Culture and Contracts: The Historical Legacy of Forced Labour · lected data on inter-household agriculture insurance agreements. On average, we might expect Hutu and Tutsi to make

Culture and Contracts: The Historical Legacy of

Forced Labour

Arthur Blouin

University of Toronto

March 4, 2020⇤

Can divide-and-rule colonial policy be responsible for contemporary eth-nic tension? This paper empirically investigates the role of a divisive andextractive colonial policy on Hutu-Tutsi discord in Rwanda and Burundi.It shows that Hutu with a family history of subjugation to forced labourby Tutsi chiefs are less trusting of Tutsi today and less willing to partnerwith Tutsi for a cooperative task. This may have implications for agricul-ture insurance agreements since Hutu are more agrarian and Tutsi are morepastoral. Indeed, Hutu with a forced labour family history make fewer inter-household agricultural sharing agreements and experience more insurancefailure.

JEL Codes: J15; L14; Q12; Z13

⇤Thanks to Robert Akerlof, Sascha Becker, Daniel Berger, Gustavo Bobonis, Guilhem Cassan,David Hugh-Jones, Rob Jensen, Victor Lavy, Rocco Macchiavello, Rob McMillan, Guy Michaels,Stelios Michalopoulos, Ted Miguel, Sharun Mukand, Nathan Nunn, Elias Papaioannou, FabianWaldinger and Chris Woodru↵ for very helpful comments and suggestions. I received tremendoussupport from my field teams in both Rwanda and Burundi, who number too many to thankindividually. Thanks as well to seminar participants at Brock University, University of Essex,Harvard University, Lancaster University, McMaster University, l’Universite de Montreal, ParisSchool of Economics, Universitat Pompeu Fabra, Queen’s University, University of Toronto,University of Illinois Urbana-Champaign, l’Universite d’Aix-Marseille, l’Universite de Namur,University of Warwick, and University of Waterloo as well as conference participants at ASREC.I am extremely grateful for financial support from the British Academy and the Centre forCompetitive Advantage in the Global Economy. Contact: [email protected]; 150 St. GeorgeAve., rm. 305. Toronto, Ontario, Canada. M5S 3G7. 416-946-3404.

1. Introduction

One of the most contentious inter-group relationships during the past 30 years

has been between the Hutu and the Tutsi.1 This may be surprising based on

both linguistic similarity (Desmet et al., 2011) and trade incentives (Jha, 2013)

since the Hutu and Tutsi speak the same language, practice the same religion,

and trade incentives have favoured cooperation. One prominent explanation for

the animosity has been that Belgian colonizers imposed arbitrary ethnic divisions,

favoured the Tutsi and drove a wedge between the groups (Lemarchand, 1966).

If true, this suggests drastic, long-run consequences of identity manipulation by

colonial authorities to extract resources. This argument is oft-applied to Belgian

colonialism but similar claims have been used to explain countless instances of civil

strife around the world.2 For instance, Morrock (1973) notes that “divide-and-

rule [is] a policy that has played a crucial part in ensuring the stability -indeed,

the viability- of nearly every major colonial system.” Yet, the ramifications have

not been empirically investigated. Accordingly, this paper aims to measure ethnic

attitudes in Rwanda and Burundi, to see if they are related to one particular

divisive and extractive colonial policy: the subjugation of the Hutu to forced

labour by the Tutsi.

In 1931 Belgian colonial authorities introduced a uniform village co↵ee quota

as part of their rent-extraction e↵orts. This policy mandated the production of

viable export crops, co↵ee in particular, to facilitate taxation by encouraging the

use of money instead of barter. To satisfy the co↵ee quota, Tutsi chiefs were

required to subject Hutu farmers to forced labour, being told “you whip the Hutu

or we whip you” (Watson, 1991). Whether the co↵ee quota was binding depends

on land characteristics. In particular, pre-quota co↵ee production depends on the

suitability of the land for both co↵ee and its alternatives, as well as the value

of each crop. So, whether the quota was binding in a particular region may be

viewed as plausibly exogenous variation in forced labour. The analysis in this

paper therefore investigates whether ancestral land characteristics that hindered

co↵ee production are associated with worse ethnic attitudes for the descendants

of Hutu who would have been subjected to forced labour to meet the quota.

To implement this strategy, I use data from the trust game, family history,

1The post-independence death-toll is well over one-million in Rwanda/Burundi, including thegenocide in Rwanda that saw 70% of Tutsi murdered (Burnet, 2008).

2A few examples include the Hindu-Muslim conflict in India; the Kikuyu-Luo rivalry in Kenya.Both have been blamed on British colonial policy (Tharoor, 2017). The French in south Asiaalso engaged in divide-and-rule strategies (McCoy, 1971), as did America in the Philippines(Morrock, 1973) and the Dutch in Indonesia (Nawawi, 1971).

1

and economic agreements among a sample of Hutu and Tutsi farmers across

Rwanda and Burundi. The trust game in this setting is played face-to-face be-

tween strangers, to observe the rule of thumb people use vis-a-vis trustworthiness

when interacting with someone of a di↵erent ethnicity.3 I then link the trust and

ethnicity data to respondents’ family history.

The empirical strategy is to examine the relationship between the historical

costliness of allocating land to co↵ee and contemporary ethnic attitudes. This

works in part because the data strongly suggests that the costliness of allocating

land to co↵ee influenced the implementation of forced labour in a particular non-

linear way. Several di↵erent measures of forced labour intensity suggest a kink-like

relationship with co↵ee costliness. The empirical strategy exploits this to examine

whether inter-ethnic attitudes are related to the co↵ee costliness in the same non-

linear way. I expect this to be true only for the descendants of the Hutu co↵ee

farmers who would have been most influenced by forced labour (henceforth forced

labour Hutu). This strategy sets up a number of natural falsification tests to help

to identify forced labour as the causal mechanism.

Indeed, I find that forced labour Hutu made trust game o↵ers to Tutsi that were

about 20% lower than other Hutu from their district, and were also less willing to

partner with Tutsi in the lab. Both e↵ects are non-linear in land characteristics in

the same way as forced labour. Meanwhile, Tutsi who would have been targeted

had they been Hutu show similar levels of trust as their peers. Furthermore, forced

labour Hutu only make lower trust game o↵ers to Tutsi; co-ethnic trust appears

una↵ected.

To get a sense of whether these e↵ects are economically meaningful, I col-

lected data on inter-household agriculture insurance agreements. On average, we

might expect Hutu and Tutsi to make good insurance partners because of high

ethnic agricultural specialization (Destexhe, 1995).4,5 In this context, partnering

exclusively within-ethnicity would likely reduce insurance, given an increased like-

lihood that both partners simultaneously experience poor agricultural output. I

find evidence consistent with partner-selection induced reductions in agricultural

insurance.

These results make two contributions to the literature. The first is to show

that divide-and-rule colonial policy contributed to Hutu-Tutsi enmity. Lowes and

Montero (2018) show rubber concessions in the Democratic Republic of Congo

3Group heuristics is one definition of culture (Boyd and Richardson, 2005, Nunn, 2012)).4Tutsi are typically more pastoral; Hutu are more agrarian.5These agreements are also attractive because insurance agreements are not often made out

of village due to monitoring di�culties (Kinnan, 2014).

2

are associated with worse economic outcomes but more pro-sociality. There has

been work showing that slave-trade intensity had long-lasting economic impacts

(Nunn, 2008) as a result of lower generalized trust (Nunn and Wantchekon, 2012).

Work more specific to ethnic attitudes has documented that people behave dif-

ferently towards those of another ethnicity (Knack and Keefer, 1997, Alesina and

La Ferrara, 2002, Lowes et al., 2015) however, empirical analysis of the variation

in this e↵ect has been limited. Notable contributions in this vein include Miguel

(2004), who shows that Tanzanian policy was e↵ective in bringing together eth-

nically diverse communities; Shayo and Zussman (2011) who show that terrorism

influences out-group bias; and Voigtlander and Voth (2014) who show that in-

frastructure investment in Germany was positively associated with Naziism.6 The

second contribution is an analysis of how forced labour and inter-ethnic distrust

relate to economic agreements. The literature on the economic importance of

general trust is large,7 however, microeconomic evidence of how inter-ethnic trust

influences economic interactions in particular ways is sparse. Notably Hjort (2014)

shows that poor inter-ethnic co-operation in a Kenyan firm generated production

ine�ciencies.

2. Historical Background

There is little evidence of Hutu-Tutsi conflict prior to the mid-19th century. They

lived in segregated communities (Nyirubugara, 2013) that were economically and

politically undeveloped. Communities relied on prominent local lineages for public

goods, with non-monetary goods being voluntarily exchanged for protection and

representation (Newbury, 1988). This traditional clientship transformed for the

first time under king Rwabugiri (r. 1863� 1895) in a few ways.

Rwabugiri was Tutsi, and ushered in a wave of Tutsi chiefs, even in traditionally

Hutu villages. Preferential treatment towards Tutsi citizens meant that Hutu

took a subservient role in society for the first time (Nyirubugara, 2013). Second,

Rwabugiri implemented mandatory taxation to replace the lineage system which,

for the most part, was paid using cattle (Umuheto or Ubuhake).8 Since Hutu did

not traditionally keep cattle, mandatory payment with labour (Ubureetwa) was

a substitute. “Of the various services performed for chiefs, Ubureetwa ‘was the

6Notably Eifert et al. (2010), Bazzi and Gudgeon (2018) study the intersection of ethnicconflict and politics.

7For a review of the literature see Alesina and LaFerrara (2005); for one on culture andinstitutions see Alesina and Giuliano (2015).

8In Umuheto a client would purchase ‘protection’ from the Chief in exchange for a cow.Ubuhake involved the loan of pasture land for protection (Newbury, 1988).

3

most hated and humiliating.’ It symbolized the servitude of the Hutu.” (Newbury,

1988).

When Belgium took control of the colony after World War I, their priority was

was “modernization.” Their immediate goals were twofold: to phase the economy

into the monetary system and away from bartering; and to abolish what they

believed to be antiquated local institutions, likeUmuheto, Ubuhake and Ubureetwa.

Accordingly, they made several changes, two of which inadvertently impacted

Hutu-Tutsi relations. First, they scaled back forced labour requirements under

Ubureetwa, with the plan to phase out the practice completely. In 1927 “Hutu

labour corvee or forced labour requirement due to Tutsi nobles is reduced by

Belgian administrators from two days in five to one day in seven” (Page and

Sonnenburg, 2003). Concurrently, Belgium pursued an aggressive export strategy.

One pillar of this strategy was the imposition of uniform agricultural requirements

on each village, designed to boost exports (Bonaventure, 2010). Co↵ee was the

export crop most suited to production in the region, and a regulation in 1931

made co↵ee a required crop (Page and Sonnenburg, 2003, page 664).

“Belgium also decided to initiate co↵ee production on Hutu land under

the corvee system: compulsory labor demanded by a lord or king.

Hutus were subjected to ten lashes a day to ensure a solid work ethic,

in case these ‘inferior’ people strayed from their assigned duties. By the

time of Rwandan independence in 1962, the Hutus were a subjugated

population, manipulated by both their fellow Rwandans and colonial

powers.” (Mendis, 2014)

The uniform co↵ee quotas had the largest impact on the regions that were least

suitable for co↵ee. So why did Belgium believe that growing co↵ee in these regions

was a good idea? While co↵ee was not heavily grown at the time, the fact was

that almost all regions could (and arguably should) grow at least some co↵ee. In

fact in Rwanda and Burundi very few villages are completely unsuitable for co↵ee

according to data from Fischer et al. (2012) (figure A1).9

Co↵ee quotas influenced agriculture dramatically after 1931 (figure A2).10 Cof-

fee went from being one of twenty modestly produced crops to dominating the

industry, mainly replacing subsistence crops like manioc and maize. The pressure

from the chiefs to meet the quotas was especially burdensome for Hutu co↵ee

9Data are from FAO-GAEZ: description in section 3.C.10This is based on data transcribed from Belgian colonial yearbooks. Description in Appendix

B.5.

4

farmers who were best positioned to scale production in regions where co↵ee was

not already heavily produced.

“This was ubureetwa, one ‘imposed specifically on Hutu’ and left un-

reformed because o�cials argued that to do away with it would be to

‘undermine the chiefs’ authority over the population.’ The chief who

came out of the interwar period was expected to enforce and supervise

obligatory cultivation of food exports...and even to become majority

co↵ee producers by using corvee labour.” (Mamdani, 2014)

Because Belgium simultaneously abolished traditional labour requirements in some

regions, and used them to suit their export strategy in others, I limit the analysis

to the study of large or small changes in forced labour intensity in various places,

and the associated di↵erences in ethnic relations. In fact, it may not be surprising

to find that Hutu inter-ethnic attitudes generally improved in regions without

forced labour as the result of the abolishment of exploitative practices from the

Rwabugiri era.

3. Data

The analysis primarily relies on survey, lab-in-the-field, and geographic informa-

tion system (GIS) data. The outcomes and many of the biographical controls come

from a survey and a set of lab exercises implemented in Burundi and Rwanda in

2013. Data was collected from a total of 869 farmers from 143 di↵erent villages.11

We brought together 4-5 individuals from each of 4-5 di↵erent villages in any given

data collection session, and conducted two such sessions per day, of about 20 peo-

ple each. Villages were primarily in the south in Rwanda, in part to ensure that

some Tutsi would attend,12 while in Burundi most of the country was covered.

Protocol details are in Appendix B.2.

