinked Transactions and Pass-Through:Experimental Evidence from Sierra Leone,” mimeo Stanford, 2014.

Chetty, Raj, John N. Friedman, Søren Leth-Petersen, Torben Heien Nielsen,and Tore Olsen, “Active vs. Passive Decisions and Crowd-Out in Retirement SavingsAccounts: Evidence from Denmark,” The Quarterly Journal of Economics, August 2014,129 (3), 1141–1219.

Chipchase, Jan, Mark Roston, Cara Silver, and Joshua Blumenstock, “In TheHands of God: A Study of Risk and Savings in Afghanistan,” Technical Report 2013.

, Panthea Lee et al., “Mobile Money: Afghanistan,” innovations, 2011, 6 (2), 13–33.

Choi, James J., David Laibson, Brigitte C. Madrian, and Andrew Metrick, “Forbetter or for worse: Default effects and 401 (k) savings behavior,” in “Perspectives on theEconomics of Aging,” University of Chicago Press, 2004, pp. 81–126.

Collier, Paul and Jan Willem Gunning, “Why has Africa Grown Slowly?,” The Journalof Economic Perspectives, 1999, pp. 3–22.

and Marguerite Duponchel, “The Economic Legacy of Civil War Firm-level: Evidencefrom Sierra Leone,” Journal of Conflict Resolution, 2013, 57 (1), 65–88.

Davis, Donald R. and David E. Weinstein, “Bones, Bombs, and Break Points: TheGeography of Economic Activity,” American Economic Review, 2002, 92 (5), 1269–1289.

Delavande, Adeline and Hans-Peter Kohler, “Subjective expectations in the contextof HIV/AIDS in Malawi,” Demographic Research, 2009, 20 (31), 817–875.

, Xavier Gine, and David McKenzie, “Measuring subjective expectations in develop-ing countries: A critical review and new evidence,” Journal of Development Economics,March 2011, 94 (2), 151–163.

Demirguc-Kunt, Asli and Leora Klapper, “Measuring Financial Inclusion: The GlobalFindex Database,” Technical Report, World Bank 2012.

, , Dorothe Singer, and Peter Van Oudheusden, “The Global Findex Database2014: Measuring Financial Inclusion around the World,” Technical Report, World Bank,Washington, DC 2015.

Dermish, Ahmed, Christoph Kneiding, Paul Leishman, and Ignacio Mas, “Branch-less and mobile banking solutions for the poor: a survey of the literature,” innovations,2011, 6 (4), 81–98.

BIBLIOGRAPHY 113

Dixit, Avinash K, Lawlessness and Economics: Alternative Modes of Governance, Prince-ton University Press, 2007.

Dupas, Pascaline and Jonathan Robinson, “Why Don’t the Poor Save More? Evidencefrom Health Savings Experiments,” American Economic Review, 2013, 103 (4), 1138–71.

Fafchamps, Marcel, “Trade Credit in Zimbabwean Manufacturing,” World Development,1997, 25 (5), 795–815.

, “Market institutions in sub-Saharan Africa: Theory and evidence,” MIT Press Books,2004, 1.

, Jan Willem Gunning, and Remco Oostendorp, “Inventories and Risk in AfricanManufacturing,” The Economic Journal, 2000, 110 (466), 861–893.

Filkins, Dexter, “Letter from Kabul: The Great Afghan Bank Heist,” The New Yorker,2011, 31.

Ghani, Tarek and Tristan Reed, “Competing for Relationships: Markets and InformalInstitutions in Sierra Leone,” 2015.

Greif, Avner, “Reputation and Coalitions in Medieval Trade: Evidence on the MaghribiTraders,” Journal of Economic History, 1989, 49 (04), 857–882.

, “Contract Enforceability and Economic Institutions in Early Trade: The MaghribiTraders’ Coalition,” American Economic Review, 1993, pp. 525–548.

, “History Lessons: The Birth of Impersonal Exchange: The Community ResponsibilitySystem and Impartial Justice,” Journal of Economic Perspectives, 2006, 20 (2), 221–236.

GSMA, “State of the Industry: Mobile Financial Services for the Unbanked,” TechnicalReport 2015.

Hamdard, Javid, “The state of telecommunications and internet in Afghanistan six yearslater 2006-2012,” USAID Assessment Report, 2012.

Jack, William and Tavneet Suri, “Risk Sharing and Transactions Costs: Evidence fromKenya’s Mobile Money Revolution,” American Economic Review, 2014, 104 (1), 183–223.

Jensen, Robert, “The digital provide: Information (technology), market performance, andwelfare in the South Indian fisheries sector,” The quarterly journal of economics, 2007,122 (3), 879–924.

Johnson, Raymond G, Mohamed Kandeh, Abdulai Jalloh, Gerald C Nelson, andTimothy S Thomas, “Sierra Leone,” IFPRI book chapters, 2013, pp. 323–352.

BIBLIOGRAPHY 114

Johnson, Simon, John McMillan, and Christopher Woodruff, “Property Rights andFinance,” American Economic Review, 2002, 92 (5), 1335.

Karlan, Dean, Margaret McConnell, Sendhil Mullainathan, and Jonathan Zin-man, “Getting to the Top of Mind: How Reminders Increase Saving,” Working Paper16205, National Bureau of Economic Research July 2010.

, Robert Darko Osei, Isaac Osei-Akoto, and Christopher Udry, “Agriculturaldecisions after relaxing credit and risk constraints,” 2012.

Keynes, JM, The General Theory of Employment, Interest and Money, London: Macmil-lan, 1936.

