Crowdseeding Conflict Data: An Application of an SMS-based Data ...

Crowdseeding Conflict Data: An Application of an SMS-based Data System to Estimate the Conflict Effects of Development Aid Peter van der Windt∗

Macartan Humphreys

September 27, 2012

Abstract Poor quality data about conflict events weakens humanitarian responses and hinders academic research on the dynamics of violence. To address this problem we piloted a data gathering system in Eastern Congo in which reporters in randomly selected villages reported on events in real-time. We describe the data generated through this system and use it to implement a downstream experiment that illustrates how the data can be used. We take advantage of exogenous variation in the allocation of development aid to assess whether aid is associated with increased or reduced violence. Our data suggests that aid had a negative effect on conflict. Critically, by exploiting the continuous nature of our data, we also highlight the sensitivity of estimates of effects to the timing of measurement.

∗ Corresponding author: [email protected]. We would like to thank Junior Bulabula, Simon CollardWexler, Yair Ghitza, Stefan Lehmeier, Steven Livingston, Neelanjan Sircar and Patrick Meier. We also thank Tracy Longacre, Desire Cimerhe and Alain Kabeya for outstanding in-country work. Voix des Kivus has been presented at a number of occasions: 2009 International Conference for Conflict Mapping (Cleveland University), UNOCHA’s Coordination Meeting (Bukavu, Congo), 2010 Folke Bernadotte Conference (Columbia University), 2011 World Bank Workshop (Bukavu, Congo), 2011 Conference on Good Governance in Areas of Limited Statehood (Free University Berlin), 2012 MPSA (Chicago), and the University of British Columbia. We thank the participants at these meetings. The Voix des Kivus pilot project was funded by USAID and benefited from support from the group’s learning network; we thank especially Borany Penh and John Berry. Thanks too to IRC for their partnership on the larger study of the development interventions in East Congo. Most of all we thank our 54 phoneholders who, by sending thousands of SMS-messages, made it possible to learn about events that take place in Eastern Congo.

1

1

Introduction

Information technology is becoming increasingly visible as a means for gathering data on social trends. Web-based social networks are used to organize, mobilize and coordinate activities—from riots in London to regime change in the Middle East. Satellite imagery is used by the UNHCR to track population movements during conflicts or droughts, and by the ICC to learn about the presence of mass-graves in Bosnia. And cellphones, combined with geo-mapping software, are used to create early-warning systems. Information technology is now also employed prominently in research. Satellite imagery has been used to study patronage (Posner (2010)), Twitter to learn about the Gaza Conflict (Zeitzoff (2011)), and GIS data to assess many elements of conflicts (see for example Lujala (2010), Raleigh and Hegre (2009), Lujala et al. (2005), Weidmann et al. (2010) and Raleigh et al. (2010)). Micro-level conflict data is normally collected by conducting a survey or interviews. These approaches produce well-recognized challenges. First, conflict events take place in insecure areas that are often off-limits to research teams for extended periods of time. Second, these areas are often remote and difficult to access. As a result, data is often only gathered years after the fact and is heavily dependent on the ability of subjects to recall complex events. Finally, the areas where conflict takes place are often characterized by high levels of distrust. As a result, current methods of collecting data on conflict are likely to suffer from a set of biases that impact the quality of research—we will discuss these in more detail in the next section. We seek to investigate how information technology can overcome the problems associated with the collection of conflict data. In particular, we try to understand how cellphone technology can be used to obtain high-quality data on local-level conflict events. The benefits of cellphones seem obvious: most populated parts of the globe have—or will soon have—phone coverage; using cellphones is cheap—while research teams have problems crossing bad roads or washed-away bridges, cell phone signals do not; and finally, an SMSmessage sent is received instantaneously—information no longer has to travel for days on a motorbike while exposed to vicissitudes induced by the weather or by fighting factions. The use of cellphones to collect data is not new. In recent years many data-collection projects have been set up following an approach called “crowdsourcing”: anyone can send an SMS-message to a central platform in which messages are gathered, stored and visualized on a map. The Ushahidi platform has pioneered this approach and supported it for a fast-growing range of applications. Though clearly a compelling system for “fire alarm” monitoring, there are reasons to worry that the data generated through such systems is not representative.1 Reporters might send incorrect information—for example they might overreport hardship, hoping for humanitarian intervention. The data may also be unrepre1 Berinsky et al. (2012), for example, find that information collected via Amazon’s Mechanical Turk—a particularly popular crowdsourcing system to collect data for experimental research in political science—is not representative of the wider population. The MTurk subject pool is no worse than convenience samples.

2

sentative for more innocent reasons; only people with access to a cellphone (and who have heard about the project) can send messages. Furthermore, the system will only receive messages from people that are willing to pay the cost of an SMS.2 To address these concerns we piloted a “crowdseeding” system that depended on preidentified informants in randomly sampled locations. The project, Voix des Kivus, was implemented between 2009 and 2011 in one of the war-torn province of South Kivu in Eastern Democratic Republic of Congo. Large parts of the province are hard-to-access and although there is reasonable cellphone coverage, access to cellphone technology is limited in the more remote areas. In this context there are two key advantages of a crowdseeding approach. First because the phoneholders have been pre-selected (and provided with the means to report information) in a set of randomly-selected villages, the information has a stronger claim to being representative. Second there are potential benefits for the quantity and quality of the data from the fact that reporters enter a relationship with the research team. In this paper we introduce data from the Voix des Kivus system and demonstrate its utility in practice. We first discuss the concerns for data quality when researching conflict areas (Section 2). We then describe our experiences with Voix des Kivus and discuss how the crowdseeding approach may be able to overcome important limitations of other data-collection efforts present in conflict areas (Section 2.2). The data and data quality is discussed in Section 2.3 and 2.4, respectively. Section 3 illustrates the use of this data in the context of a downstream experiment on aid and conflict. We do this by combining the Voix des Kivus information with information on the presence of a randomized development intervention in the region. Despite the small sample, we find relatively strong evidence for a negative relationship between aid and conflict suggesting that the presence of aid projects, served to protect communities from local violence. Furthermore, our data also suggests that the assessment of this causal quantity is highly sensitive to the timing of measurement in a way that cannot be assessed from traditional cross-sectional data.

2

Real-time Conflict Data

While over the last years there has been increased focus on tackling violence, the development of micro-level measures of violence remains weak. There have been improvements in data collection on violence but this data is typically at a high a level of aggregation. As Verwimp et al. (2009) argue: “At a fundamental level, conflict originates from individuals’ behavior and their repeated interactions with their surroundings, in other words, from its micro-foundations.” Recent work by Autesserre illustrates the point: Although the war in 2 A Ushahidi-based crowdsourcing platform was introduced in Congo in 2008 onwards to provide a platform to report on conflict events. The system fell out of use with only a handful of messages being received, despite many ongoing events, thus illustrating the risks of a system that depends on possibly weak supply incentives.