Measurement is a challenge in this context for a variety of reasons. First, the

Rwandan government does not typically allow research that mentions, or even

primes ethnicity, so we must rely on interactions in the lab. Second is the ob-

servation of ethnicity itself - both by the researcher and the respondents. Unless

both are able to make reasonable inferences about ethnicity, even the lab-based

inter-ethnic attitudes measures will be unreliable. Finally, forced labour variation

exists at the ancestral location, and measuring this variation necessitates collect-

11A map of villages is in figure B112The approximate ethnic distribution is in figure B2

5

ing family migration and crop production data going back generations, which is

imperfectly known by respondents.

3.A. Data Challenge 1: Measuring Outcomes

i) The trust game The trust game is a standard method to elicit ethnic attitudes

(Fershtman and Gneezy, 2001). In our implementation, a randomly matched

pair sat down with an enumerator to play face-to-face.13 Pairs played only one

round to remove strategic considerations. Because I only allowed pairings between

people who were from di↵erent villages and for whom the field team confirmed

had never met, respondents could only use decision heuristics to determine the

trustworthiness of their partner.

The trust game worked as follows: one partner was randomly assigned to be

‘the sender’ and the other was assigned to be ‘the receiver.’ The sender was given

600RWF (approximately $1 USD),14 which they could share with their partner

or keep to themselves.15 They were given 6 $100 notes ($200 in Burundi) of

Monopoly money, and asked to pass as much money as they wanted to their

partner. Whatever they chose to share was matched by the enumerator and given

to the receiver, who then decided how to share that sum. The amount o↵ered by

the sender is used in the analysis as a measure of how much the sender trusts the

receiver. The intuition is that if the sender completely trusts the receiver, they

should o↵er all of the money, since the receiver can then return all or more of the

original 600RWF. If the sender does not trust the receiver (i.e. plays the Nash

equilibrium) they should o↵er nothing to protect themselves against the possibility

that the receiver returns nothing. In practice, nobody in Rwanda and two people

in Burundi did this. On average just under 300RWF out of 600RWF was shared

(Table B1).

ii) Partner Selections The partner selection task was implemented between the

survey and lab exercises. Respondents were told that they would be partnered with

someone, and that their ability to cooperate with their partner could increase their

payo↵. They were given a chance to submit a list of five people that they would

prefer to be partnered with for this exercise. Respondents looked around the room

at all of the people in their session (having not yet interacted with anyone) and

listed the ID tag numbers of their choices. ID numbers were matched to presumed-

13Enumerator instructions appear in Appendix B.314Represented by Monopoly money. This is 1200BUF in Burundi.15Mean daily wage for the sample was about $1.05.

6

ethnicity ex post. I am interested in the share of people the respondent selected

that are not in their ethnic group.

(1)

Preference for inter-ethnic partner =number of choices from other ethnic group

min{5, total other ethnic group}

The numerator is the number of people on the five person list that are not in the

respondent’s ethnic group. On its own, this reflects both respondent preferences

and the number of people from another ethnic group at the respondent’s session.

For instance, if there was only one Tutsi at a session, then a Hutu selecting one

Tutsi would not indicate anti-Tutsi preferences. To account for this I divide by

the smaller of the number of out-group members available to be selected, and the

total number of selections made.

iii) Contracts The third outcome focuses on real world informal contracts. The

idea was to examine a context where inter-group agreements may be beneficial,

and to assess the outcomes in these agreements. To this end, I surveyed respon-

dents about inter-household agricultural insurance agreements. We may expect

agricultural shocks to be less correlated across ethnic groups than within them,

since Tutsi are largely pastoral and Hutu largely agrarian. I focus primarily on

the number of agreements, and what I will label the insurance failure rate, which

is equal to (1 + defaults)/(1 + agreements). By default, I mean an expected gift

/ transfer that never materialized. The insurance failure rate is equal to one if

no agreements are made, since there would be no transfer in the event of a shock.

The intent is to measure how insured they are, taking account of both not making

agreements, and making low quality agreements. The insurance failure rate would

be small for individuals with agreements that were never defaulted on.

Notably, the insurance failure rate is conceptually much di↵erent than a vari-

able measuring contractual default, or contractual failure. Rather than considering

the breakdown of a relationship that had already formed, here my interest lies in

understanding how informally insured an individual is. Of less interest is dis-

tinguishing between a reduction in insurance due to a breakdown of an existing

relationship, or because the anticipated breakdown of the relationship was su�-

ciently high that no relationship was formed in the first place. The reason for this

is that low trust could plausibly manifest in either outcome, so it seems unclear

ex ante which may drive results.

7

3.B. Data Challenge 2: Ethnicity

In Burundi, asking about ethnicity is permitted, and enumerators did so there. In

Rwanda it is typically not permissible to ask individuals about ethnicity - either

their own, or of people they interact with. So, I use the ethnicity proxy described

in Blouin and Mukand (2019): the eligibility for a genocide survivor fund, which

is available only to Tutsi in genocide regions. In combination with our sampling

strategy, which was limited to regions where the fund operates, a positive response

to the eligibility question signifies Tutsi status (see Appendix B.4 for more details).

In particular, individuals in Rwanda may receive funds from the government

under a variety of di↵erent programs, including a fund called the FARG (“Fonds

d’Assistance pour Rescapees du Genocide”). FARG is exclusively targeted to fam-

ilies who are ‘survivors’ of the 1994 genocide - a category that coincides almost

perfectly with the Tutsi in the places where FARG gives money. We asked par-

ticipants in Rwanda their sources of income from various aid programs (including

FARG). For each aid program they were asked whether they (a) were aware of it

(b) eligible for it, and (c) received money from it. In order to make the question

on FARG eligibility less salient, the section about FARG eligibility was part of a

subsection about income sources from government funds, which in turn was part of

a larger income module. In Burundi ethnicity is directly measured, so robustness

is evaluated using only that sample.

Also important is the identification of ethnicity by the respondents themselves.

Respondents play in either a co-ethnic or inter-ethnic trust game, and cannot be

primed on which it is. So, the game is only a reliable measure of ethnic attitudes

if respondents can infer ethnicity based on a brief interaction with their partner.

There are stereotypical physical di↵erences between Hutu and Tutsi (Gourevitch,

1998), however misattribution almost certainly occurred. Measurement error is

likely to bias results towards zero as it seems unlikely that attribution errors are

correlated with family forced labour history.

3.C. Data Challenge 3: Variation in Forced Labour

To measure forced labour exposure I need the locations of the ancestors of re-

spondents during the colonial era. Accordingly, the survey included a module

on family migration history going back three generations. I use the location of

the parents birth, which typically gives the location of the grandparents during

the colonial era. Since there is some expectation that co↵ee farmers may have

been disproportionately targeted for forced labour, I also asked about parent and

8

grandparent crop production, and rely on respondent estimates of grandparent

co↵ee cultivation intensity (share of land devoted to co↵ee).

Grandparent location was geocoded and matched to crop suitability data (fig-

ure B1). The land characteristic data comes from Fischer et al. (2012) (FAO

GAEZ) who provide GIS data on the potential produceable tonnes per hectare

for each crop across the globe.16 This was matched to colonial price data for each

crop, which was transcribed by hand from Belgian colonial yearbooks (details in

Appendix B.5).

Obtaining a measure of actual forced labour was also a challenge. It was a

policy with a strong ethnicity and power element, and after consulting with local

partners, it was determined that it would negatively prime ethnicity, so it was

not mentioned in the survey. Instead, to get variation in actual forced labour, I

scraped Google Books for any digitized colonial-era documentation in French that

referenced both a particular district and some of the language typically used to

discuss forced labour. The aim was to target the reports that were written by

administrators in the region as closely as possible.

I investigate three terms that, after reading through reports by Belgian ad-

ministrators (e.g. Rwa (Scan Date: October 17, 2012)), appeared to be frequently

used in reference to forced labour. They are (1) Ikiboko which is a local word for

whips with a hippo hide. This was the punishment for refusal to comply with

labour requirements; (2) corvee which is french for forced labour; and (3) Presta-

tions Coutumieres Dues which is french for traditional labour requirements. For

each term, I collected the share of documents about each grandparent-district that

mentioned the search term.

4. Empirical strategy

I am interested in the average causal e↵ect of forced labour on attitudes for those

with a family history of forced labour. Consider the case of trust, and denote inter-

ethnic trust as T and forced labour as FL. Two dimensions may have determined

a respondent’s forced labour family history.

The first, ✓i, captures that some individuals (denoted i) had grandparents

who were more likely to be selected into forced labour by the chief, within each

village. For instance, Tutsi chiefs may have selected people disliked by the Tutsi

community in a village. ✓i is endogenous, and in any case is not directly observed

16Estimates are available for various input levels. To match historical conditions for Rwanda-Burundi, data chosen was for low-input and rain-fed conditions. The resolution is at the 5arc-minute level.

9

because, as previously mentioned, this type of ethnically divisive question was

problematic to collect in this context. I therefore need to proxy for ✓i, and propose

considering that co↵ee farmers were mostly selected to work to meet the quota.

The variable Ci denotes whether the respondent’s grandparents produced any

co↵ee on their own land, and is used as a proxy for ✓i.

One concern may be whether respondents to reliably know what their grand-

parents grew, and even if so, whether they are recollecting co↵ee production before

or after forced labour. This is quite important, since we might expect Hutu co↵ee

production under forced labour to be endogenously related to ethnic attitudes.

However, it would be surprising if co↵ee production independent of forced labour

was related to ethnic attitudes. A suggestive test is available. If respondents rec-

ollect whether their grandparents grew co↵ee prior to forced labour, we should

expect that relative co↵ee suitability is positively correlated with co↵ee produc-

tion. If on the other hand, they recall (the more endogenous) co↵ee production

under forced labour, we would expect either no correlation, or possibly even the

reverse. Table B3 reveals that observed Hutu co↵ee production was much less

common in forced labour regions, consistent with recollection of co↵ee production

prior to the implementation of forced labour.

The second main factor determining forced labour, µlgp , captures that some

grandparent locations (denoted lgp) were exposed to forced labour and others were

not. These may also have been selected based on pre-existing inter-ethnic trust.

However, land characteristics may be related to FL through µlgp if forced labour

was used to meet the co↵ee quota, as suggested by the history literature. All

villages have some level of unconstrained equilibrium co↵ee production, which

may be above or below the quota. This production is a function of the region-

specific returns to co↵ee relative to all other crops. Using the FAO data on crop

suitability, I observe produceable tons per hectare for each crop (denoted qFAO

lgp,s for

each crop s), which I matched to colonial crop prices (denoted ps).17 This allows

me to compute returns to producing each crop s as: ⇡lgp,s = qFAO

lgp,s ps.

Now, consider how crop returns (⇡lgp,s) might have related to the regional dis-

tribution of forced labour (µlgp). Hypothetically, if the best crop in the region

returned 500% of what co↵ee returned, then co↵ee would be unlikely to be pro-

duced naturally, and the quota would increase production dramatically, largely

using forced labour. On the other hand, if the best co↵ee alternative returned

only 20% of what co↵ee returned, the quota likely would not bind, so less forced

17source: M. le Premier Ministre (1927(-1945))

10

labour would be required. Accordingly, consider:

(2) ⇧lgp =max{⇡lgp,s|s 6= c}

⇡lgp,c

Where c denotes co↵ee and s can be any crop.18 A histogram of this variable can

be seen in figure B3. The figure suggests that the median value of ⇧lgp is around

0.9, and of the 122 districts 62 are above ⇧lgp = 0.9.

I hypothesize that cov(µlgp ,⇧lgp) > 0 because the higher the returns to the

best non-co↵ee crop relative to co↵ee, the less likely co↵ee is to be grown prior

to the quota, the more likely the quota was binding, and the more forced labour

we might expect in that location. Table C1 examines the relationship between

forced labour and crop returns. Indeed, there is a positive correlation between the

costliness of allocating land to co↵ee and forced labour / mistreatment for each

measure of forced labour.

Each of the three forced labour measures, as well as the average of the three,

are plotted against the costliness of a co↵ee quota in figure 1. Interestingly, each

plot produces the same non-linear pattern.19 All four measures are increasing in

co↵ee costliness, consistent with table C1. But for each, the increasing relationship

is driven almost exclusively by values of ⇧lgp starting just below 1; with the slope

consistently turning from near-zero to strongly positive between 0.9 and 1. There

is a vertical line in each graph at ⇧lgp = 0.975, corresponding to where the slope

turns positive for the mean of the three measures.20 This is not a stark threshold,

in panels (a) and (b) of figure 1 the increasing relationship emerges a bit before this

value, and in panel (c) a bit after. I interpret ⇧lgp = 0.975 as an approximation

of the typical point where the quota binds, and define lgp = (⇧lgp > 0.975). I

define ⇣i,lgp,lr = Ci,lgp,lr · lgp , which allows the main specification used throughout

the analysis to be written as:

(3) Ti,lgp,lr = �0 + �1⇣i,lgp,lr + �2 lgp + �3Ci,lgp,lr +⇤lr +Xi,lgp�+ ✏i

Where ⇤lr represents respondent district fixed e↵ects and Xi,lgp is a matrix of

individual level covariates such as gender, age, risk aversion, raven score, and

enumerator fixed-e↵ects, all of which might influence observed attitudes.21 It also

18The other crops and their relative frequency appears in table B2.19The analogous binscatter plots can be seen in figure C1.20This is not driven by outliers. There are 51 districts out of a total of 122 with ⇧lgp > 0.975,

see figure B3.21A figure showing the variation in each of gender, age risk and raven score appear in figure

C2. There are di↵erences, especially in gender and age near the ⇧lgp > 0.975 but none are

11

captures characteristics of the grandparent location that might be correlated with

⇧lgp , like suitability of each crop, a dummy for the best produceable crop and

distance to the capital. The main identifying assumption required to interpret

�1 causally is that the only way ⇣i,lgp,lr matters for co↵ee farmers in non-co↵ee

regions is through forced labour. I take two strategies to assess the validity of this

assumption.