Kish, Leslie, “A procedure for objective respondent selection within the household,” Jour-nal of the American Statistical Association, 1949, 44 (247), 380–387.

Macchiavello, Rocco and Ameet Morjaria, “The Value of Relationships: Evidence froma Supply Shock to Kenyan Flower Exports,” 2012.

and , “Competition and Relational Contracts: Evidence from Rwandas Coffee Mills,”mimeo Harvard, 2014.

MacLeod, W Bentley and James M Malcomson, “Implicit contracts, incentive com-patibility, and involuntary unemployment,” Econometrica: Journal of the EconometricSociety, 1989, pp. 447–480.

Madrian, Brigitte C., “Matching Contributions and Savings Outcomes: A BehavioralEconomics Perspective,” Working Paper 18220, National Bureau of Economic ResearchJuly 2012.

and Dennis F. Shea, “The Power of Suggestion: Inertia in 401(k) Participation andSavings Behavior,” The Quarterly Journal of Economics, November 2001, 116 (4), 1149–1187.

Maimbo, Samuel Munzele, The money exchange dealers of Kabul: a study of the hawalasystem in Afghanistan, World Bank Publications, 2003.

Mas, Ignacio and Dan Radcliffe, “Scaling mobile money,” Journal of Payments Strategy& Systems, 2011, 5 (3), 298–315.

Mbiti, Isaac and David N Weil, “Mobile banking: The impact of M-Pesa in Kenya,”Technical Report, National Bureau of Economic Research 2011.

McMillan, John and Christopher Woodruff, “Dispute Prevention Without Courts inVietnam,” Journal of Law, Economics, and Organization, 1999, 15 (3), 637–658.

BIBLIOGRAPHY 115

and , “Interfirm Relationships and Informal Credit in Vietnam,” Quarterly Journal ofEconomics, 1999, pp. 1285–1320.

Mel, Suresh De, David McKenzie, and Christopher Woodruff, “Returns to capital inmicroenterprises: evidence from a field experiment,” The Quarterly Journal of Economics,2008, 123 (4), 1329–1372.

Michaillat, Pascal and Emmanuel Saez, “A Theory of Aggregate Supply and AggregateDemand as Functions of Market Tightness with Prices as Parameters,” 2013.

Miguel, Edward and Gerard Roland, “The long-run impact of bombing Vietnam,”Journal of Development Economics, 2011, 96 (1), 1 – 15.

Mueller, Hannes, “The Economic Cost of Conflict,” Working Paper, 2013.

Posner, Richard A, “The Social Cost of Monopoly and Regulation,” Journal of PoliticalEconomy, 1975, 83 (3), 807–828.

Singh, Prakarsh, “Impact of Terrorism on Investment Decisions of Farmers: Evidence fromthe Punjab Insurgency,” Journal of Conflict Resolution, 2013.

Suri, Tavneet, William Jack, and Thomas M Stoker, “Documenting the birth ofa financial economy,” Proceedings of the National Academy of Sciences, 2012, 109 (26),10257–10262.

Sutton, John and Bennet Kpentey, An Enterprise Map of Ghana, International GrowthCentre, 2012.

and Donath Olomi, An Enterprise Map of Tanzania, International Growth Centre,2012.

and Gillian Langmead, An Enterprise Map of Zambia, International Growth Centre,2013.

and Nebil Kellow, An Enterprise Map of Ethiopia, International Growth Centre, 2010.

Thaler, Richard H. and H. M. Shefrin, “An Economic Theory of Self-Control,” Journalof Political Economy, April 1981, 89 (2), 392–406.

and Shlomo Benartzi, “Save More TomorrowTM: Using Behavioral Economics to In-crease Employee Saving,” Journal of Political Economy, February 2004, 112 (S1), S164–S187.

Tullock, Gordon, “The Welfare Costs of Tariffs, Monopolies, and Theft,” Economic In-quiry, 1967, 5 (3), 224–232.

BIBLIOGRAPHY 116

United Nations High Commissioner on Refugees, “2014 UNHCR country operationsprofile-Pakistan,” 2014.

Venables, Anthony J, “Economic Geography and African Development,” Papers in Re-gional Science, 2010, 89 (3), 469–483.

Voors, Maarten, Eleonora Nillesen, Philip Verwimp, Erwin Bulte, RobertLensink, and Daan Van Soest, “Violent Conflict and Behavior: A Field Experimentin Burundi,” The American Economic Review, 2012, 102 (2), 941–964.

Webb, Matthew D, “Reworking wild bootstrap based inference for clustered errors,” Tech-nical Report, Queen’s Economics Department Working Paper 2013.

Williamson, Oliver E, “Markets and Hierarchies,” New York, 1975.

, “Transaction Cost Economics: The Governance of Contractual Relations,” Journal ofLaw and Economics, 1979, pp. 233–261.

World Bank, World Development Report 2011: Conflict, Security, and Development, WorldBank, 2011.

Zwane, Alix Peterson, Jonathan Zinman, Eric Van Dusen, William Pariente,Clair Null, Edward Miguel, Michael Kremer, Dean S. Karlan, Richard Horn-beck, Xavier Gine, and others, “Being surveyed can change later behavior and relatedparameter estimates,” Proceedings of the National Academy of Sciences, 2011, 108 (5),1821–1826.

Documents

Essays on Markets and Institutions in Emerging Economies · 2018-10-10 · Under the monopoly manufacturer, late deliveries are common due to outside de- ... violence retain more