3

the DRC officially came to an end in 2003, thousands of civilians still die each week. A plausible reason is that the peace settlement has been reached with a focus solely at the regional and national levels; ignoring the local level where conflicts over land and political power increasingly became self-sustaining and autonomous from the national and regional tracks.3

2.1

Challenges in the Collection of Conflict Data

Obtaining detailed micro-level information about conflict events in real-time is difficult because such events cluster in areas that have three characteristics. First, they are insecure, which limits the ability of researchers to visit them for extended periods of time. Second, conflict often takes place in areas that are physically hard-to-access (indeed difficult terrain is significantly related to a higher incidence of civil war (Fearon and Laitin, 2003)). Third, conflict areas may exhibit high levels of societal suspicion and distrust due, for example, to recent conflict experiences or the influx of new, unknown people due to displacement. These characteristics can create different types of biases that impact data quality. First are selection biases. Areas are often off-limit while conflict events take place. The researcher can therefore access and enlist the cooperation of only a (non-random) fraction of the research populations.4 Such selection biases may arise in subtle ways—for example, even if all areas are covered by a survey, security considerations may result in scheduling adjustments so that different regions are only visited when it is safe to do so. Particular selection biases also obtain with retrospective data when testimonies are only taken from the living. Second are risks of recall bias arising from the gathering of data after the conclusion of conflict. Coughlin (1990) provide an overview of the literature. While time cues might lessen this bias in some cases, De Nicola and Gin´e (2012) find evidence that the use of time cues can sometimes worsen accuracy. One example is the problem of ‘telescoping’ where respondents run together several events and lose track of when they actually happened. The importance of prompt data collection then depends on the number of events taking place. Third are risks of reporting bias that may arise when sensitive questions are asked in settings characterized by distrust and suspicion. Trust has to be created to obtain high-quality responses. However, this is especially difficult in conflict areas where datacollection is often a one-off activity.5 Finally, reporting may also be biased because of deliberate attempts to manipulate information.6 3

Autesserre (2009) and Autesserre (2010). Cohen and Arieli (2011) discuss how the populations in conflict areas are hidden from and hard to reach for researchers, and argue how the “snowball surveying”-method works to access such populations. 5 Various approaches to get at sensitive information in these environments have recently been developed. See for example Corstange (2008). 6 Luyendijk (2009), for example, discusses how in the Middle East the actors provide neatly packaged story lines to media reporters. A more subtle way to influence the data is for government or rebel groups 4

4

2.2

Crowdseeding

In principle a crowdseeding approach may alleviate some of these concerns. Rather than relying on punctual retrospective data gathering, crowdseeding relies on continuous information flows from preselected reporters from a random set of villages. Fundamentally, this approach combines approaches traditionally used by intelligence groups with sampling approaches more commonly employed in social scientific survey research. The approach promises a number of advantages. First data generated in this way may be more representative. If participating sites are selected at random from a given population then the resulting data is representative of that population at the village level. Moreover if the reporters are pre-identified and given the means to provide information, risks of self-selection at the reporter level are minimized. In addition, because a larger part of the research population stays available for data-collection, worries about selection bias might be be lessened—even if an area is off-limits to a research team, SMS-messages can still be sent and received. Also in the event of (forced) displacements the reporter can continue providing information via his or her portable phone. Second, the real time nature of crowdseeding removes concerns related to recall and allows for the possibility of a detailed examination of conflict dynamics. Third, risks of reporting bias may be reduced because of the creation of a relationship with reporters. Repeated interactions can reduce the incentive for phoneholders to send incorrect information and allows for ready auditing and verification of data. To pilot this approach we implemented the Voix des Kivus (literally “Voice of the Kivus”) pilot project starting in 2009 in the province of South Kivu in the war-torn East of the Democratic Republic of Congo. The pilot ran in eighteen villages spread over four territories (Kalehe (5), Mwenga (4), Uvira (3), and Walungu (6)).7 Figure 1 shows the approximate area of operation.8 Within these villages the Voix des Kivus system worked as follows. In each village there were three reporters (“phoneholders”) identified: one representing the traditional leadership (generally the chief of the village), one representing women’s groups (generally the head of the women’s association), and one elected by the community.9 to grant a researcher access to only certain areas, and deny it to others. 7 The territory is the administrative unit below the province and above the chiefdom. Villages 1-4 were selected purposively to ensure some conflict and some non conflict areas; villages 5 - 18 were selected using stratified random sampling from these territories. 8 Note that to safeguard the safety of the phoneholders the village names are omitted. Moreover, locations within chiefdoms have been scrambled and so the actual location of participating villages can not be inferred from the location of points on this map. We provided information about village location only to precleared organizations. We thank UNOCHA for the shapefiles used to construct this map. 9 The idea behind these three reporters is that the chief is the go-to person for many affairs in a village ranging from land conflict to marital disputes. The head of the women’s association, on the other hand, is often the go-to person for issues such as domestic violence. Finally one person is elected to decrease concerns of capture. These reporters can also be selected randomly to obtain a representative sample within the

5

Figure 1: Voix des Kivus area of operation: South Kivu.

Note: Squares denote Voix des Kivus villages, where solid (hollow) squares denote villages that (do not) participate in the development program.

6

Holders were trained on the use of the phone and how to send messages to the system. They were provided with a phone, weekly credit, and a codesheet that lists possible events that can take place in the village.10 Phoneholders automatically received weekly phonecredit that they could freely use, and were reimbursed for the number of messages sent. Sending messages to the system was therefore free but it was also voluntary—while users did not have to pay for each message, they did not get any financial rewards for sending content to the system either. On the receiving side a standard cellphone linked to a laptop comprise the necessary equipment. With freely available software (FrontlineSMS, R and LaTeX), messages received were automatically filtered, coded for content, cleaned to remove duplicates, and merged into a database. Graphs and tables were generated automatically and then mounted into bulletins with different levels of sensitivity. Translations of non-coded text messages (often from Swahili or one of the local languages into French and English) were undertaken manually.11 Our first goal in implementing the pilot was to assess the feasibility of data collection of this form. Establishing feasibility was important in light of a number of obvious challenges. A first concern was that capacity would be too weak for users to implement the project. Indeed, 19 of our 54 reporters are only primary-school educated and with the exception of two participants, nobody had received education beyond secondary school. This concern is likely to be present in many hard-to-access areas where conflict takes place. We mitigated this concern through a (minimum) two-day training by the field coordinator and the use of relatively simple code-books with pre-coded events. In fact, over the eighteen months of operation Voix des Kivus received—in addition to SMS-messages with codes—1,244 textmessages in Swahili or one of the local languages; this suggests that the codebook, while useful, was not necessary for many holders. A second concern was technical capacity. On the receiving side a cheap netbook, a phone and free software was sufficient. However, things were different on the sending side. Phone-coverage was not a problem—of the eighteen villages no village had to be replaced after selection for lack of coverage. Electricity, however, was a problem with phoneholders walking up to three hours to charge their phone. The problem was solved by purchasing $25 solar chargers and donating one to each Voix des Kivus village. These concerns exist for many phone-based projects that operate in a hard-to-access area. A third concern, however, is specific to crowdseeding systems: would participation in the system produce security risks for phone holders? We put in place multiple measures to reduce risks and rolled out the system slowly and with high levels village. 10 All documents can be found on the project’s website: www.cu-csds.org/projects/event-mapping-incongo/. This also includes the computer code to create bulletins and a “Voix des Kivus Implementation Guide” for organizations that want to set up a similar system. 11 While Voix des Kivus started as an academic exercise, such a system also has the potential to empower participating communities, and provide a basis for response to practitioners and policy advocates. As a result, we set up a system in which each Monday a bulletin with event-information was produced and disseminated to organizations that had received clearance from Voix des Kivus and its phoneholders. Many of these organizations are based in Bukavu and have the means to respond.

7

of monitoring to assess these risks. We found that at no point did any user indicate any security concern of any form arising from their participation in the project. However it is important to note that the concern for the security of the phoneholders is particularly present in systems based upon a crowdseeding approach because of the direct link between phoneholder and project. In Section 4 we discuss this in more detail.