The first focuses on the non-linearity in the relationship between co↵ee returns

and forced labour. If ethnic attitudes begin to decline just before ⇧lgp = 1 for

descendants of Hutu co↵ee farmers, and there is no di↵erence to the left of this

threshold, that lends credence to the idea that it had to be forced labour generating

the di↵erence. The logic is similar to a kink-design, although in this case I have

neither the data nor the sharp boundary to estimate an actual kink.

The second strategy is to examine the Tutsi as a falsification group. I examine

the attitudes of Tutsi with grandparents from a forced labour village, who would

have likely been selected for forced labour given their agricultural activity, had

they been Hutu. If there are strong selection e↵ects into co↵ee production in

a region where co↵ee is a secondary crop, then it may be reasonable to assume

that these pressures exist similarly for both Hutu and Tutsi. Of course Tutsi

and Hutu have di↵erent agricultural profiles, and may have faced di↵erences in

migration restrictions, so this strategy is limited to the extent that Hutu and Tutsi

experienced di↵erent selection forces.

5. Main Results

5.A. Trust Game Results

Given that that mistreatment by the Tutsi towards the Hutu appears to have

been most likely in regions where the co↵ee quota had the most bite, it seems

reasonable to expect respondents with grandparents that grew co↵ee in these

villages to exhibit the least inter-ethnic trust. This hypothesis is investigated

in table 1 panel A, which presents evidence of di↵erential trust game o↵ers by

the descendants of farmers believed most likely to have been exposed to forced

labour. Columns 1 and 2 of table 1 panel A suggest that forced labour Hutu make

trust game o↵ers that are about 20% lower than other Hutu who currently live

in the same districts, but whose grandparents were likely not exposed to forced

significant.

12

labour.22 Columns 1 and 2 di↵er by the controls included. In column 2 we include

income and education, but not in column 1 because education and income have

been shown to be endogenous in other contexts.23

A more fundamental concern might regard the type of individual who grows

co↵ee in regions that are less suitable for co↵ee in the first place, so I take a few

approaches to deal with that.24 First, consider the Tutsi who similarly grew some

co↵ee in these same regions. Tutsi o↵ers can be seen in panel B, columns 1 and

2. If anything, these Tutsi made larger o↵ers than their non-co↵ee counterparts,

though neither estimate is close to statistically ruling out zero.25 The co-ethnic

sample is an opportunity for a similar exercise. If we thought - independent of

forced labour - that people who grew co↵ee in places where co↵ee was a secondary

crop were more marginalized, then it might be natural to think of this as a general

trust e↵ect rather than as a specifically inter-ethnic trust e↵ect. However, columns

3 and 4 of panels A and B show no evidence that co-ethnic trust has changed for

either Hutu or Tutsi. However, the Tutsi-Tutsi estimate is a less conclusive null

result, since the point estimate is large (although, positive) and the sample is

small.

Another approach to investigating selection into treatment is to exploit the

non-linearity between land characteristics and forced labour in figure 1. It may

well be that as co↵ee becomes a less viable option, that di↵erent types of farmers

grow it. However, it seems reasonable to assume this e↵ect to be linear in land

characteristics. Or at least, it seems unlikely that - if it had nothing to do with

forced labour - it would become strongly non-linear just around the point where

forced labour ramps up. Accordingly, di↵erences in inter-ethnic trust beginning

near this threshold might help to empirically rule out an e↵ect generated purely

by di↵erences the composition of co↵ee farmers in di↵erent villages.26

22Appendix C.1 shows that trust game di↵erences do not appear due to altruism or reciprocity.They also demonstrate that the results are not driven by di↵erential trustworthiness.

23Bobonis and Morrow (2014), Lowes and Montero (2018) show that forced labour causeslower education; Dell and Olken (2019), Lowes and Montero (2018) show forced labour influenceswealth.

24The ethnicity proxy in Rwanda and the genocide in Rwanda may also be concerns. I examineresults separately for Rwanda and Burundi (Appendix C.2), and show robustness of each resultto the Burundi-only sample to rule-out Rwanda-specific concerns. I also investigate the role ofthe genocide (Appendix C.3).The estimates are similar, but the standard errors are smaller, asexpected. Still though we fail to reject that the e↵ect is zero.

25Note that the inability to rule out zero could be due to power issues, as the Tutsi sample isvery small. To check this I run a pooled-by-ethnicity version of the specification in table C3.

26We might also expect compositional di↵erences if people fled forced labour. This seemsunlikely based on migration restrictions. Section 6.C shows, indeed, Hutu were less likely toleave forced labour regions, but forced labour Hutu were not di↵erentially likely to leave.

13

In figure 2 panel (a) I plot the parameter estimates from the Hutu-Tutsi sample

(y-axis) for various possible thresholds in ⇧lgp to define a ‘forced labour village’

(x-axis). Each point denotes the estimate and confidence intervals for �1 from the

main regression equation. Estimates are always from separate regressions. Each

regression uses a sample-window around the threshold being considered. This is

because, for example, if I consider the false-threshold of ⇧lgp = 0.7, I do not want

respondents that I consider actually treated, e.g. those with ⇧lgp = 1.2, to generate

a negative estimate where none is expected. Using a window centred around each

threshold means that we might expect movement in the estimate slightly below the

actual threshold defining forced labour, since it is unavoidable to start including

some actually treated people in the treatment group for thresholds near, but below

the actual threshold. Note that I use the phrase ‘actual threshold’ for simplicity,

but this is not precise, so it may be reasonable to consider any e↵ects beginning

nearby in either direction as plausibly related to forced labour.

Estimates are stable and close to zero, up to a threshold of approximately

⇧lgp = 0.9, which is just below where there was consistent movement in forced

labour (again, see figure 1). For thresholds beyond ⇧lgp = 0.9, estimates become

increasingly large. This is consistent with low observed inter-ethnic trust having

been caused by the forced labour or associated mistreatment. Panel (b) shows the

same exercise for the Tutsi sample, where we fail to observe a similar pattern.27

5.B. Partner Preference Results

One disadvantage of the trust game is I get an estimate only for the subset of

respondents who played with a respondent from the other ethnic group. This

may not be a first order concern since partners were randomized, but since the

partner preference measure is available for all Hutu, I can use that variable to

both address the validity of the randomization, and show robustness to another

measure of inter-ethnic attitudes. Panel A of table 2 shows forced labour Hutu

are 13p.p. less likely to choose a Tutsi partner in the full Hutu sample (columns 1

and 2), and 18p.p. less likely in the inter-ethnic trust-game sample of individuals

who were assigned to the inter-ethnic trust game instead of the co-ethnic trust

game (columns 3 and 4).28 The results highlight the robustness of the e↵ect on

attitudes to both di↵erent measures and samples.

Taking a look at the main robustness exercises, it appears that the non-linearity

27I examine robustness to a trust survey question in Appendix C.428I also check for preference for other observable characteristics. However I find no evidence

of di↵erential preference on any other dimension (Appendix C.5)

14

of the e↵ect is quite similar to both the inter-ethnic trust e↵ect and the descriptive

hockey-stick pattern between quota-costliness and actual forced labour. There is

a significant negative e↵ect for definitions of a forced labour village only at or

beyond the threshold ⇧lgp = 0.9 (figure 2 panel b and c). As before, this lends

some credibility to the causal interpretation. Similarly, there appears to be no

e↵ect on attitudes for Tutsi with grandparents who likely would have been exposed

to forced labour had they been Hutu (table 2 panel B). This is true for for both

the full (columns 1 and 2) and inter-ethnic trust-game samples (columns 3 and 4).

6. Analysis of Related Hypotheses and Robustness

6.A. Insurance Agreements Results

Based solely on lab measures it is often di�cult to assess magnitudes, because the

context is artificial. To address this, I examine agricultural insurance agreements

traditionally made between Hutu and Tutsi. I start by looking at the number of

agreements made, to see if the e↵ect on trust is large enough to have real world

consequences. That said, it should be noted that there are a number of possible

interpretations of the insurance results, especially given that these outcomes are

not tied directly to insurance partners.

Table 3 panel A shows that forced labour Hutu make fewer agreements (col-

umn 1 and 3), but not Tutsi with grandparents who would have been exposed

had they been Hutu (column 2 and 4). If we think of the inter-ethnic agree-

ments as being more valuable, then the fact that forced labour Hutu make fewer

agreements is consistent with their lower inter-ethnic trust. Insurance could be

low either because individuals make fewer agreements, or because they experience

more default. Accordingly I also look at the insurance failure rate which captures

both sources (panel B). Forced labour Hutu are less insured using that measure

as well (columns 1 and 3), but observationally equivalent Tutsi are not (columns

2 and 4). Suggestive evidence implies that increased defaults are primarily driven

by a selection into less suitable Hutu-Hutu agreements (Appendix C.6).

Again, this section is suggestive of real world implications in the sense that the

results are consistent with possible within-ethnicity sorting. Without collecting

data on real world insurance partner ethnicity, it is perhaps not possible to pin this

mechanism down entirely and rule out alternative explanations that may also be

consistent. Nevertheless, we are unable to rule out reduced agricultural insurance

due to forced labour based on the data available.

15

6.B. Genocide

The genocide is clearly a related and very important issue, and there are a number

of reasons why it has not featured more prominently in the analysis so far. Perhaps

the most pressing is the way that the ethnicity proxy in Rwanda was collected.

Because the ethnicity proxy in Rwanda measures whether individuals are ‘genocide

survivors’, the sampling procedure in Rwanda targeted genocide regions to reduce

the likelihood that any Tutsi would claim to be ineligible for the fund on the

basis that their region did not experience genocide violence. The implication

of this is that in the data I have only intensive margin variation in genocide,

which could lead to imprecise, inconclusive estimates. Beyond that, it is not

entirely clear that controlling for the genocide is a reasonable approach, given it

is plausibly endogenous. Since the sampling strategy generated data that is not

particularly conducive to exploring genocide violence, and controlling for it may

have introduced endogeneity to the main empirical model, a consideration of the

genocide has been relegated to a robustness exercise.

That said, I would be remiss to not discuss it. A more complete examination

of the implication of the genocide appears in section C.3. To summarize those

results, each of the main estimates are very similar in Rwanda-only and Burundi-

only samples, suggesting that the genocide is not the main mechanism (table

C5). Furthermore, controlling for the genocide does not meaningfully impact the

results, and if anything the genocide is positively related to attitudes in Rwanda

(tables C6 and C7). The positive correlation between attitudes and violence is

interesting, as it is consistent with work on the pro-sociality of violence (Bauer

et al., 2016), but extends results to the inter-ethnic context. Still though, given the

result is not particularly robust, I hesitate to read too much into it. For instance,

estimates in table C7 are not precise enough to be conclusive for either partner

preferences or trust o↵ers.

Genocide persecutions do, however, seem to have been more intense in forced

labour regions by about 110 people per sector (p-value< 0.01). It seems reasonable

to expect the genocide to potentially be more intense in regions with a history of

ethnic exploitation, as authorities openly referenced historical Tutsi repression to

mobilize violence (Straus, 2013). However, again, since the result is not robust,

we leave the issue as inconclusive and leave a more exhaustive investigation for

future work.

16

6.C. Migration

Migration could be another concern. If a substantial number of Hutu fled forced

labour, then there may be concern that those who were unable to flee forced labour

might be systematically di↵erent than the other (non-forced labour) Hutu in their

regions. Colonial administrators were reportedly quite aware of the threat of

flight, and responded with migration restrictions (352, 1925a,b) and punishments

for fleeing (Newbury, 1978). However, both were inconsistently applied, and it is

very di�cult to find clear and consistent formal rules or punishments regarding

forced labour evasion (Butamire, 2012).

We can, however, check whether there was an exodus from forced labour regions

in the data. In non-forced labour regions I expect individuals to migrate as usual,

however likely not into forced labour regions. Migration restrictions for Hutu may

have applied to (1) all Hutu; to (2) Hutu in forced labour regions; to (3) just

forced labour Hutu; or (4) not enforced. The biggest concern would be di↵erential

migration by forced labour Hutu, which could result from either case (3) or (4). In

either of those cases forced labour Hutu would plausibly be unobservably di↵erent

from their comparison group: the other Hutu from those same places.

Table C2 examines whether co↵ee farmers whose families started in a forced

labour region are more likely to now live elsewhere. We see in columns 1 and 3 that

all Hutu from forced labour villages were much less likely to migrate, and there

was no di↵erential e↵ect for forced labour Hutu. So, neither of the concerning

cases seem to be the case for Hutu. First, migration restrictions seem to have

been enforced, essentially shutting down all Hutu migration out of forced labour

villages. Second, the migration restrictions did not seem to have influenced forced

labour Hutu di↵erentially.

For Tutsi, the estimates warrant a bit more caution. Already, the Tutsi fal-

sification exercise had been plagued by di↵erent selection pressures into co↵ee as

a result of the di↵ering agricultural profiles of Hutu and Tutsi. Additionally, ta-

ble C2 columns 2 and 4 suggest that Tutsi may have faced less severe migration

restrictions. As a result, Tutsi who would have been selected for forced labour

had they been Hutu fled those regions di↵erentially. This is consistent with Tutsi

observing the forced labour imposed on the Hutu, being afraid that forced labour

would expand to Tutsi (historically this had been discussed (Jefremovas, 2002)),

and fleeing the worst areas while they still could. The results suggest that mi-

gration restrictions may have been more severe for Hutu, a conclusion that is not

altogether surprising.

17

Together the results imply that some caution is warranted in interpreting the

Tutsi falsification estimates. However, the main Hutu estimates seem less prob-

lematic than they might have been. Colonial administrators do appear to have

been e↵ective in limiting Hutu migration, mitigating the ability of forced labour

Hutu to select out of treatment.

7. Discussion

Divide-and-rule is still considered by some as a viable strategy for America to

pursue in the Middle East (Pernin et al., 2008, page 101), despite long-standing

arguments that variations of divide-and-rule have been responsible for countless

civil strifes around the world in the colonial context. Notable cases include Hindu-

Muslim tension in India, sectarian rivalry in Ireland, and Kikuyu-Luo conflict in

Kenya. However, there has been little to no empirical evidence to document the

long-run implications of this type of policy on inter-group relations. Rwanda and

Burundi have experienced both divisive colonial policies and extreme inter-group

conflict. This paper has attempted to demonstrate that these may be causally

related, and that even 50 years post-independence, the implications of divide-and-

rule colonialism for inter-group relations continues to have consequences.

18

References

Colonial era burdens and opportunities, government taxes. Center for ResearchLibraries, Scan Date: October 17, 2012.

RWA 352. Resident of kisenyi to resident of ruanda. Archives Africaines (Belgium),May 1925a.

RWA 352. Resident of gatsibu to the resident of ruanda. Archives Africaines(Belgium), May 1925b.

Alberto Alesina and Paola Giuliano. Culture and institutions. Journal ofEconomic Literature, 53(4):898–944, 2015. doi: 10.1257/jel.53.4.898. URLhttp://www.aeaweb.org/articles.php?doi=10.1257/jel.53.4.898.

Alberto Alesina and Eliana La Ferrara. Who trusts others? Journal of PublicEconomics, 85(2):207–234, 2002.

Alberto Alesina and Eliana LaFerrara. Ethnic diversity and economic performance.Journal of Economic Literature, 43(3):762–800, September 2005. doi: 10.1257/002205105774431243.

Michal Bauer, Christopher Blattman, Julie Chytilova, Joseph Henrich, EdwardMiguel, and Tamar Mitts. Can war foster cooperation? Journal of EconomicPerspectives, 30(3):249–274, Aug 2016. ISSN 0895-3309. doi: 10.1257/jep.30.3.249. URL http://dx.doi.org/10.1257/jep.30.3.249.

Samuel Bazzi and Christopher Blattman. Economic shocks and conflict: Evidencefrom commodity prices. American Economic Journal: Macroeconomics, 6(4):1–38, 2014.

Samuel Bazzi and Matthew Gudgeon. The political boundaries of ethnic divisions.Technical report, National Bureau of Economic Research, 2018.

Arthur Blouin and Rocco Macchiavello. Strategic default in the internationalco↵ee market. The Quarterly Journal of Economics, 134(2):895–951, May 2019.

Arthur Blouin and Sharun Mukand. Erasing ethnicity? propaganda, nation build-ing, and identity in rwanda. Journal of Political Economy, 127(3):1008–1062,June 2019.

Gustavo J Bobonis and Peter M Morrow. Labor coercion and the accumulationof human capital. Journal of Development Economics, 108:32–53, 2014.

Mureme K. Bonaventure. Manuel d’Histoire du Rwanda a l’epoque coloniale: suiv-ant le modele de Mgr Alexis Kagame. Harmattan, 2010.

Robert Boyd and Peter Richardson. The Origin and Evolution of Cultures. OxfordUniversity Press, New York, 2005.

19

Jennie E Burnet. Gender balance and the meanings of women in governance inpost-genocide rwanda. African A↵airs, 107(428):361–386, 2008.

Pan Butamire. How colonialism begat self-denigrating rwandans. New Times,2012. URL https://www.newtimes.co.rw/section/read/49838.

Melissa Dell and Benjamin A Olken. The development e↵ects of the extractivecolonial economy: The dutch cultivation system in java. The Review of Eco-nomic Studies, forthcoming, 2019.

K. Desmet, I. Ortuno-Ortın, and R. Wacziarg. The political economy of linguisticcleavages. Journal of Development Economics, 2011.

Alain Destexhe. Rwanda and Genocide in the 20th Century. Pluto Press, 1995.

Benn Eifert, Edward Miguel, and Daniel N Posner. Political competition andethnic identification in africa. American Journal of Political Science, 54(2),2010.

Chaim Fershtman and Uri Gneezy. Discrimination in a segmented society: Anexperimental approach. The Quarterly Journal of Economics, 116(1):351–377,2001. doi: 10.1162/003355301556338. URL http://qje.oxfordjournals.org/content/116/1/351.abstract.

Gunther Fischer, Freddy O Nachtergaele, Sylvia Prieler, Edmar Teixeira, GezaToth, Harrij Van Velthuizen, Luc Verelst, and David Wiberg. Global agro-ecological zones (gaez v3. 0). Technical report, Food and Agriculture Organi-zation, 2012.

Philip Gourevitch. We Wish to Inform You That Tomorrow We Will be KilledWith Our Families: Stories from Rwanda. Picador, 1998.

Jonas Hjort. Ethnic divisions and production in firms. Quarterly Journal ofEconomics, 129(4):1899–1946, 2014.

Villia Jefremovas. Brickyards to Graveyards: From Production to Genocide inRwanda. SUNY Press, 2002.

Saumitra Jha. Trade, institutions, and ethnic tolerance: Evidence from south asia.American Political Science Review, 107(04):806–832, 2013.

Cynthia Kinnan. Distinguishing barriers to insurance in thai villages. 2014.

Stephen Knack and Philip Keefer. Does social capital have an economic payo↵?a cross-country investigation. Quarterly Journal of Economics, 112(4), 1997.

Rene Lemarchand. Political instability in africa: The case of rwanda and bu-rundi/l’instabilite politique en afrique: Le cas du rwanda et du burundi. Civil-isations, pages 307–337, 1966.

20

Sara Lowes and Eduardo Montero. Concessions, violence, and indirect rule: Evi-dence from the congo free state. Unpublished manuscript, 2018.

Sara Lowes, Nathan Nunn, James A. Robinson, and JonathanWeigel. Understand-ing ethnic identity in africa: Evidence from the implicit association test (iat).American Economic Review Papers and Proceedings, 105(5):340–345, 2015.

Ministre des Colonies M. le Premier Ministre. Rapport sur L’administration Belgedu Rwanda-Urundi Pendant L’ann‘’ee 1927(-1945). Societe anonyme d’editionsjuridiques et cientifiques, 1927(-1945).

Mahmood Mamdani. When victims become killers: Colonialism, nativism, andthe genocide in Rwanda. Princeton University Press, 2014.

Alfred McCoy. Laos: War and Revolution, chapter French Colonialism in Laos,1893-1945. New York, 1971.

Nikhita Mendis. Rwanda: The quest for reconciliation and jus-tice, April 2014. URL http://brownpoliticalreview.org/2014/04/rwanda-the-quest-for-reconciliation-and-justice/.

Edward Miguel. Tribe or nation? nation building and public goods in kenya versustanzania. World Politics, 56(03):328–362, 2004.

Richard Morrock. Heritage of strife: The e↵ects of colonialist divide and rulestrategy upon the colonized peoples. Science and Society, 37(2):129–151, 1973.ISSN 00368237, 19432801.

Mohd. A. Nawawi. Punitive colonialism: The dutch and the indonesian nationalintegration. Journal of Southeast Asian Studies, 2(2):159–168, 1971. ISSN00224634, 14740680. URL http://www.jstor.org/stable/20069916.

Catherine Newbury. The Cohesion of Oppression: Clientship and Ethnicity inRwanda, 1860-1960. New York: Columbia University Press, 1988.

M Catharine Newbury. Ethnicity in rwanda: the case of kinyaga. Africa, 48(1):17–29, 1978.

N. Nunn. Culture and the historical process. Economic History of DevelopingRegions, 27(S1):108–126, 2012.

Nathan Nunn. The long-term e↵ects of africa’s slave trades. Quarterly Journal ofEconomics, 123(1):139–176, 02 2008.

Nathan Nunn and Leonard Wantchekon. The slave trade and the origins of mis-trust in africa. American Economic Review, 101(7):3221–52, 2012.

Olivier Nyirubugara. Complexities and Dangers of Remembering and Forgettingin Rwanda. Sidestone Press, 2013.

21

Melvin E Page and Penny M Sonnenburg. Colonialism: an international, social,cultural, and political encyclopedia. AM. Vol. 1, volume 1. ABC-CLIO, 2003.

Christopher G Pernin, Brian Nichiporuk, Dale Stahl, Justin Beck, and RickyRadaelli-Sanchez. Unfolding the Future of the Long War: Motivations,Prospects, and Implications for the US Army. Rand Corporation, 2008.

Moses Shayo and Asaf Zussman. Judicial ingroup bias in the shadow of terrorism.Quarterly Journal of Economics, 126(3):1447–1484, 2011.

Scott Straus. The order of genocide: Race, power, and war in Rwanda. CornellUniversity Press, 2013.

Shashi Tharoor. The partition: The british game of ’divide andrule’, 2017. URL https://www.aljazeera.com/indepth/opinion/2017/08/partition-british-game-divide-rule-170808101655163.html.

Nico Voigtlander and Hans-Joachim Voth. Highway to hitler. Working Paper20150, National Bureau of Economic Research, May 2014. URL http://www.nber.org/papers/w20150.

Catherine L Watson. Exile from Rwanda: background to an invasion. US Com-mittee for Refugees, 1991.

22

(a) “Ikiboko” (b) “Obligatory Labour Requirement” (this isenglish translation)

(c) “Corvee” (d) Mean of (a), (b), (c)

Figure 1: E↵ect of forced labour on trust, by definition of forced labourNotes:: Each subfigure plots the descriptive relationship between the returns to allocating land to co↵ee, which isdefined as the quantity produceable times price of the best non-co↵ee crop divided by the same value for co↵ee. Avalue of 1.2 means that other crops return 120% the value of co↵ee, so co↵ee is not likely to be the most dominantcrop in the region in the absence of quotas. A value of 0.6 means that the next best crop to co↵ee returns 60% ofthe value of co↵ee, so co↵ee is expected to be heavily produced in these regions. The y-axis is the share of colonialera french texts on Google Books about a district in Rwanda or Burundi that mention a particular phrase. Ineach subfigure we report a di↵erent phrase. Ikiboko is used in panel a; Prestations Coutumieres Dues in panelb; and corvee in panel c. Panel d reports the mean of the three. In each case we plot the relationship usingkernel weighted local polynomial smoothing. All subfigures use a bi-weight kernel of bandwidth 0.3, and using apolynomial of degree 1. A vertical line is placed at 0.975 on the x-axis to denote the approximate point at whichthe relationship becomes upward sloping in each figure. To ensure outliers are not driving the e↵ect, the datapresented in the figure are winsorized at the 5% level.

23

(a) Hutu Trust (b) Tutsi Trust

(c) Hutu Preference for Inter-ethnic Partner (d) Tutsi Preference for Inter-ethnic Partner

(e) Hutu Agreements (f) Tutsi Agreements

Figure 2: E↵ect of forced labour on ethnic attitudes, by definition of forced labourNotes:: Each subfigure plots regression estimates from a version of equation (3). For each subfigure, the x-axisrepresents the estimate of the e↵ect of having a Grandparent who farmed co↵ee where the quota could be binding,using di↵erent definitions of how much co↵ee needed to return relative to other crops for the quota to bind. Theregressions all use a sample-window approach, where the sample used in the regression is constrained to be withina bandwidth of the threshold. This prevents ‘false-positives’ at the low end of the distribution associated withthe fact that individuals who should be considered treated appear in the falsification treatment group. Thebandwidth used is 0.4 for Hutu, but 0.6 for Tutsi since there is consistently no Tutsi e↵ect, and the Tutsi sampleis much smaller. The panels on the left (panel a, c, e) all show results for the Hutu subsample, while panels on theright (panel b, d, f) all show results for the Tutsi subsample. The top two panels (panel a and b) use the log trustgame o↵er as the dependent variable; the middle two panels (panel c and d) use the partner preference measureas the dependent variable; the bottom two panels (panel e, f) use the number of agreements as the dependentvariable.

24

Table 1: E↵ect of forced labour on trust

(1) (2) (3) (4)

Panel A: Hutu decisions

Dependent Variable: log(Trust Game O↵er)

Sample: Hutu - Tutsi Hutu - Hutu

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -0.227 -0.228 -0.0486 -0.0376(0.106) (0.108) (0.0865) (0.0799)

Grandparent from a village where quotas are thought to be binding ( lgp) 0.188 0.195 0.0791 0.0820(0.170) (0.170) (0.107) (0.0948)

Grandparent farmed co↵ee (Ci,lgp,lr) 0.00461 0.0100 0.0587 0.0401(0.0554) (0.0637) (0.0307) (0.0317)

Mean of dependent variable 6.21 6.21 6.35 6.35Cluster 1: number of grandparent villages 64 64 64 64Cluster 2: number of respondent villages 77 77 56 56

N 258 258 361 361R2 0.375 0.376 0.270 0.294

Panel B: Tutsi decisions

Dependent Variable: log(Trust Game O↵er)

Sample: Tutsi - Hutu Tutsi - Tutsi

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) 0.169 0.171 0.221 0.179(0.165) (0.166) (0.292) (0.283)

Grandparent from a village where quotas are thought to be binding ( lgp) -0.156 -0.153 -0.591 -0.569(0.159) (0.152) (0.264) (0.258)

Grandparent farmed co↵ee (Ci,lgp,lr) -0.0721 -0.0815 0.170 0.215(0.118) (0.124) (0.133) (0.130)


N 128 128 121 121R2 0.793 0.794 0.707 0.711

Respondent district FE Yes Yes Yes YesGrandparent village controls Yes Yes Yes YesEnumerator FE Yes Yes Yes YesIndividual controls Yes Yes Yes YesEducation and Income No Yes No Yes

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in all columns is the log of the trust game o↵er. The log of the o↵er is constructed usingthe hyperbolic sine transformation to deal with zeros. Respondent district fixed e↵ects are included in eachregression. Grandparent village controls included are the suitability for each crop, and an indicator for the cropin the village with the highest return. Distance to the capital and distance to the nearest major city are alsoincluded in columns 2 and 4. Enumerator fixed e↵ects are included in each specification. Individual controlsincluded throughout include gender, age, score on a raven IQ test, and the response to a survey question onrisk preference (hypothetical, not incentivized). In columns 2 and 4 we also include years of education andself-reported income. Samples are split according to own ethnicity and partner ethnicity in the trust game. Allethnicity data is either self-reported (Burundi) or based on self-reported eligibility for a genocide survivors fund(Rwanda), as described in the text.