2.3

The Data

To facilitate the sending of messages we distributed code-books with pre-assigned codes to events, organized in ten categories: (1) presence of military forces, (2) attacks on the village, (3) deaths related to armed combat, (4) local violence and property loss, (5) displacement, (6) health events, (7) natural disasters, (8) development and NGO activities, (9) social events and (10) special codes. An overview of these codes is given in Table 1. To our phoneholders we distributed this list in French and Swahili. If no event took place during the week reporters would send the code “00”. If an event took place that was not listed on the code-book or the reporter prefered to provide more detail, the code “98” followed by text could be send. Voix des Kivus was launched in August 2009 and operated during the first twelve months in only four villages. Then from August 2010 onwards the pilot project was expanded to eighteen villages.12 In this short period the project received a total of 4,783 non-empty SMS messages. The phoneholders sent messages about a total of 5,293 events of which 4,623 were unique—many village events were thus reported by more than one phoneholder. Of these non-empty SMS messages 1,244 were text-messages.13 Conflict is a dynamic phenomena with periods of more and less violence. One major advantage of a system like Voix des Kivus is that it captures this dynamic component. To illustrate the temporal variation, the top panel of Figure 2 plots the number of conflict (solid line) and total events (dashed line). The center panel shows the cumulative number of messages sent per phoneholder. As the linear lines indicate the rate of receipt is sustained over time, showing that phoneholders display no sign of reporting fatigue.14 Each week Voix des Kivus sent a fixed amount of phonecredit ($1.5, or around a day’s 12

Dates that Voix des Kivus started: 1-Aug-09 (3x), 17-Aug-09, 10-Aug-10, 13-Aug-10 (2x), 9-Oct-10 (2x), 26-Oct-10, 30-Oct-10, 1-Nov-10, 2-Nov-10 (2x), 4-Nov-10, 6-Dec-10, 8-Dec-10, 10-Dec-10. The village that started on 17-Aug-09 started later than expected because not enough people were present at the first general assembly. We had as rule that at least 40% of the village had to be present, be informed of the pilot project and give their consent. 13 This is based on the FrontlineSMS export of August 1, 2011. Table 3 in Section 5 provides summary information per event where the unit of analysis is the village-month. Identical messages received from the same reporter within 30 minutes of each other, and identical messages received from different representatives in the same village within 24 hours of each other, are treated as single events. 14 Reporting by one phoneholder (the upper-line) is anomalous. We noticed after one month that messages by this reporter had been sent to the wrong phonenumber. The reporter had recorded his messages in a notebook however and subsequently uploaded them to the system (resulting in the steep slope in October 2009).

8

Figure 2: Top panel shows the average number of events per month by type over time. Middle panel shows reporter stamina—the cumulative number of events sent per reporter; this panel also indicates when reporters started and stopped sending messages. The lower panel shows the level of “Internal Validation”—that is the share of messages sent that reported on events that were also reported by other reporters.

9

Code

Event

Code

PRESENCE OF MILITARY FORCES

Event HEALTH

11

Presence of MONUSCO

61

New outbreak of disease

12

Presence of men in uniform

62

Civilian death due to disease

13

Presence of other armed groups

63

Civilian death due to natural causes

ATTACKS

64

Sexual violence against women

21

Attacks on village by MONUSCO

65

Sexual violence against men

22

Attacks on village by men in uniform

23

Attacks on village by rebel group

71

Flooding/heavy rain

24

Attacks on the village by unknown group

72

Large forest or village fire

DEATHS RELATED TO ARMED COMBAT

73

Earthquake

31

Civilian deaths (man)

74

Drought

32

Civilian deaths (woman)

75

Crop failure/ plague

33

Civilian deaths (child)

34

MONUC deaths

81

Complaint against NGO

35

FARDC deaths

82

Practice message (for Voix des Kivus)

36

Other rebel group deaths

85

Constr./reparation/rehab. of school/health center

LOCAL VIOLENCE AND PROPERTY LOSS

86

Construction/reparation/rehab. of church/mosque

41

Rioting

87

Other constr./reparation/rehab. (eg roads, wells)

42

Looting/ property damage

88

Organization of security patrols

43

Violence between villagers

89

Work to improve agricultural productivity

44

Violence between villagers due to a land conflict

45

Domestic violence

91

Funeral

46

Ethnic violence

92

Wedding/ Other celebrations

47

Forced labor by men in uniform

93

Visit/meeting organized by nat. or prov. authorities

48

Forced labor by other army groups

94

Visit/meeting organized by territoire authorities

DISPLACEMENT

95

Visit/meeting organized by local authorities

51

Kidnapping by men in uniform

96

Visit/meeting organized by a political party

52

Kidnapping by rebel group

97

Visit/meeting organized by a traditional leader

53

Arrival of Refugees or IDPs

54

Departure of villagers as IDPs

98

Unclassifiable issue (followed by text)

55

Disappearances

99

Security alert

56

Villagers were forced to move

57

Villagers decided themselves to move

00

NOTHING TO REPORT

NATURAL DISASTERS

DEVELOPMENT ACTIVITIES/ NGOs

SOCIAL

SPECIAL CODES

Table 1: Voix des Kivus Code-book wages) to the phoneholders that they could use as they saw fit (to call friends, call the hospital, start a company, etc.). In order to receive this credit the holder had to send at least one (possibly empty) SMS per week. To avoid phoneholders only sending one message per week, Voix des Kivus reimbursed the phoneholders for all SMS messages sent. At the end of the project we had the opportunity to learn about the importance of the weekly transfer of $1.5—an incentive beyond simply giving the means to provide information (which is the cellphone and the reimbursement). On January 31, 2011, the project officially ended under the USAID grant (indicated in Figure 2 by the vertical line). However, because of the interest by the phoneholders we continued the project for several months; but with one difference. We sent a one-time $20 of phonecredit to the phoneholders with the message that this phonecredit was for them to use in any way they want, that they would still be reimbursed for their messages and that we would continue to collect SMS messages and distribute the bulletins. However, we also told them that from that moment onwards they would no longer receive the weekly $1.5 nor would they receive any feedback from the system. As can be seen from the top panel in Figure 2, the result was a gradual decline in 10

the messages sent from January 2011 onwards reaching zero by July 2011.