25

Table 2: E↵ect of forced labour on partner ethnicity preference

(1) (2) (3) (4)

Panel A: Hutu decisions

Dependent Variable: Preference for inter-ethnic partnership

Sample: Full Sample Inter-ethnictrust-game sample


Grandparent from a village where quotas are thought to be binding ( lgp) -0.0416 -0.0365 0.0172 0.0140(0.0393) (0.0392) (0.0879) (0.0884)

Grandparent farmed co↵ee (Ci,lgp,lr) 0.0749 0.0846 0.0907 0.118(0.0408) (0.0402) (0.0697) (0.0678)


N 619 619 258 258R2 0.336 0.339 0.333 0.338

Panel B: Tutsi decisions

Dependent Variable: Preference for inter-ethnic partnership

Sample: Full Sample Inter-ethnictrust-game sample

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -0.0161 -0.00786 0.0853 0.0846(0.0460) (0.0462) (0.0728) (0.0705)

Grandparent from a village where quotas are thought to be binding ( lgp) -0.0473 -0.0419 0.0271 0.0356(0.0630) (0.0689) (0.115) (0.114)

Grandparent farmed co↵ee (Ci,lgp,lr) 0.00669 0.00358 -0.0822 -0.0858(0.0382) (0.0383) (0.0738) (0.0709)


N 249 249 128 128R2 0.413 0.425 0.508 0.510


Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in all columns is the share of partner choices made that are not of the respondent’s ethnicity.Respondent district fixed e↵ects are included in each regression. Grandparent village controls included are thesuitability for each crop, and an indicator for the crop in the village with the highest return. Distance to thecapital and distance to the nearest major city are also included in columns 2 and 4. Enumerator fixed e↵ects areincluded in each specification. Individual controls included throughout include gender, age, score on a raven IQtest, and the response to a survey question on risk preference (hypothetical, not incentivized). In columns 2 and4 we also include years of education and self-reported income. Samples are split according to own ethnicity andpartner ethnicity in the trust game. All ethnicity data is either self-reported (Burundi) or based on self-reportedeligibility for a genocide survivors fund (Rwanda), as described in the text.

26

Table 3: E↵ect of forced labour on insurance agreements

(1) (2) (3) (4)

Sample: Hutu Tutsi Hutu Tutsi

Panel A: Number of Agreements

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -3.884 1.856 -3.983 2.611(1.524) (2.056) (1.544) (2.348)


Grandparent farmed co↵ee (Ci,lgp,lr) 1.859 -1.930 2.022 -2.189(0.717) (1.779) (0.695) (1.876)


N 619 249 619 249R2 0.147 0.344 0.149 0.349

Panel B: Insurance Failure Rate

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) 0.141 -0.0133 0.148 -0.0138(0.0481) (0.0948) (0.0466) (0.0903)

Grandparent from a village where quotas are thought to be binding ( lgp) 0.0214 -0.155 0.0177 -0.165(0.0479) (0.0790) (0.0469) (0.0773)

Grandparent farmed co↵ee (Ci,lgp,lr) -0.0797 -0.0479 -0.0900 -0.0474(0.0363) (0.0801) (0.0357) (0.0744)


N 619 249 619 249R2 0.354 0.429 0.356 0.438

Respondent district FE Yes Yes Yes YesGrandparent village controls Yes Yes Yes YesEnumerator FE Yes Yes Yes YesIndividual controls Yes Yes Yes YesEducation and Income No No Yes Yes

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable is the number of insurance agreements in panel A, and the insurance failure rate, constructedas (1+ defaults)/(1+ agreements), in panel B. Respondent district fixed e↵ects are included in each regression.Grandparent village controls included are the suitability for each crop, and an indicator for the crop in thevillage with the highest return. Distance to the capital and distance to the nearest major city are also includedin columns 2 and 4. Enumerator fixed e↵ects are included in each specification. Individual controls includedthroughout include gender, age, score on a raven IQ test, and the response to a survey question on risk preference(hypothetical, not incentivized). In columns 2 and 4 we also include years of education and self-reported income.Samples are split according to own ethnicity. All ethnicity data is either self-reported (Burundi) or based onself-reported eligibility for a genocide survivors fund (Rwanda), as described in the text.

27

Appendix A. Historical Background Appendix

A.1. Additional Evidence Referenced in the Historical Background Section

Figure A1: Map of co↵ee suitability in Rwanda and BurundiNotes: This map plots raw data on produceable tonnes per acre of co↵ee for Rwanda and Burundi usingdata from the FAO. Overlaid is administrative boundaries for each country, and the locations of respondentgrandparents, in order to provide a sense of the variation being exploited in the paper.

Appendix Page 1

Figure A2: Crop production and prices at the introduction of co↵ee quotasSource: Author calculation based on Belgian colonial publications.Description: This figure shows that the crop quotas were a binding constraint in at least some regions. Afterco↵ee quotas were introduced in 1931 the amount of land devoted to co↵ee increased dramatically relative toother export crops. (exp) denotes export crops, which are represented by solid lines and refer to the y-axis on theleft. (sub) denotes subsistence crops which are represented by dashed lines and refer to the y-axis on the right.Subsistence crop data starts being collected starting in 1933.

Appendix Page 2

Appendix B. Data Appendix

Note that this section borrows very heavily from Blouin and Mukand (2019) andexcludes some descriptions of details that were intended for use only in Blouin andMukand (2019), and that are already described in detail in that paper.

B.1. Additional Evidence Referenced in the Data Section

Table B1: Summary Statistics

Did grandparent farm co↵ee where quotas are thought to be binding? Yes No

Variable Mean Std. Dev. N Mean Std. Dev. N

Panel A: Hutu

Inter-ethnic Trust Game O↵er 263.5 108.5 52 273.8 104.0 206Co-ethnic Trust Game O↵er 286.4 116.6 59 316.2 123.2 302Partner Preference 48% 38% 111 39% 36% 508Insurance Agreements 1.98 4.24 111 3.12 7.51 508Insurance Failure Rate 73% 34% 111 58% 35% 508

Gender: Female 25% 0.43 111 37% 0.48 508Country: Burundi 71% 0.45 111 38% 0.49 508Age 40.0 12.6 111 40.5 13.2 508Education Years 5.35 2.88 111 5.45 3.31 508Cognitive score 5.18 1.73 111 4.91 1.97 508Risk Survey 1.51 0.50 111 1.55 0.49 508Distance to Capital 52.5 31.0 111 53.3 47.5 508

Panel B:Tutsi

Inter-ethnic Trust Game O↵er 313.6 132.0 22 316.9 126.9 106Co-ethnic Trust Game O↵er 286.9 132.5 23 268.4 98.0 98Partner Preference 62% 33% 45 63% 29% 204Insurance Agreements 2.07 2.94 45 4.25 15.11 204Insurance Failure Rate 64% 36% 45 63% 35% 204

Gender: Female 47% 0.50 45 53% 0.50 204Country: Burundi 60% 0.49 45 45% 0.49 204Age 43.6 13.1 45 43.2 12.91 204Education Years 5.64 3.18 45 5.78 3.83 204Cognitive score 5.02 1.83 45 4.88 1.97 204Risk Survey 1.58 0.50 45 1.59 0.49 204Distance to Capital 45.1 26.3 45 46.3 35.1 204

Notes: The table shows descriptive statistics for Hutu and Tutsi farmers in Rwanda and Burundi, who par-ticipated in a data collection session. Among Hutu the sample is balanced on covariates except for gender andcountry. Forced labour Hutu are more likely to be men and more likely to be from Burundi than other Hutu.Tutsi with the same agricultural profile are also much more likely to be from Burundi, but the Tutsi sample isbalanced on gender (and all other covariates as well). The gender di↵erence is driven by co↵ee farming and notregion. The country e↵ect is due to the fact that overall, 80% of respondents in Burundi reported that theirgrandparents grew some co↵ee relative to just 18% in Rwanda.

Appendix Page 3

Table B2: Summary Statistics II: Best Non-Co↵ee Crops by Predicted Returns

Crop Frequency Percentage of sample

Tea 428 49.31%Manioc 240 27.65%Maize 193 22.2%Cocoa 3 0.4%Sweet Potato 4 0.46%

Notes: The table shows the alternative best crop options, matched to each grandparent location. This is thecrop that determines the returns of the non-co↵ee crop that appears in the main source of exogenous variation inforced labour. In particular, the returns to each of these crops appears in the denominator in equation (2) withthe frequency listed in this table.

Appendix Page 4

Table B3: Di↵erences in historical co↵ee production by co↵ee suitabilityFull Sample Hutu-only Tutsi-only

t-test: mean grandparent co↵ee production by relative returns to co↵ee 0.055 0.076 -0.006(0.034) (0.040) (0.065)

Notes: The table shows the di↵erences in historical co↵ee production by co↵ee profitability. We should expect,if respondents knew whether their grandparents produced co↵ee, that historical co↵ee production is correlatedwith the suitability of co↵ee to other crops. We do see that, at least for Hutu respondents, that grandparentproduction is higher in regions where co↵ee is more suitable. However the same is not true for Tutsi, perhapssuggestive of the lower overall co↵ee production among Tutsi.

Appendix Page 5

(a) Respondent locations (b) Ancestral location relative to forced labour

Figure B1: Maps of relevant respondent locationsNotes: Panel (a): Dots represent the location where each respondent lives. Dots are clustered because individualsfrom 4-5 villages were brought to a district capital to be surveyed and play lab games in a data collection session.This was necessary to be able to ensure that respondents would be able to play against strangers in the one-shottrust game.Panel (b): This map shows ancestral village and forced labour regions, defined as regions where co↵ee profitabilityis less than or equal to that of the most profitable non-co↵ee crop. It shows that respondents ancestors arescattered across both Rwanda and Burundi, and populate both forced labour regions (dark) and no forced labourregions (light).

Appendix Page 6

Percentage of Native Population that's Tutsi by Commune (1991)

Legend

ScaleNo Data

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

0.11

0.12

0.13

0.14

0.15

0.16

0.17

0.18

0.19

0.2

0.21

0.22

0.24

0.25

0.27

0.29

0.32

0.33

0.47

Figure B2: Distribution of Rwandan Tutsi population in 1991Notes: This figure maps the Tutsi share by administrative commune, as reported in the 1991 Rwandan census.This was the last time a census collected ethnicity information in Rwanda. The map shows the share of Tutsi ineach pre-genocide commune. After the genocide the administrative boundaries were redrawn.

Appendix Page 7

Figure B3: Histogram of Predicted Relative Co↵ee ReturnsNotes: This figure displays the share of grandparent districts in the data within various bins of predictedrelative co↵ee returns. The histogram shows a non-trivial share of the population drives the hockey-stick patternobserved in the data. Indeed, of the 122 grandparent districts, 62 are above a threshold of 0.9 while 51 are abovea threshold of 0.975. There is one extreme outlier in the sample, so the data presented in the figure are winsorizedat the 1% level. Still there appear to be outliers above a value of 2, so to be very safe that these are not drivingthe main results, we further winsorize to the 5% when displaying the hockey-stick pattern in forced labour.

Appendix Page 8

B.2. Protocol and Experiment Instructions

Enumerators were hired from a local country-specific pool used by the firm thatwe hired to help us manage the data collection. The two data collections e↵orts,one in Rwanda and another in Burundi, were nearly identical. The main di↵erencewas that subjects were directly asked about their ethnicity in Burundi, while thiswas not permitted in Rwanda. In addition, all instructions (written or oral) werein Kinyarwanda in Rwanda and in Kirundi in Burundi. Data was collected forthree projects in mind. The first is Blouin and Mukand (2019) the second is theproject described in this paper and the third is a yet unwritten project.

On a given day there was a morning and an afternoon data collection session.Typically the same villages were used for the morning and afternoon sessions.In any given session we typically have 4-5 people from any given village and 4-5villages present (20 people total), but overall, in the data we have 8-10 subjectsfrom each village (4-5 from the morning and 4-5 from the afternoon).

Data collection sessions took place in a town hall. There was a survey portionand an experimental portion to the session. The surveys took place first, andthe experiments took place second. Surveys were completed sitting down at atable in private with a subject. In the experiment portion of the session there4 were experimental stations and a waiting area. Subjects were in one of thesetwo locations throughout the experiments. In the waiting area there was a largeposter board that listed the partnerships for the trust game. The poster boardwas updated with the o↵ers of the trust game throughout the day if the trustgame was assigned to the public treatment.