2.4

Data Quality

Voix des Kivus was implemented to assess the feasibility of using crowdseeding to gather high-quality data. In the absence of a gold standard for conflict event data it is difficult to formally validate the approach employed here (we describe approaches that might be adopted for formal validation below). Nevertheless four considerations on data quality bear emphasis. First, in its first year Voix des Kivus employed a field coordinator to visit each of the phoneholders at least once every two weeks to assess the quality of the sent-messages.15 The coordinator would assess whether holders interpreted codes in the same way as researchers, whether respondents employed the correct codes given the events they wished to report and so on. Moreover, the field coordinator would verify whether events had actually taken place and whether there were other events that took place that were not reported. Second, because Voix des Kivus distributed phones to three people per village we have a measure of internal validation. The bottom panel of Figure 2 shows the share of events that were reported by at least two phoneholders out of all events reported. We find that internal validation was around 30% at the outset but increased after the project start until August 2010—the moment the project expanded to more villages.16 Third, we have a limited ability to compare event information received from the Voix des Kivus system with that from data from a survey we implemented in some of the same regions in a similar period. In the latter we asked each village chief from a list of potential events—events similar to those in the Voix des Kivus code-book—how many took place in the village in the preceding month. Our survey villages overlapped with just three Voix des Kivus villages during the lifespan of the project.17 We consider three ways of assessing the congruence of the survey and the Voix des Kivus data. First we compare the average number of each type of event across villages under each approach. On average we find nearly twice as many reported events according to the survey, a difference that is significant at the 95% level. In particular, there were higher numbers of children’s deaths due to conflict, citizen deaths due to natural causes, disease outbreaks, and funerals reported in the survey. 15 To do so Voix des Kivus would prepare a sheet with the messages sent by that phoneholder in the preceding two weeks. For security reasons these were decoded in such a way that the information only made sense to the field coordinator. 16 The improvements and subsequent decline in validation may in part result from our communications with phoneholders. Consulting phoneholders suggested that the low internal validation early on was in part due to phoneholders informally adopting a division of labor where one would not send a message if they knew another had. In the early months with the first villages we encouraged holders not to do this and to report whether or not they thought others did also. This same messaging was not employed with the later villages as they came online. 17 And we note that even for these three the survey took place during the final ‘voluntary’ phase in which payments were no longer being made to reporters (see Section 2.3).

11

Second, despite the difference in means, the correlation between these two measures of average incidence is 46% which is substantively large and statistically significant. This suggests that similar types of events get reported to the same relative extent under the two systems. Third, we generate measures of the average incidence of conflict and nonconflict events for each of the three village, yielding 6 data points. The correlation between the Voix des Kivus and survey estimates of these 6 numbers is low largely because of one outlying observation in which a village reported a much higher incidence of army visits under the Voix des Kivus system. If this outlier is dropped the correlation between the Voix des Kivus and the survey based measure is 95%. These correlations are encouraging although we note that survey measures are also subject to measurement error and bias and so on their own the correlations do not establish the reliability of the data.18 Fourth, a particular benefit of crowdseeding—in contrast to crowdsourcing—is that we can learn whether particular types of reporters provide different information. Surprisingly, village chiefs sent substantially more messages than the women’s representative or the elected holder: respectively 12.17, 8.73 and 9.30 messages per person per month on average.19 Overall, around 30% of the messages sent concern conflict events—with elected holders sending relatively the most.20 Regressing the number of conflict events reported per month on the type of phoneholder and the number of nonconflict events reported in that month—with the chief as baseline and standard errors clustered at the village level— confirms this. We obtain estimates equal to 0.28 (0.62) and 1.56 (1.36) for the women’s representative and the elected holder (standard errors in parentheses). While the difference between reporter-position and messaging is not statistically significant, the number of nonconflict events reported is positively and significantly associated with the number of conflict events reported at the 95% confidence level—with an estimated marginal effect (and standard error) of 0.45 (0.18). Finally, a problem for any data-collection method in conflict areas, including surveys, is censorship. For example, informants might report less conflict events, not because this is the truth but precisely because of the presence of an armed group and high levels of conflict.21 A particular benefit of a system like Voix des Kivus is the relationship that is build with the phoneholders over time. Given our experience with Voix des Kivus, which 18

We also note that in this case since the chief was also involved with Voix des Kivus the comparison of systems involves comparing reports from the same actor, which may give rise to concerns that the two reports would be more similar than they would be if recorded from different subjects. 19 There is also a large variation over time for these three types of phoneholders, with standard deviations of respectively 14.34, 7.68, and 13.76. 20 The share of conflict events reported over total events reported is 0.27 (0.29), 0.32 (0.32), 0.34 (0.33) respectively for the chief, the women’s representative, and the elected phoneholder (standard deviation in parentheses). On average they sent respectively 2.84 (5.94), 2.19 (2.50), 3.20 (7.45) messages concerning conflict per month. As we discuss in more detail in Section 3.2, conflict events are measured as the messagecodes: 21-56, 64-65, and 99. 21 As will be discussed in more detail in Section 3.1, in South Kivu we find that development aid is associated with greater troop presence, but also with fewer violent events.

12

included regular contact with phoneholders, at no point in time did we think censoring to be a concern. To examine the possibility of censorship we look at the correlations between the presence of MONUSCO, FARDC and rebel troops, and a set of violent events that have received a lot of media-attention in recent years: attacks on the village by armed groups different than MONUSCO or FARDC, civilian deaths, sexual violence against women, and kidnapping.22 Are reporters less likely to report these events when different actors are present. To assess the possibility we regress the incidence of the violent event on the presence of MONUSCO, FARDC and rebel troops simultaneously (and a month lag of these variables, including the dependent variable). We cluster the errors at the village level. Table 2 presents the results of these regressions. We find evidence that—as expected—the presence of the United Nations peacekeeping army (MONUSCO) in the previous month has a statistically significant, negative correlation with the incidence of attacks on the village and the incidence of sexual violence. Moreover, we find a positive, and statistically significant, contemporaneous association between the presence of peacekeepers and attacks on the village in that month. Table 2 shows also how rebel presence is positively correlated with attacks and kidnappings. This is not surprising and the large magnitude—particularly compared to those of MONUSCO and FARDC—is telling. While statistically not significant, there is also a positive association between rebel group presence and civilian deaths, while these coefficients are negative for MONUSCO and FARDC. Finally, the results for the presence of the FARDC are not significant and point in different directions. This is not surprising. Indeed, while on the one hand the government troops are present to provide security, their relation to the local populations is often contentious. Messages such as “FARDC passed by [Village Omitted ] and shot with guns at 5:30 am.” and “A FARDC soldier shot a woman.” are not uncommon. And messages such as “FARDC kidnapped people who were then forced to transport things.” and “FARDC kidnapped people to build a shelter for the soldiers.” were received almost weekly by Voix des Kivus.23 These correlations suggest that reporters were willing to report on conflict events in moments in which (according to their reports) different actors were present; it is still possible however that in some situations some reporters simultaneously failed to report on both troop presence and conflict events.

2.5

Research Applications

A system like Voix des Kivus thus provides a rich dataset to learn about violence in Eastern Congo, and conflict areas in general. A set of interesting relationships between different events could be analyzed as we did in Section 2.4. Particular use can be made of the data’s 22

News reports reporting such events are common with regards to Eastern Congo—and South Kivu in particular. See, for example, BBC (2009) and BBC (2011). 23 The first two messages were received on 23 April 2010 by a village chief and 23 May 2010 by one of the democratically-elected phoneholders. The contentious relation between the FARDC and local populations has also been highlighted in numerous reports by agencies such as the United Nations and the International Crisis Group (See, for example, ICG (2009) and UNSC (2012)).

13

Presence

Attack

Civ. death

Sex. violence

Kidnapping

MONUSCO

0.09***

-0.01

-0.02

-0.06***

(0.01)

(0.03)

(0.01)

(0.02)

-0.07***

-0.08

-0.02***

0.02

(0.02)

(0.07)

(0.01)

(0.03)

MONUSCO (Lag) FARDC FARDC (Lag) REBEL REBEL (Lag) DEP. VAR. (Lag) N:

0.01

-0.03

0.03

0.13

(0.06)

(0.11)

(0.02)

(0.09)

-0.01

-0.01

-0.01

-0.11

(0.05)

(0.06)

(0.01)

(0.10) 1.19***

0.53**

0.34

0.00

(0.23)

(0.43)

(0.10)

(0.40)

-0.16**

0.99

-0.03

-0.33***

(0.07)

(1.09)

(0.06)

(0.11)

0.03

0.39*

0.11

0.55***

(0.11)

(0.19)

(0.08)

(0.06)

193

193

193

193

Notes: Standard errors, clustered at the village level, are provided in parentheses. One, two or three asterisks indicate, respectively, significance levels at the 90%, 95% and 99%. Conflict event (code): Attack (23+24), Civ. death (31+32+33), Sex. violence (64), and Kidnapping (51+52).

Table 2: Presence of MONUSCO, FARDC and Rebel Groups panel-component to learn about the dynamics and impact of certain events over time.24 Another interesting application of a phone-based system is to combine it with external datasources, and use the data collected by phone as an outcome measure. The local United Nations offices could provide historic information about patrol routes which can then be used to learn whether these patrols pacify a region or simply move violence by a rebel group from one area to another.25 The data could be combined with satellite information about weather patterns to learn about the impact of, for example, rainfall on different types and levels of conflict. To learn about election violence, data could be collected from Congo’s National Election Committee on the location of voting booths (in November 2011 the presidential elections took place in Congo and was in some areas characterized by high levels of violence). A final application is to use a Voix des Kivus style system as a treatment in order to examine the effects of transparency on conflict outcomes of interest. In the next section we illustrate an approach in which Voix des Kivus data provides outcome data for assessing the conflict impact of development. 24

See, for example, Zeitzoff (2011) who draws data on hourly conflict intensity from Twitter and other social media to learn about the Gaza Conflict (2008—2009). 25 This is an important question that we should also ask in Section 3 where we analyze the conflict impact of a development program. Does this program decrease the level of violence, or is violence simply moved to another area? We do not answer this question in the paper.

14

3

Application: The Conflict Impact of Development Programs

We illustrate the use of this data by examining the effects of a randomized development intervention on local conflict events. There is broad recognition that development and conflict are closely related. Major development interventions focus on countries emerging from war and include reintegration, reconstruction, capacity building and other initiatives. It remains unclear however if and how these investments are effective. On the one hand, it is expected that by improving the level of development these projects could reduce risks of violence. Miguel et al. (2004), for example, find that economic growth is strongly negatively related to civil conflict. And De Ree and Nillesen (2009) show how foreign aid is found to have a direct negative impact on the probability of an ongoing civil conflict to continue.26 However, the very introduction of development actors, financing, and projects could also have adverse effects on violence. Nunn and Qian (2012), for example, suggests that U.S. food aid increases the incidence, onset and duration of civil conflicts in recipient countries and related arguments have been made by Anderson (1999), De Waal (2009) and Polman (2011). There are multiple reasons why these relations–or opposite relations may obtain. These include changing the resource available for looting, changing the capacity of local communities to respond, changing the incentives of local communities to participate in violence, and increasing the extent to which communities are—or are believed to be— monitored by external actors. Much of this development-conflict literature is at the macro, often the national, level. Conflict research at the micro-level, and in particular the impact of development aid on local levels of violence, is sparse and what does exist has up to recently been largely of a descriptive nature.27 This is not surprising because attempts to find out whether these aid programs actually have an impact on the level of violence face two major challenges. First, there is a selection problem: the fact that poverty and the incidence of violence are intertwined makes identifying causal linkages especially challenging.28 Thus if programs are implemented in poorer communities that have higher (or lower) levels of violence than comparison groups, a comparison of outcomes will lead to inaccurate estimates of program effects. This problem can be addressed through random assignment of communities to interventions which helps ensure that program (“treatment”) villages and comparison (“control”) villages are alike in all respects except for the intervention itself. To give one example, using a large-scale randomized field experiment, Beath et al. (2011) suggests that 26

See also Collier (2003). In addition, little work exists that ties both levels together. See for a particular strong illustration of the disconnect between the micro and macro levels in the DRC: Autesserre (2010). 28 While we referred to the literature discussing the impact of development on conflict, there is also a large literature discussing the impact of conflict on development. Bates (2009), for example, argues that it is the reorganization of coercion—not its extinction—that underpins the security needed for investment and prosperity. 27

15

a development programs in Afghanistan lead to reports of an improved security situation.29 The second problem is one of measurement and has been the focus throughout this paper. It is difficult to obtain micro-level information about violent conflict. While there have been improvements in data collection on violence at a more aggregate level, development projects often take place at the village or community-level—some villages receive the project and other (maybe even neighboring) villages do not. It is therefore necessary to obtain conflict data that distinguishes conflict at such a low level. Current studies usually make use of surveys to obtain such data which—as has been discussed in detail in Section 2—is prone to a set of biases that impact the quality of this data. We therefore turn to Voix des Kivus, and the randomized intervention we examined in Eastern Congo. Between 2007-2011 a DFID-funded randomized development intervention was implemented by the International Rescue Committee and CARE International throughout Eastern Congo—including the province of South Kivu. The villages were selected into the development program by a public lottery—thus in expectation treatment and control villages are alike in all respects except for the intervention itself. This helps address the first problem. The development project provided communities with financing of up to $70,000 to construct local projects such as school rooms or clinics—part of this money was received and managed by the communities directly. The intervention followed a “do no harm” approach. During the selection of the Voix des Kivus villages we took this development project into account. That is, we sampled the Voix des Kivus villages stratified upon the villages’ treatment status in order to obtain balance between the number of Voix des Kivus villages that had the development program and those that did not. Now by combining this treatment status information with the Voix des Kivus pilot data we hope to, first, illustrate how a crowdsourcing system can be used as a measure (thereby alleviating problem 2). Second, we task ourselves to understand the downstream effects of development aid on conflict. That is, although the purpose of the random assignment was not undertaken to allow for a study of the effects of the intervention on conflict outcomes, it nevertheless makes this possible (Green and Gerber (2008)).

3.1

Disaggregated Results

To investigate whether the development project might have had an impact on any of the events listed in the Voix des Kivus codebook we regressed the incidence of each event on the project treatment variable, clustering the standard errors at the village level. Figure 3 provides a visual representation of the results. The circles indicate the coefficient on the 29

This positive effect, however, is not observed in districts with high levels of initial violence. Crost et al. (2012) estimated the causal effect of a large development program on conflict casualties in the Philippines by making use of a regression discontinuity design. They find that eligible municipalities experienced a large but temporary increase in conflict casualties, which was suffered equally by government troops and insurgents. Berman et al. (2011), on the other hand, by making use of panel data from Iraq, find that improved service provision reduces insurgent violence, particularly for smaller projects and since the insurgency began in 2007.

16

treatment variable; that is, the point estimate or magnitude of the development project on the incidence of that particular event. The lines indicates the 95% confidence intervals around these point estimates. Figure 3 shows that, while the standard error is large, development aid is negatively associated with crop failures and plagues (75). Furthermore, Figure 3 suggests a negative relation between the intervention and public goods projects (notably the construction, maintenance and rehabilitation of churches (code 85) and schools and health centers (86), other projects (87), and agricultural productivity projects).30 The effects of the project on conflict events point in different directions. On the one hand the project is associated with greater troop presence from both Congolese government soldiers (FARDC) and the United Nations peacekeeping army (MONUSCO). On the other hand, development aid seems to be associated with fewer violent events: most of the coefficients in Figure 3’s panels “03: Deaths Combat”, “04: Local Violence” and “05: Attacks” have a negative sign. One simple explanation might be that the international intervention resulted in more troop presence which had a pacifying effect. Another, more worrying possibility is that more troop presence leads to more censoring on the part of the phoneholders—an issue we discussed in more detail in Section 2.4.