In each data collection session, there were well-defined roles for our 8 enu-merators. They were specialists in that, for example, the person who was in theEnumerator 1 role was in that role for every session. It will help to label theseroles as follows, and we’ll refer to them as E1-E4:

• Enumerator 1 (1 person)

1. responsible for greeting subjects as they arrived and handling consent.

2. responsible for matching trust game partners as well as roles (sender /receiver) for the trust game

3. responsible for collecting partner preferences and briefing subjects onhow the experiments would run.

4. responsible for managing the flow of experiments, ensuring subjectsknew where to go, etc.

5. responsible for payment at the end of the day

• Enumerator 2 (3 people)

1. responsible for completing surveys with subjects.

Appendix Page 9

2. responsible for assisting E1 with task 4 above (i.e. logistics / flow ofexperiments).

• Enumerator 3 (3 people)

1. responsible for completing surveys with subjects

2. responsible for completing the trust game with subjects

3. responsible for completing a separate task (for another paper) withsubjects

• Enumerator 4 (1 person)

1. responsible for completing surveys with subjects

2. responsible for completing the SIT with subjects (see Blouin and Mukand(2019) for details)

The timeline of events for data collection was as follows:

• Between 8:00 and 8:30am the enumeration team travelled from the hotel in4 SUVs to the town hall where the data collection session took place.

• The team unloaded materials and started arranging tables, chairs, posters,and other materials needed for the survey and experiments.

• While set-up was taking place each driver drove an SUV to met subjects neartheir own villages at a pre-determined meeting location, and drove them tothe town hall.

• As subjects arrived (typically 4-5 at a time) they were greeted by E1 whodescribed in general terms the survey, experiments and the purpose of thestudy before distributing and reading the consent agreement.

• After collecting the consent agreements, each subject was given an ID cardthat they pinned to their shirt which listed a letter that corresponded to theregion they live in (a geographic cluster close to the pick-up location), anda number identifying each individual from that region. ID tags were in bagscorresponding to the region (i.e. letter) and were dolled out randomly uponsubject arrival (conditional on letter).

• As they entered the town hall, each subject was paired up with any oneof seven enumerators (i.e. E2-E4) to complete the survey. These enumera-tors simply lined up near the entrance of the town hall and paired-up withsubjects as they entered.

• As subjects were completing the surveys with one of the other enumerators,E1 matched each subject to another for the trust game and determinedroles (who was to play as sender / receiver). This matching was done usingnumbers from the ID tags, and was done without any consideration (orknowledge of) the ethnic identity of the subjects. At this stage the only

Appendix Page 10

consideration when matching two subjects was logistical - to ensure thatsubjects who were matched to each other were not from the same village.Once E1 knew the exact composition of the session (i.e. how many subjectsfrom each village) she ensured that, for instance, whoever had ID tag A2would be partnered with whoever got B5. An ‘A’ (i.e. from region A) wasnever partnered with another ‘A’, and likewise for every other letter. So asa first step, we assumed that if two people were not from the same village itwas unlikely they would know each other. This matching process typicallytook no more than 15 minutes.

– The only reason this was not completely trivial (and hence why it typi-cally took more than 30 seconds) is each subject could only be partneredwith another individual once, to prevent outcomes from one game in-fluencing another. For example, subjects played the trust game twice,once as a sender and once as a receiver. E1 always made sure that thatany partnership only ever took place once.

• As subjects finished the surveys, they were sent back to E1. She briefedeach subject (sequentially and one at a time) on logistics and also elicitedinformation in the following order:

– Partner-selection task: at this point the partner-preferences were takenfor each subject. Each subject was asked to look around at the IDtags of the people at the session, and list the top 5 individuals thatthey would like to be partnered with to take part in a cooperative task.They were told that a few people would be matched with a partner oftheir choice for the last game of the day (which was the task under E3item 3 in the list of responsibilities above)

– Payment Protocol. Each subject was informed that the monopolymoney he/she received, represented real money. Every dollar of monopolymoney represented a Rwandan or Burundian Franc. At the end of eachexperiment an enumerator would write down on a piece of paper howmuch money was earned in the experiment, signed the back of the pieceof paper, and put the piece of paper in a sac that the subjects were re-sponsible for. At the end of all experiments, subjects would reach intothe sac and pull out one piece of paper, and would be paid in cash,the amount listed. It was stressed that total money earned would de-pend on outcomes of the various lab exercises, but that they would earnmoney from only one specific exercise, chosen at random.

– Question on previous acquaintance: As noted above, subjects were as-signed trust-game partners randomly. Since the partner pairings weredone while surveys were being completed, at this point subjects wereinformed of who their partner would be. The enumerator asked sub-jects (individually) if they had ever met their partner before. If eitherhad, the protocol dictated that new partners be found.

Appendix Page 11

• Subjects were then taken to an area of the town hall that had seats for themto wait until they were called to participate in an experiment. They oftenhad to briefly wait for an assigned partner to complete the survey.

• As subjects became available, E1 called out the ID tags of subjects andbrought each subject them over to E3 or E4, at one of the experiment stationsdepending on who was free.

• When a subject arrived at E4’s work station, he implemented the SIT forthe subject (see Appendix F).

• In the case of the trust-game, as subjects arrived and sat down at the trust-game station, E1 (or E2) told E3 which subject was player 1 and which wasplayer 2.29 This is to to ensure that each player got to play once as ‘sender’and the other time as ‘receiver’.

– E3 explained the experiment to both subjects and gave out the Monopolymoney to player 1. There was a script that they read, and then they askedboth subjects if they understood both the rules and the implications of thedecisions. The enumerator was free to explain the trust game in their ownwords, if they felt that subjects did not understand the game after beingread the script.

• After playing the trust game, the subject was sent back to the waiting areauntil they could participate in another experimental exercise.

• Part of the job of E1 and the three E2s was to keep an eye out for idleexperimental stations and subjects that had not yet participated in thosestations. E.g. if E4 was idle and a participant in the waiting area had notcompleted the SIT, E1 or E2 would be responsible for making sure that E4completed the SIT with the subject at that time.

• Once a subject finished all the experiments, they were called from the wait-ing area, and they pulled from their sac a piece of paper that listed theirpayment. They were given this payment plus a participation fee plus a caseof soap. They returned the ID tag and were free to leave.

• As they exited the town hall, the drivers were waiting for them to take themback to their village. When drivers dropped o↵ one set of subjects anotherset was waiting at the same location for the afternoon session.

• While drivers dropped o↵ and picked up subjects, enumerators typicallywent into town to have lunch. In the evening session when drivers droppedo↵ subjects, the enumerators organized all of the surveys and experimentalmaterials. Drivers returned, the SUV was packed up, and the team typicallywent for dinner near the hotel.

29Recollect that after the survey was completed, all the E2’s had only one responsibility -namely to help E1 with logistics.

Appendix Page 12

B.3. Original Experiment Instructions:

T1) Player 1 ID:

T2) Player 2 ID:

For all respondents, read the following:

Here’s how the exercise works: Player 1 receives 600RWF. You can put somemoney onto the table, and whatever is placed on the table will be doubled. ThenPlayer 2 gets to decide how divide all of the money on the table between the twoof you.

Enumerator instructions:

• Make sure they understand the game.

• Explain in your own words if necessary.

• Give the 6 bills of Monopoly money to player 1.

T4) Player 1: How much of your 600RWF would you like to share? Circle one

a) 0 RWFb) 100 RWFc) 200RWFd) 300RWFe) 400RWFf) 500RWFg) 600RWF

Instruction to Enumerator: Now double what was given by player 1 and givethe money to player 2.

Player 2: I have taken the money shared by player 1 and have doubled it. Youcan now decide how to divide this money between the two of you. How muchwould you like to keep and how much would you like player 1 to have, in additionto what (s)he has already kept?

T5) Player 2 keeps:

T6) Player 2 shares with player 1:

Enumerator Instructions: Now write down on a ‘payment stub’ for each person,how much that person earned in the experiment. Sign or initial the back of thestub and place it in the sac that they have to hold the stubs.

Appendix Page 13

B.4. Ethnicity Proxy Rwanda

In Rwanda it is not feasible to directly ask individuals their ethnicity. We discov-ered that information on ethnicity is typically easy to collect (and often alreadyis) using income surveys that ask respondents their sources of income.

The Rwandan government is strict about the criterion for FARG eligibility,and the term ‘survivors of the genocide’ has become a political issue. It currentlycoincides with Tutsi ethnicity and completely excludes any Hutu a↵ected by thegenocide. International human rights agencies define someone as a victim of thegenocide (and family members are genocide ‘survivors’) if they died in Rwanda dueto violence between April 7th and July 15th 1994, regardless of ethnicity. However,this is not so in Rwanda. The genocide has been relabelled as the “genocideagainst the Tutsi”, and Hutu who died during the genocide are not considered asits ‘victims.’ Consequently, only the Tutsi can be considered ‘genocide survivors’.In our sample, everyone in Rwanda was aware of FARG. In Burundi we did not ask.Of course, there was no constraint on asking ethnicity in Burundi. In contrast toRwanda, in Burundi as part of a background survey, we directly asked all subjectsabout their ethnic background.

Appendix Page 14

B.5. Colonial Price Data

Colonial era production and price data was collected for Rwanda and Burundi.The quantity data is used descriptively while the price data is used to define⇧lpg. The source for both is Belgian colonial yearbooks (M. le Premier Ministre,1927(-1945)).

Summaries varied in what they included by year, but for main crops a yearlyquantity and value produced was available.30 An example of one of the pages fromthe colonial yearbooks can be seen in figure B4.

These tables were transcribed for all available crops to build a crop-year panelcontaining 368 observations between 1925 and 1950. Because we are concernedabout the e↵ect of the supply shock on local prices, we only consider the meanprices prior to 1931.

This data was matched to the FAO data to get a sense of which crops wouldhave been most lucrative. In some cases the relevant crop data do not exist inthe colonial yearbooks. In these cases we supplement using export commodityprice data from Bazzi and Blattman (2014), and use the closest overlapping yearto normalize the units. Specifically, I take the closest year where data exists forboth the colonial yearbook and the Bazzi and Blattman (2014) data and dividethe yearbook data by the Bazzi and Blattman (2014) data. I then multiply thisratio by the Bazzi and Blattman (2014) data for the year that is missing in thecolonial yearbook.

30The available crops: maize, rice, sorghum, manioc, potato, sweet potato, banana, peas,vegetables, sugarcane, peanut, sesame, cotton, jute, co↵ee, cocoa. All but jute are availablefrom FAO.

Appendix Page 15

Figure B4: Example of Prices and Quantities in Colonial Yearbooks

Appendix Page 16

Appendix C. Strategy and Results Appendix

(a) “Ikiboko” (b) “Obligatory Labour Requirement” (this isenglish translation)

(c) “Corvee” (d) Mean of (a), (b), (c)

Figure C1: E↵ect of forced labour on trust, by definition of forced labourNotes:: Each subfigure plots the binscatter between the returns to allocating land to co↵ee, which is defined asthe quantity produceable times price of the best non-co↵ee crop divided by the same value for co↵ee. A value of1.2 means that other crops return 120% the value of co↵ee, so co↵ee is not likely to be the most dominant cropin the region in the absence of quotas. A value of 0.6 means that the next best crop to co↵ee returns 60% of thevalue of co↵ee, so co↵ee is expected to be heavily produced in these regions. The y-axis is the share of colonialera french texts on Google Books about a district in Rwanda or Burundi that mention a particular phrase. Ineach subfigure we report a di↵erent phrase. Ikiboko is used in panel a; Prestations Coutumieres Dues in panel b;and corvee in panel c. Panel d reports the mean of the three. In each case we plot the relationship using kernelweighted local polynomial smoothing. All subfigures plot 0.02 bins for all bins less than 1 as any bin reportingvalues of 1 contain one district without many books, and distorts the y-axis scale. A vertical line is placed at0.975 on the x-axis to denote the approximate point at which the relationship becomes upward sloping in eachfigure. To ensure outliers are not driving the e↵ect, the data presented in the figure are winsorized at the 5%level.

Appendix Page 17

(a) gender (b) age

(c) raven score (d) risk

Figure C2: Di↵erences in individual covariates around the thresholdNote: The figure shows the di↵erences in individual covatiates around the threshold. All covariates are includedas controls in each regression, and while there are some di↵erences near the threshold in gender and age, noneare significant.

Appendix Page 18

Table C1: Determinants of forced labour

(1) (2) (3) (4)

Dependent Variable: [Count of “District Name + Search Term”] / Count of “District Name”

Search Term “Ikiboko” “Obligatory Labour “Corvee” Mean ofRequirement” (1)-(3)

(french translated)

Costliness of devoting land to co↵ee 0.05 0.20 0.26 0.17(0.026) (0.067) (0.097) (0.059)

Age control Yes Yes Yes YesDistance to capital Yes Yes Yes Yes

N 78 78 78 78R2 0.03 0.06 0.05 0.03

Notes: Standard errors are clustered by bins sized 0.25 units of the costliness of devoting land to co↵ee. Theunit of observation is a colonial era district.

Appendix Page 19

Table C2: E↵ect of Forced Labour on Migration

(1) (2) (3) (4)

Dependent Variable: Current Village Di↵erentfrom Grandparent Village

Sample: Hutu Tutsi Hutu Tutsi

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) 0.00697 0.517 -0.000297 0.536(0.0370) (0.0951) (0.0348) (0.0949)


Grandparent farmed co↵ee (Ci,lgp,lr) -0.00725 -0.183 0.00456 -0.190(0.0355) (0.0593) (0.0306) (0.0664)

Respondent district FE Yes Yes Yes YesGrandparent village controls Yes Yes Yes YesEnumerator FE Yes Yes Yes YesIndividual controls Yes Yes Yes YesEducation and Income No No Yes Yes


N 619 249 619 249R2 0.540 0.707 0.542 0.713

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in each column is an indicator for whether the respondent currently lives in the same districttheir father was born in. Grandparent village controls included are the suitability for each crop, and an indicatorfor the crop in the village with the highest return. Respondent district fixed e↵ects are included in each regression.Distance to the capital and distance to the nearest major city are also included. Enumerator fixed e↵ects areincluded in each specification. Individual controls included throughout include gender, age, score on a raven IQtest, and the response to a survey question on risk preference (hypothetical, not incentivized). All ethnicity datais either self-reported (Burundi) or based on self-reported eligibility for a genocide survivors fund (Rwanda), asdescribed in the text.