3.2

Summary Results and Temporal Trends

The real promise of a real-time system however is that it allows one to assess trends over time. Here we use the data to illustrate treatment effects estimated over time. The ability to undertake such analysis has two advantages. First, it provides a finer mapping of intervention effects; for example it is possible that an intervention has effects in some periods and not others, depending on changing local conditions, on the changing nature of the project, or on the timing of maturation or decay of effects. An ability to assess such fluctuating effects can provide insight both into the durability of effects and into the conditions under which they are strongest or weakest. The first panel in Figure 4 gives the estimated impact of development aid on the incidence of conflict over three month periods. To do so we regressed the incidence of conflict on the treatment status variable.31 The standard errors are clustered at the village 30

We are conscious that the interpretation of these results is made difficult by the large number of outcome variables. Because we have many measures it is possible to obtain false positives. That is, it is possible to obtain a statistically significant result by chance. The Bonferroni correction is a (conservative) method used to counteract this problem of multiple comparisons. The correction is based on the idea that if we have n hypotheses and one dataset, then one way of maintaining the category-wide error rate is to test each individual hypothesis at a statistical significance level of 1/n times what it would be if only one hypothesis were tested—0.05 in our case. The Bonferroni correction would use a significance level of 0.05/n. By making use of this correction our earlier results on crop failure and plagues are no longer statistically significant. The negative impact of development aid on public goods projects, however, still holds for two events: for church and agriculture productivity projects. 31 Conflict events are measured as the message-codes: 21-56, 64-65, and 99. Based on the trends apparent in the disaggregated data this construction separates the troop presence measures from the violent event measures. A definition that includes the troop presence measures shows no relation to the intervention in

17

Figure 3: Causal Impact of Development Aid

Note: Circles indicate the point estimate of the development project on the incidence of that particular event. The lines indicates the 95% confidence intervals around these point estimates.

18

level. The two red, dashed lines indicate the 95% confidence interval. For the first set of months the standard errors are large (largely because for these first months there is no historical data). We see that estimates vary over time with strong estimates of effects in early periods which then decline and reemerge, though never to initial levels. Overall, the data provides evidence of a positive impact of development aid: that is, villages with the development program have a lower level of conflict incidence. Not only are the point estimates of the development impact consistently below the zero line, so are most of the 95% confidence intervals. This result is confirmed in the center panel, where the impact of development aid is now not estimated as a three-month moving average but cumulative. That is, month five takes all the data from the preceding five months into account, month 17 the data from the preceding 17 months, et cetera. These panels illustrate the major promise of a real-time system such as Voix des Kivus by providing the treatment effects over time. We find that the intervention had stronger effects in some periods and weaker effects in others. The top and center panels of Figure 4 report results for each period. Given this collection of findings can we conclude that the project led to less conflict overall ? While in most periods the estimates are negative and statistically significant, there are periods in which the estimate is not significantly different from zero. This gives rise to a multiple comparisons problem. We address this problem using a randomization inference procedure (Fisher, 1935).32 In practice, we repeatedly randomly re-assigned our eighteen villages to the treatment and for each of these new re-assignments to treatment we calculated the average p-value over time.33 This formed our test statistic and we used the multiple reassignments to generate the reference distribution for this test statistic. We then generate a final p-value that indicates how likely it is we would have obtained results as strong or stronger than those presented in Figure 4. We find this likelihood to be very small. For both the three-month moving average and cumulative-beta analyses we find these pvalues to equal 0.07.34 We can thus conclude that the development project had an overall negative, and statistically significant, impact on the incidence of conflict. The ability to analyze treatment effects estimated over time has a second, methodological advantage. Existing studies of the effects of development interventions often produce contradictory results; yet these studies differ on many dimensions of design, including the any period. 32 Other approaches to deal with the multiple comparisons problem include using an index of measures (the approach adopted by Kling et al. (2007) for example), statistical adjustments, such as the Bonferroni ˇ ak corrections (see also Benjamini and Hochberg (1995) for methods to control the false discovery or Sid´ rate), or multi-level approaches (Gelman et al., 2009). On randomization inference see also: Barrios et al. (2012), Small et al. (2008), and Ho and Imai (2006).
33 To create our distribution of treatment vectors we do not compute all 18 = 31, 824 treatment vectors, 11 but sample 1,000 possible treatment vectors. Throughout this analysis we do not take into account the first two months for which we have particularly few observations. All replication data and code are posted online at {link}. 34 To be more precise: p=0.066 and p=0.070, respectively.

19

Figure 4: Conflict Impact of Development

Note: The top panel shows the estimated effects based on different three-month periods, including 95% confidence intervals. The center panel shows estimated effects based on cumulated data over time. The bottom panel shows the distribution of estimated effects (x-axis) based on different three-month periods.

20

temporal lag between intervention and measurement. An analysis using real-time data allows one to measure the extent to which the variance in estimated effects depends on the timing of measurement. For example, measures of conflict events that took place in the month prior to data collection might yield an estimated effect with low variance. However, while this within-period variance is low, the across-period variance might be high: very different estimates might have been produced if data were gathered a month earlier or a month later. Our analysis here suggests the timing of measurement contributes a large share of the true variance in estimates of treatment effects. Both sources of variance can be seen in the bottom panel of Figure 4, where the x-axis is the estimated treatment effects for different three-month periods, and the y-axis is the number of months. It is clear that the variance in estimated effects depends strongly on the timing of measurement. Starting from month 4, the variance in estimates of the treatment effect over months 4-24 is 194; the average variance within each period however is just 12. Things settle down after month 6, but even still the variance in estimates over time (months 7—24) is 12 while the average variance is 11.

3.3

Robustness

In the analysis above we account for within-group dependence by using cluster-robust standard errors. Without such a correction, the OLS standard errors can be downward biased. Bertrand et al. (2004), however, discuss how even with this correction a researcher is still likely to over-reject with few (less than thirty) clusters. In this study we randomly assigned the treatment to only eighteen villages and we might therefore be worried about the robustness of our results. Consequently, we re-estimate our results by wild clusterbootstrapping the standard error based on Cameron et al. (2008). Cameron et al. (2008) argue that this method provides an asymptotic refinement and show how rejection rates of 10% using standard methods can be reduced to the nominal size of 5%. The top panel of Figure 5 provides the results of re-doing the three-month moving average analysis (the robustness results for the cumulative-beta analysis can be found in Section 6). Overall, we find similar results. As a second check, we re-conduct all tests using randomization inference. To do this, we randomly re-assigned our eighteen villages to the aid treatment, estimated for each period the causal impact of aid, and thereby obtain a distribution of parameter estimates for each period under the null of no effect. Comparing this to the estimate in that period from the actual (true) assignment to treatment we calculate the probability that we would have found that same estimate (or larger) in our data. The bottom panel of Figure 5 provides these p-values over time for the three-month moving average analysis. The red, solid and dashed lines provide the 5 and 10% levels, respectively. With the exception of the first two months for which we have very few observations, all observations are around or below 30%, and most below the dashed 10% line. Combined with the results of the randomization inference regarding the impact of the program overall, we can be quite confident that the 21

results in this paper are unlikely to have arisen by chance. As a final robustness check we redo our analyses by excluding the first four Voix des Kivus villages because these initial villages were not selected randomly. Moreover, only one out of these four villages received the development project. From the center panel in Figure 5 we can see that while the loss of data results in larger standard errors, the estimates are almost identical (as highlighted by the fact that the estimates overlap the dotted line that provides the estimates from Figure 4).