Appendix Page 20

Tab

leC3:

Poo

ledTutsiregression

tocheckifunder

pow

ered

Log

trust

gameo↵

erInter-ethnic

partner

preference

Log

trust

gameo↵

erInter-ethnic

partner

preference

(1)

(2)

(3)

(4)

Grandparentfarm

edco↵ee

wherequ

otas

arethou

ghtto

bebinding(C

o-ethnic

partner)

0.221

0.0599

0.179

0.0641

(0.292)

(0.0951)

(0.283)

(0.0860)

Grandparentfrom

avillagewherequ

otas

arethou

ghtto

bebinding(C

o-ethnic

partner)

-0.591**

-0.136

-0.569**

-0.141

(0.264)

(0.115)

(0.258)

(0.116)

Grandparentfarm

edco↵ee

(Co-ethnic

partner)

0.170

-0.0292

0.215*

-0.0337

(0.133)

(0.0620)

(0.130)

(0.0554)

Grandparentfarm

edco↵ee

wherequ

otas

arethou

ghtto

bebinding(Inter-ethnic

partner)

-0.0528

0.0254

-0.00824

0.0205

(0.290)

(0.109)

(0.269)

(0.100)

Grandparentfrom

avillagewherequ

otas

arethou

ghtto

bebinding(Inter-ethnic

partner)

0.435

0.164

0.417

0.176

(0.285)

(0.117)

(0.278)

(0.120)

Grandparentfarm

edco↵ee

(Inter-ethnic

partner)

-0.242

-0.0530

-0.297

-0.0521

(0.203)

(0.0981)

(0.193)

(0.0948)

Respon

dentdistrictFE

Yes

Yes

Yes

Yes

Grandparentvillagecontrols

Yes

Yes

Yes

Yes

Enu

merator

FE

Yes

Yes

Yes

Yes

Individual

controls

Yes

Yes

Yes

Yes

Education

andIncome

No

No

Yes

Yes

Meanof

dep

endentvariab

le6.25

0.63

6.25

0.63

Cluster

1:nu

mber

ofgran

dparentvillages

6161

6161

Cluster

2:nu

mber

ofrespon

dentvillages

5757

5757

N249

249

249

249

R2

0.033

0.012

0.041

0.016

Notes:

Standard

errors

are

two-w

ayclustered

atth

eresp

onden

tsectorandgrandparentdistrictleve

ls.Thedep

enden

tva

riable

inea

chco

lumnis

anindicatorforwhether

the

resp

onden

tcu

rren

tlylives

inth

esamedistrictth

eirfath

erwasborn

in.Grandparentvillageco

ntrolsincluded

are

thesu

itabilityforea

chcrop,andanindicatorforth

ecropin

the

village

withth

ehighestretu

rn.Respon

den

tdistrictfixed

e↵ects

areincluded

inea

chregression

.Distance

toth

eca

pital

anddistance

toth

enea

rest

majorcity

arealso

included

.Enumeratorfixed

e↵ects

are

included

inea

chsp

ecifica

tion.Individualco

ntrols

included

throughoutincludegen

der,age,

score

on

arave

nIQ

test,and

theresp

onse

toasu

rvey

question

onrisk

preference

(hypothetical,not

incentivized

).Allethnicitydatais

eith

erself-rep

orted(B

uru

ndi)

orbased

onself-rep

ortedeligibilityforage

nocidesu

rvivorsfund

(Rwanda),

asdescribed

inth

etext.

Appendix Page 21

C.1. The Roles of Trustworthiness, Altruism and Reciprocity

The Trust game e↵ects appear to be driven specifically by trust and not altruism(trust game o↵ers capture trust and altruism; they also may capture risk, butrisk is included as a control throughout) since table C4 show the return o↵ersinvolving forced labour Hutu are consistently positive but imprecisely estimated(return o↵ers capture altruism and reciprocity, but neither seems to have beeninfluenced by forced labour). These results also demonstrate that the resultsare not driven by di↵erential trustworthiness of the partners randomly assignedto individuals whose grandparent farmed co↵ee where quotas are thought to bebinding.

Table C4: Analysis of Forced Labour and Return O↵ers

(1) (2) (3) (4)

Dependent Variable: log(Trust Game Return)

Sample : Tutsi return to Hutu Hutu return to Tutsi

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) 0.174 0.141 0.0506 0.0437(0.268) (0.271) [0.228] [0.233]

Grandparent from a village where quotas are thought to be binding ( lgp) -0.497 -0.564 0.166 0.159(0.581) (0.549) [0.416] [0.418]

Grandparent farmed co↵ee (Ci,lgp,lr) 0.0624 0.0992 -0.173 -0.164(0.214) (0.243) [0.176] [0.191]


Mean of dependent variable 3.18 3.18 4.79 4.79Cluster 1: number of grandparent villages 45 45 . .Cluster 2: number of respondent villages 61 61 25 25

N 170 170 111 111R2 0.900 0.901 0.971 0.971

Notes: Standard errors are in brackets. In round brackets we report two-way clustered at the respondent sectorand grandparent district levels. For the Tutsi sample we do not have enough variation to do this (note the verysmall MSE), so in square brackets we report standard errors only at the sector level. The dependent variable ineach column is inter-ethnic trust game 2nd stage (return) o↵ers. Respondent district fixed e↵ects are includedin each regression. Grandparent village controls included are the suitability for each crop, and an indicatorfor the crop in the village with the highest return. Distance to the capital and distance to the nearest majorcity are also included. Enumerator fixed e↵ects are included in each specification. Individual controls includedthroughout include gender, age, score on a raven IQ test, and the response to a survey question on risk preference(hypothetical, not incentivized). All ethnicity data is either self-reported (Burundi) or based on self-reportedeligibility for a genocide survivors fund (Rwanda), as described in the text.

Appendix Page 22

C.2. Heterogeneity by Country

Given the Rwanda-specific challenges, there may be some interest in exploringheterogeneity by country. In particular, issues relating to measuring ethnicityusing the ethnicity proxy, and thinking about how the genocide may influenceestimates, make interpreting the Rwanda estimates more di�cult. While splittingthe sample by both country and ethnicity begins to push the boundaries of samplesize limitations, it is still possible, for the most part, to separate the main e↵ectsby country, as a robustness check.

Table C5: Main results by country

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

Dependent Variable: log(H-T trust) Group Preference Agreements

Burundi Rwanda Burundi Rwanda Burundi Rwanda

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -0.183 -0.277 -0.210 -0.116 -5.872 -2.017(0.126) (0.158) (0.0829) (0.0517) (3.233) (0.983)

Grandparent from a village where quotas are thought to be binding ( lgp) 0.312 0.0917 0.153 -0.0253 0.789 -0.278(0.280) (0.304) (0.215) (0.0587) (3.182) (1.139)

Grandparent farmed co↵ee (Ci,lgp,lr) -0.0183 0.125 0.0723 0.0683 1.198 1.379(0.0466) (0.0874) (0.0604) (0.0427) (1.028) (0.904)

Respondent district FE Yes Yes Yes Yes Yes YesGrandparent village controls Yes Yes Yes Yes Yes YesEnumerator FE Yes Yes Yes Yes Yes YesIndividual controls Yes Yes Yes Yes Yes YesEducation and Income No No No No No No

Mean of dependent variable 6.12 6.39 0.59 0.25 3.00 2.85Cluster 1: number of grandparent villages 49 15 70 28 70 28Cluster 2: number of respondent villages 65 12 77 12 77 12

N 169 89 273 346 273 346R2 0.395 0.384 0.242 0.129 0.176 0.198

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in columns 1-2 is inter-ethnic trust game o↵ers. The dependent variable in columns 3-4 is theshare of people selected as a possible partner that were not of the respondents ethnicity. The dependent variablein columns 5-6 is the number of inter-household insurance agreements made by the respondent. Respondentdistrict fixed e↵ects are included in each regression. Grandparent village controls included are the suitability foreach crop, and an indicator for the crop in the village with the highest return. Distance to the capital and distanceto the nearest major city are also included. Enumerator fixed e↵ects are included in each specification. Individualcontrols included throughout include gender, age, score on a raven IQ test, and the response to a survey questionon risk preference (hypothetical, not incentivized). All ethnicity data is either self-reported (Burundi) or basedon self-reported eligibility for a genocide survivors fund (Rwanda), as described in the text.

Table C5 examines each of the Hutu-Tutsi trust game, the group preferencemeasure and the insurance agreements by country. It shows that for the two labmeasures, estimates are very similar (albeit less precise), while for the resourcesharing agreements, the results are much larger in Burundi but negative and pre-cise for both countries separately.

Overall, it is di�cult to conclude that any Rwanda-specific concern is respon-sible for driving estimates. In two of the three main outcomes (partner preferenceand agreements) the Burundi point estimate is larger, and in the third (the trustgame), the point estimate is similar and there is no statistical di↵erence betweenthem. In the Rwanda-only sample the insignificant but large negative estimateturns significant with the addition of a genocide control (see table C7). This makes

Appendix Page 23

sense as the genocide does appear to be a function of forced labour (figure C3) andthe pro-social e↵ects of civil conflict are very well known Bauer et al. (2016). Thispro-social e↵ect of civil-conflict can be replicated for our sample as well (again,see table C7)

Appendix Page 24

C.3. Genocide

I consider the genocide in this section. It is a tricky empirical issue to deal withbecause clearly an event in 1994 based on extreme ethnic conflict is plausiblyendogenous to the colonial era treatment that may have sparked ethnic conflict.The genocide might therefore be better viewed as a mechanism rather than aconfound to the estimates. In this light, if the genocide reduced the magnitudeand significance of the estimates31 it could just be because there is no e↵ect oftreatment once we control for the e↵ect of the treatment.

That said, the genocide does not make as much di↵erence on the results asone might expect, and that appears to be partly because it is not negativelycorrelated with current attitudes. It could also partly be because of the samplingprocedure in Rwanda. I sampled exclusively in the south where we knew therewould be Tutsi, and where we knew genocide fund FARG was active. The latteris important to avoid having mis-measurement of the ethnicity variable comingfrom people who were Tutsi, but who claimed not to be eligible for FARG becausethey had no presence in the region. As a result, I only capture intensive marginvariation in the genocide. Nevertheless, having noted all of these caveats, it iscertainly worthwhile examining the role of the genocide.

First, consider again table C5 which presents results separately for Rwandaand Burundi. While there has been considerable ethnic violence in Burundi therewas not a single event like the genocide, yet the results are very similar betweenRwanda and Burundi. Second, we can simply add a genocide control to see whathappens, but would strongly caveat this exercise by reiterating that the genocideis clearly endogenous. Anyway, to start, I show that adding a genocide controldoes not get rid of the e↵ect when we consider Rwanda and Burundi together(table C6).

Interestingly the genocide seems, if anything, to be associated with higherinter-ethnic trust. This is consistent with the literature which consistently findspro-social consequences of conflict (Bauer et al., 2016). Again, the estimate isnot precisely estimated. This could be because a value of zero has been imputedfor all Burundi observations where there was no genocide, or because limiting thevariation to only the intensive margin makes the exercise underpowered.

Accordingly, seeing what happens only for the Rwanda observations may beof interest. Table C7 therefore examines the same table but for Rwanda-only.

The results on genocide are somewhat mixed - the trust e↵ect is a larger pos-itive and marginally significant for trust (column 1); the partner preference pointestimate is also positive but small and imprecise (column 2); and the agreementestimate is positive and significant (column 3). The e↵ect of forced labour remainsroust in each case. In fact, the Rwanda-only forced labour result is more negativeand more precise with the genocide control (table C7 column 1) than without(table C5 column 4). The other estimates are similarly robust.

This could be because genocide is unrelated to forced labour, or because thegenocide did not di↵erentially impact co↵ee farmers. A t-test suggests that inregions where other crops returned at least 97.5% of colonial era co↵ee returns,

31This is hypothetical - in practice it does not reduce the magnitude or precision

Appendix Page 25

Table C6: Considering the Genocide I

Dependent Variable log(trust game o↵er) Partner Preference Agreements(1) (2) (3)

Genocide 0.0792 0.00357 0.514(0.109) (0.0277) (0.405)

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -0.226 -0.136 -3.837(0.107) (0.0469) (1.531)

Grandparent from a village where quotas are thought to be binding ( lgp) 0.169 -0.0418 -0.502(0.191) (0.0393) (0.671)

Grandparent farmed co↵ee (Ci,lgp,lr) 0.00942 0.0749 1.855(0.0539) (0.0409) (0.713)

Respondent district FE Yes Yes YesGrandparent village controls Yes Yes YesEnumerator FE Yes Yes YesIndividual controls Yes Yes YesEducation and Income No No No

Mean of dependent variable 6.21 0.40 2.92Cluster 1: number of grandparent villages 64 97 97Cluster 2: number of respondent villages 77 89 89

N 258 619 619R2 0.377 0.336 0.148

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in columns 1 is inter-ethnic trust game o↵ers. The dependent variable in columns 2 is theshare of people selected as a possible partner that were not of the respondents ethnicity. The dependent variablein columns 3 is the number of inter-household insurance agreements made by the respondent. Respondent districtfixed e↵ects are included in each regression. Grandparent village controls included are the suitability for eachcrop, and an indicator for the crop in the village with the highest return. Distance to the capital and distance tothe nearest major city are also included. Enumerator fixed e↵ects are included in each specification. Individualcontrols included throughout include gender, age, score on a raven IQ test, and the response to a survey questionon risk preference (hypothetical, not incentivized). All ethnicity data is either self-reported (Burundi) or basedon self-reported eligibility for a genocide survivors fund (Rwanda), as described in the text.

genocide persecutions in 1994 was higher by about 110 people per sector (p-value<0.01), so the genocide and forced labour do at least appear related on the surface.Whether this relationship exhibits the same non-linearity as the forced labourdata depends on local polynomial kernels and bandwidth choices. Given thismuch less conclusive result, the prudent thing may be to show the scatterplotwithout imposing too much interpretation. That is presented in figure C3.

In any case, the broader point is that the genocide is endogenous to colonialforced labour in theory, and plausibly endogenous to forced labour in the data. Re-gardless, overall the results are surprisingly robust to controlling for the genocidewhich seems to stem from the fact that the genocide and attitudes are positivelycorrelated. The pro-social e↵ect of conflict is an empirical finding that has supportfrom the broader conflict literature (Bauer et al., 2016).

Appendix Page 26

Table C7: Considering the Genocide II (Rwanda-only)

Dependent Variable log(trust game o↵er) Partner Preference Agreements(1) (2) (3)

Genocide 0.135 0.00507 0.995(0.0797) (0.0190) (0.278)

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -0.285 -0.115 -1.883(0.159) (0.0524) (0.998)

Grandparent from a village where quotas are thought to be binding ( lgp) 0.170 -0.0249 -0.202(0.265) (0.0578) (1.071)

Grandparent farmed co↵ee (Ci,lgp,lr) 0.153 0.0682 1.361(0.0799) (0.0429) (0.911)

Mean of dependent variable 6.39 0.26 2.85Cluster 1: number of grandparent villages 15 28 28Cluster 2: number of respondent villages 12 12 12

N 89 346 346R2 0.393 0.129 0.207

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in columns 1 is inter-ethnic trust game o↵ers. The dependent variable in columns 2 is theshare of people selected as a possible partner that were not of the respondents ethnicity. The dependent variablein columns 3 is the number of inter-household insurance agreements made by the respondent. Respondent districtfixed e↵ects are included in each regression. Grandparent village controls included are the suitability for eachcrop, and an indicator for the crop in the village with the highest return. Distance to the capital and distance tothe nearest major city are also included. Enumerator fixed e↵ects are included in each specification. Individualcontrols included throughout include gender, age, score on a raven IQ test, and the response to a survey questionon risk preference (hypothetical, not incentivized). All ethnicity data is either self-reported (Burundi) or basedon self-reported eligibility for a genocide survivors fund (Rwanda), as described in the text.

Appendix Page 27

Figure C3: Scatterplot of genocide and costliness of co↵ee quotas based on landcharacteristics

Appendix Page 28

C.4. Trust Survey

I have a trust question which is di�cult to interpret, but may be related to inter-ethnic trust. I suggest considerable caution in interpreting this measure. Weasked respondents about their trust of people from the “other community” intheir village. In other contexts (Blouin and Mukand, 2019) this measure is mostlyconsistent with other measures of ethnic attitudes, but typically provides muchnoisier estimates, and does not always capture as much nuance as other measures.32

I stress that it is not known how this was interpreted by respondents, though itmay have been that some respondents interpreted the question as a veiled questionabout inter-ethnic trust. Regardless, I can attempt to replicate the result in table1 using this survey variable instead of the trust game o↵er (table C8).

In columns 1 and 2 I show the estimates from the Hutu-Tutsi subsample (analo-gous to table 1 panel A columns 1 and 2), while in columns 3 and 4 show estimatesfrom the Tutsi-Hutu subsample (analogous to table 1 panel B columns 1 and 2).Results are quite similar to the trust game estimates using the trust survey ques-tion. Hutu respondents with grandparents that were more likely to be exposed toforced labour responded more negatively to a question about out-community trust(column 1 and 2). The result is not significant for the Tutsi, but am unable tosay that the magnitudes are statistically from di↵erent from each other (as is thecase with the trust game estimates) since the Tutsi sample produces a moderatenegative estimate as well, albeit at or less than a quarter of the magnitude of theHutu result. In any case the result is mostly consistent with the trust game, butultimately inconclusive on its own.

32That paper examines radio signal. Using the trust game, partner preference and SIT wesee expected inverse-U patterns in trust, as expected. Using the survey measure we see resultsin the areas we expect, but the expected inverse-U pattern in estimates is not noticeable, withunexpectedly large results in many control regions.

Appendix Page 29

Table C8: E↵ect of Forced Labour on Trust Survey Response

(1) (2) (3) (4)

Dependent Variable: Trust (survey)

Sample Hutu Tutsi


Grandparent from a village where quotas are thought to be binding ( lgp) 0.375 0.356 -0.433 -0.488(0.204) (0.209) (0.346) (0.359)


Respondent district FE Yes Yes Yes YesGrandparent village controls Yes Yes Yes YesEnumerator FE Yes Yes Yes YesIndividual controls Yes Yes Yes YesEducation and Income No No No No


N 257 257 128 128R2 0.409 0.412 0.510 0.541

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable selections to whether people trusted others in their village from another community. Answerscould be one of: ‘Not at all’; ‘Just a little’; ‘Somewhat’; ‘A lot’. There is one missing observation in the Hutusubsample. Respondent district fixed e↵ects are included in each regression. Grandparent village controls includedare the suitability for each crop, and an indicator for the crop in the village with the highest return. Distanceto the capital and distance to the nearest major city are also included. Enumerator fixed e↵ects are included ineach specification. Individual controls included throughout include gender, age, score on a raven IQ test, andthe response to a survey question on risk preference (hypothetical, not incentivized). All ethnicity data is eitherself-reported (Burundi) or based on self-reported eligibility for a genocide survivors fund (Rwanda), as describedin the text.

Appendix Page 30

C.5. Partner Preference Robustness

The partner preference measure reveals that forced labour Hutu have a di↵eren-tial preference for partnering with other Hutu, relative to Hutu who live in thesame village but do not have a family history of forced labour. However, thereare observable di↵erences between Hutu and Tutsi, and it could be any of theseobservable di↵erences that are generating the e↵ect.

For instance, Tutsi have historically had higher income, and been more edu-cated. Although neither is true in our data (see relevant means and a t-test ofthe di↵erence in table C9, this may be a stereotype of Tutsi, and individuals mayhave been selecting on the basis of perceived income or education, rather thanon Tutsi status directly. In our data gender and age are di↵erent by ethnicity,with Tutsi being slightly older, and much more likely to be a woman. It certainlyseems possible that we might capture an increased partner preference for womenor for older individuals that looks like a Tutsi-preference. There is no observabledi↵erence for risk or raven score, and no obvious ethnic-stereotype or way to tellwhether someone might be smart or high risk, independent of income or education.

Table C9: Hutu-Tutsi di↵erences by Observable Characteristics

(1) (2) (3)

Hutu Mean Tutsi Mean p-value of di↵erence

Income (RWF / yr) 255,465 229,336 0.53(22,971) (32,334)

Education 5.4 5.8 0.19(0.13) (0.24)

Gender: Female=1 0.348 0.522 0.000(0.019) (0.032)

Age 40.4 43.3 0.004(0.53) (0.82)

Raven Score 4.96 4.91 0.71(0.08) (0.13)

Risk 0.55 0.59 0.25(0.02) (0.03)

Notes: For each observable characteristic we show Hutu and Tutsi means, and standard deviations in bracketsbelow. Column 3 reports a t-test for the di↵erence between each observable characteristic by ethnicity. In eachcase there are 616 Hutu and 249 Tutsi observations.

To account for the possibility that forced labour Hutu were selecting individualson some characteristic other than ethnicity that may be correlated (or perceivedto be correlated) with ethnicity, I directly compute the averages across partnerchoices for each of income, education, gender and age. The results are in C10and they indicate no evidence that forced labour Hutu did not make selections

Appendix Page 31

Table C10: E↵ect of forced labour on preference for partners with various charac-teristics

(1) (2) (3) (4)

Dependent Variable: average of choices for Income Education Gender (female) AgeSample: Hutu

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) -7,296 0.270 0.0731 -0.228(35,730) (0.307) (0.0446) (2.519)

Grandparent from a village where quotas are thought to be binding ( lgp) 55,679 0.825 -0.0277 2.535(36,604) (0.377) (0.0539) (2.050)

Grandparent farmed co↵ee (Ci,lgp,lr) -12,353 0.140 -0.0853 0.845(22,696) (0.191) (0.0283) (2.055)

Respondent district FE Yes Yes Yes YesGrandparent village controls Yes Yes Yes YesEnumerator FE Yes Yes Yes YesIndividual controls Yes Yes Yes YesEducation and Income No No No No

Mean of dependent variable 209,441 5.6 0.4 43Cluster 1: number of grandparent villages 97 97 97 97Cluster 2: number of respondent villages 89 89 89 89

N 619 619 619 619R2 0.299 0.198 0.207 0.233

Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable in each column is the mean of the choices of the respondent on a number of observabledimensions. Grandparent village controls included are the suitability for each crop, and an indicator for the cropin the village with the highest return. Respondent district fixed e↵ects are included in each regression. Distanceto the capital and distance to the nearest major city are also included. Enumerator fixed e↵ects are included ineach specification. Individual controls included throughout include gender, age, score on a raven IQ test, andthe response to a survey question on risk preference (hypothetical, not incentivized). All ethnicity data is eitherself-reported (Burundi) or based on self-reported eligibility for a genocide survivors fund (Rwanda), as describedin the text.

to target any of income, education, gender or age. The closest is age, whichwould rule out zero using an 85% confidence interval, however, the e↵ect goes inthe opposite e↵ect to what we might be concerned about. The share of women islarger among Tutsi than Hutu in the data (again, table C9, and forced labour Hutushowed a distaste for Tutsi but a preference for women. If a gender preferencewas confounding the ethnicity preference that I am interested in, it would be aconcern if forced labour Hutu did not want to partner with women. While itremains possible that forced labour Hutu have a preference for unobservable traitsthat are correlated with ethnicity, I do not find evidence that there is a preferencefor any observable characteristic, other than ethnicity.

Appendix Page 32

C.6. Additional Contracts Results

Evidence in table C11 suggests that examining overall failure rates hides someheterogeneity. While forced labour Hutu experience more insurance failure dueboth to the financial ability of their partner and for strategic reasons, they aremore likely to experience default due to the ability of their partners to make aninsurance transfer / provide a gift.

The original idea behind collecting this information was that the default rea-son might be indicative of partner types. If forced labour Hutu experienced moredefault due to the financial ability of their partner to make a gift / transfer, thenthis might indicate that they are partnering with less suitable partners, i.e. thosewith more positively correlated economic shocks. If on the other hand they ex-perienced more default for strategic reasons, this might indicate low relationshipvalues (Blouin and Macchiavello, 2019), which may be characteristic of partner-ships between people of low trust.

The estimates are certainly not di↵erent enough to say anything conclusive,but the point estimate on the financial ability defaults is about 30% higher theestimate on strategic defaults (table C11 panel A). This may suggest that onemargin of adjustment is for forced labour Hutu to abandon Hutu-Tutsi insurancearrangements in favour of Hutu-Hutu arrangements where partners produce moresimilar agricultural products. We do not see the same pattern for Tutsi (if anythingthe opposite: table C11 panel B).

Appendix Page 33

Table C11: E↵ect of forced labour on insurance failure, by cause

(1) (2) (3) (4)

Dependent Variable: Financial Inability Strategic

Panel A: Hutu

Grandparent farmed co↵ee where quotas are thought to be binding (⇣i,lgp,lr) 0.134 0.134 0.103 0.108(0.0603) (0.0603) (0.0473) (0.0467)

Grandparent from a village where quotas are thought to be binding ( lgp) 0.0506 0.0506 0.0105 0.00964(0.0434) (0.0434) (0.0457) (0.0467)

Grandparent farmed co↵ee (Ci,lgp,lr) -0.102 -0.102 -0.0720 -0.0807(0.0444) (0.0444) (0.0357) (0.0360)


N 619 619 619 619R2 0.277 0.277 0.382 0.384

Panel B: Tutsi





N 249 249 249 249R2 0.378 0.379 0.439 0.439


Notes: Standard errors are two-way clustered at the respondent sector and grandparent district levels. Thedependent variable is the number of insurance agreements in panel A, and the insurance failure rate, constructedas (1+ defaults)/(1+ agreements), in panel B. Respondent district fixed e↵ects are included in each regression.Grandparent village controls included are the suitability for each crop, and an indicator for the crop in thevillage with the highest return. Distance to the capital and distance to the nearest major city are also includedin columns 2 and 4. Enumerator fixed e↵ects are included in each specification. Individual controls includedthroughout include gender, age, score on a raven IQ test, and the response to a survey question on risk preference(hypothetical, not incentivized). In columns 2 and 4 we also include years of education and self-reported income.Samples are split according to own ethnicity. All ethnicity data is either self-reported (Burundi) or based onself-reported eligibility for a genocide survivors fund (Rwanda), as described in the text.

Appendix Page 34

Documents

Culture and Contracts: The Historical Legacy of Forced Labour · lected data on inter-household agriculture insurance agreements. On average, we might expect Hutu and Tutsi to make