4

Conclusion

The past decade has seen much effort to better understand the causes and consequences of violent conflict. Much of the empirical analysis, however, makes use of data at high levels of aggregation (often the country and year level). It is therefore not surprising that in their survey of the civil war literature Blattman and Miguel (2010) argue that “a major goal of civil war researchers within both economics and political science in the coming years should be the collection of new data, especially extended panel micro-data sets. . . ” Original micro-level data is normally collected by conducting a survey or interviews. The collection of conflict data in particular faces a set of challenges because of the nature of the areas in which conflict events take place. First, these events take place in insecure areas and are therefore off-limits to the research team and its partners for extended periods of time. Second, these areas are often remote and difficult to access. In consequence, data is often only gathered years after the fact and is heavily dependent on the ability of subjects to recall complex events. And, finally, the areas where conflict takes place are often characterized by high levels of distrust. These factors together do not only make it difficult to collect detailed micro-level information over time, it makes this collected data prone to a set of biases that negatively affect its quality. In this paper we investigated whether a SMS-based “crowdseeding” system can be used to obtain high-quality, micro-level panel data on conflict. To address the question we piloted the Voix des Kivus system between 2009 and 2011 in one of the most war-torn provinces of Congo, South Kivu. Phoneholders were pre-selected from randomly sampled areas to provide regular reports to the system. This approach holds out the promise of multiple benefits for data quality, most of important of which is a claim to representativeness. By implementing such a crowdseeding system we hope to overcome the problems that come with current methods to collect data on conflict. Our first objective in this research was to probe the feasibility of employing a crowdseeding approach to gathering high-quality data on conflict events in real-time. We showed how the data collected shows the great promise of a crowdseeding system. Users showed a capacity and willingness to engage at high levels, the technical implementation both at the village level and the processing level were smooth, and the costs required were relatively modest. Moreover initial probes suggest that the data is of high quality. Three types of

22

Figure 5: Robustness Analysis: Three-month Moving Average

Note: The top panel shows the estimated effects making use of CGM wildboot clustered errors. The center panel excludes the first four villages. The bottom panel provide p-values from the randomization inference per period.

23

validation—by our agent in the field, through cross-reporter comparison of reports, and through comparison with survey data—suggest that the data is of high quality. However there is scope for further work to validate the measure. In particular, satisfactory validation requires comparison against a gold standard, which we have not done here. Our second goal was to implement a downstream experiment to illustrate how the data can be used. Between 2007-2011 a donor-funded randomized development intervention was implemented in Eastern Congo. The development project provided villages, including those in South Kivu, with financing to construct local projects such as school rooms or clinics. We take advantage of exogenous variation resulting from the random implementation of this development project to assess whether aid is associated with increased or reduced violence. We implemented Voix des Kivus in such a way that some Voix des Kivus villages received the development project, and others not. Despite the small sample, we find relatively strong evidence for a negative relationship between aid and conflict. Critically, by exploiting the continuous nature of our data, we also highlight the sensitivity of estimates to the timing of measurement in a way that cannot be assessed from traditional cross-sectional data. We close with reflections on the ethical implications of taking a project like this to scale. Throughout the project our most pressing concern was for the security of phoneholders. Over the course of eighteen months Voix des Kivus received thousands of messages many of which were sensitive, including information on various types of abuses perpetrated by different actors. Because of the close connection between individual holders and the project, there was a concern of reprisals against phoneholders for sharing information. Beyond regular monitoring we used three strategies to address this concern. First, we operated in just four villages at first and subsequent villages were added only after one year of careful monitoring. Second, the weekly bulletins were generated in two versions. One version contained sensitive information (with village identifiers), the other did not. The sensitive bulletins were only disseminated to a very restricted set of organizations. Third, the system allowed phoneholders to include a code (1-4) to a message to indicate the event’s level of sensitivity, and with whom the information was to be shared (1. Only with Voix des Kivus, 2. also with close partners, 3. also with MONUSCO, 4. Everybody.) All in all, we faced no incidents that threatened the safety of the phoneholders. However, this might be different when the project is scaled up and the attention of armed groups is drawn to it. For both humanitarian and research purposes a project such as Voix des Kivus becomes truly useful only when it is taken to scale; but those are precisely the conditions which might create the greatest risks. We did not assess these risks because we could not bear them ourselves but given the importance and utility of the data these are risks that others might be better placed to bear.

24

References Anderson, M. B. (1999). Do No Harm: How Aid Can Support Peace - or War. Lynne Rienner, London. Autesserre, S. (2009). Hobbes and the Congo: Frames, Local Violence, and International Intervention. International Organization, 63(Spring 2009):249–280. Autesserre, S. (2010). The Trouble with the Congo. Local Violence and the Failure of International Peacebuilding. Cambridge University Press. Barrios, T., Diamond, R., Imbens, G. W., and Koles´ar, M. (2012). Clustering, Spatial Correlations and Randomization Inference. Journal of the American Statistical Association, (Forthcoming). Bates, R. H. (2009). Prosperity & Violence: The Political Economy of Development. W. W. Norton & Company, New York City, 2nd edition. BBC (2009). Dozens Killed in DR Congo Raids. (13 May 2009). BBC (2011). DR Congo Mass Rape in Fizi: ’170 attacked’. (24 June 2011). Beath, A., Fotini, C., and Enikolopov, R. (2011). Winning Hearts and Minds through Development Aid: Evidence from a Field Experiment in Afghanistan. Working paper. Benjamini, Y. and Hochberg, Y. (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological), 57(1):289–300. Berinsky, A. J., Huber, G. A., and Lenz, G. S. (2012). Evaluating Online Labor Markets for Experimental Research: Amazon.com’s Mechanical Turk. Political Analysis, (Forthcoming). Berman, E., Shapiro, J. N., and Felter, J. H. (2011). Can Hearts and Minds Be Bought? The Economics of Counterinsurgency in Iraq. Journal of Political Economy, 119(4):766– 819. Bertrand, M., Duflo, E., and Mullainathan, S. (2004). How Much Should We Trust Difference-in-difference Estimates? Quarterly Journal of Economics, 1(February):249– 275. Blattman, C. and Miguel, E. (2010). Civil War. Journal of Economic Literature, 48(1):3– 57.

25

Cameron, A. C., Gelbach, J. B., and Miller, D. L. (2008). Bootstrap-based Improvements for Inference with Clustered Standard Errors. Review of Economics and Statistics, 90(August):414–427. Cohen, N. and Arieli, T. (2011). Field Research in Conflict Environments: Methodological Challenges and Snowball Sampling. Journal of Peace Research, 48(4):423–435. Collier, P. (2003). Breaking the Conflict Trap: Civil War and Development Policy. Oxford University Press, New York City. Corstange, D. (2008). Sensitive Questions, Truthful Answers? Modeling the List Experiment with LISTIT. Political Analysis, 17(1):45–63. Coughlin, S. S. (1990). Recall Bias in Epidemiologic Studies. Journal of Clinical Epidemiology, 43(1):87–91. Crost, B., Felter, J. H., and Johnston, P. (2012). Aid Under Fire: Development Projects and Civil Conflict. Working paper. De Nicola, F. and Gin´e, X. (2012). How Accurate Are Recall Data? Evidence from Coastal India. Working paper. De Ree, J. and Nillesen, E. (2009). Aiding Violence or Peace? The Impact of Foreign Aid on the Risk of Civil Conflict in Sub-Saharan Africa. Journal of Development Economics, 88(2):301–313. De Waal, A. (2009). Famine Crimes: Politics & the Disaster Relief Industry in Africa. Indiana University Press, Bloomington. Fearon, J. D. and Laitin, D. D. (2003). Ethnicity, Insurgency, and Civil War. American Political Science Review, 97(1):75–90. Fisher, R. A. (1935). The Design of Experiments. Oliver and Boyd, London. Gelman, A., Hill, J., and Yajima, M. (2009). Why We (Usually) Dont Have to Worry About Multiple Comparisons. Working paper. Green, D. P. and Gerber, A. S. (2008). Get Out the Vote: How to Increase Voter Turnout. Brookings Institution Press, New York City, 2nd edition. Ho, D. E. and Imai, K. (2006). Randomization Inference With Natural Experiments: An Analysis of Ballot Effects in the 2003 California Recall Elections. Journal of the American Statistical Association, 101(475):888–900. ICG (2009). Congo: Five Priorities for a Peacebuilding Strategy. (May).

26

Kling, J. R., Liebman, J. B., and Katz, L. F. (2007). Experimental Analysis of Neighborhood Effects. Econometrics, 75(1):83–119. Lujala, P. (2010). The Spoils of Nature: Armed Civil Conflict and Rebel Access to Natural Resources. Journal Of Peace Research, 47(1):15–28. Lujala, P., Gleditsch, N. P., and Gilmore, E. (2005). A Diamond Curse? Civil War and a Lootable Resource. The Journal of Conflict Resolution, 49(4):538–562. Luyendijk, J. (2009). People Like Us: Misrepresenting the Middle East. Soft Skull Press, New York City. Miguel, E., Satyanath, S., and Sergenti, E. (2004). Economic Shocks and Civil Conflict: An Instrumental Variables Approach. Journal of Political Economy, 112(4):725–753. Nunn, N. and Qian, N. (2012). Aiding Conflict: The Impact of U.S. Food Aid on Civil War. American Economic Review, (Forthcoming). Polman, L. (2011). The Crisis Caravan: What’s Wrong with Humanitarian Aid? Picador, New York City. Posner, D. N. (2010). Using Satellite Imagery of Night Lights to Study Patronage and Politics in Africa. Working paper. Raleigh, C. and Hegre, H. (2009). Population Size, Concentration, and Civil War. A Geographically Disaggregated Analysis. Political Geography, 28(4):224–238. Raleigh, C., Linke, A., Hegre, H., and Karlsen, J. (2010). Introducing ACLED: An Armed Conflict Location and Event Dataset: Special Data Feature. Journal of Peace Research, 47(5):651–660. Small, D. S., Ten Have, T. R., and Rosenbaum, P. R. (2008). Randomization Inference in a Group-Randomized Trial of Treatments for Depression. Journal of the American Statistical Association, 103(481):271–279. UNSC (2012). Report of the Secretary-General on the United Nations Organization Stabilization Mission in the Democratic Republic of the Congo. (May). Verwimp, P., Justino, P., and Bruck, T. (2009). The Analysis of Conflict: A Micro-Level Perspective. Journal of Peace Research, 46(3):307–314. Weidmann, N. B., Rod, J. K., and Cederman, L.-E. (2010). Representing Ethnic Groups in Space: A New Dataset. Journal of Peace Research, 47(4):491–499. Zeitzoff, T. (2011). Using Social Media to Measure Conflict Dynamics: An Application to the 2008-2009 Gaza Conflict. Journal of Conflict Resolution, 55(6):938–969. 27

5

Appendix A: Summary Table Code 11 12 13 21 22 23 24 31 32 33 34 35 36 41 42 43 44 45 46 47 48 51 52 53 54 55 56 57 61 62 63 64 65 71 72 73 74 75 81 82 83 84 85 86 87 88 89 91 92 93 94 95 96 97 98 99

Event Presence of MONUSCO Presence of FARDC Presence of other armed groups Attacks on village by MONUSCO Attacks on village by FARDC Attacks on village by rebel group Attacks on the village by unknown group Civilian deaths (man) Civilian deaths (woman) Civilian deaths (child) MONUSCO deaths FARDC deaths Other rebel group deaths Rioting Looting/property damage Violence between villagers Violence between villagers due to a land conflict Domestic violence Ethnic violence Forced labor by FARDC Forced labor by other army groups Kidnapping by men in uniform Kidnapping by rebel group Arrival of Refugees or IDPs Departure of villagers as IDPs Disappearances Villagers were forced to move Villagers decided themselves to move New outbreak of disease Civilian death due to disease Civilian death due to natural causes Sexual violence against women Sexual violence against men Flooding/heavy rain Large forest or village fire Earthquake Drought Crop failure/plague Complaint against NGO Practice message Unclassifiable issue (followed by text) Security Alert Construction, reparation or rehabilitation of a school or health center Construction, reparation or rehabilitation of a church or mosque Other construction, reparation or rehabilitation Organization of security patrols Work to improve agricultural productivity Funeral Wedding/Other celebrations Visit or meeting organized by national or provincial authorities Visit or meeting organized by territoire authorities Visit or or meeting organized by the chiefdom or locality authorities Visit or meeting organized by the representative of a political party Visit or meeting organized by the Mwami Unclassifiable issue (followed by text) Security Alert

Mean

St. Dev.

Min.

Max.

1.31 1.72 0.18 0.04 0.26 0.20 0.09 0.42 0.44 0.24 0.01 0.17 0.02 0.35 1.10 0.19 0.60 0.11 0.02 0.28 0.08 0.43 0.20 0.16 0.18 0.03 0.05 0.02 1.30 2.22 0.35 0.22 0.25 0.60 0.51 0.19 0.17 1.78 0.20 0.40 1.05 0.13 0.10 0.10 0.50 0.09 0.31 0.58 0.46 0.12 0.04 0.18 0.26 0.17 0.53 0.27

3.12 4.33 0.55 0.29 0.76 0.67 0.38 1.21 1.63 0.94 0.14 0.77 0.23 1.22 2.41 0.81 2.46 0.49 0.14 0.98 0.47 1.14 0.96 0.56 0.71 0.17 0.33 0.22 2.40 3.34 1.64 0.73 1.61 1.25 1.52 0.61 0.51 3.13 0.62 1.83 2.36 0.53 0.38 0.36 1.70 0.47 0.88 1.59 1.24 0.52 0.24 0.58 1.02 0.92 1.32 0.87

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

18 23 3 3 4 5 3 12 16 9 2 6 3 9 19 9 25 5 1 8 6 7 11 4 7 1 3 3 16 21 22 6 18 7 12 4 3 24 4 17 12 4 3 2 12 4 6 11 9 5 2 5 10 10 8 8

Table 3: Summary of Events. Unit is Village-Month. N =211.

28

6

Appendix B: Robustness Cumulative Analysis

Results from the cumulative analysis by wild cluster-bootstrapping the standard error (top panel of Figure 6) are similar to those in Figure 4. The center panel of Figure 6 shows that while the standard errors are large when excluding the first four Voix des Kivus villages, the estimates have the expected sign and are consistant with our overall results. Finally, from the randomization inference (bottom panel of Figure 6) we find that, with the exception of the first two months, all results cluster around the 5 and 10% lines—and most are at or below the solid 5% line. Again, we are thus quite confident that the results in this paper are unlikely to have arisen by chance. Figure 6: Robustness Analysis: Cumulative

29