segregation, ethnic favoritism, and the strategic...velasquez, 2013) — the two studies that...
TRANSCRIPT
Segregation, Ethnic Favoritism, and the StrategicTargeting of Local Public Goods1
SIMON EJDEMYRStanford University
ERIC KRAMONGeorge Washington University
AMANDA LEA ROBINSONThe Ohio State University
July 8, 2017
1The authors thank Armstrong Chavula and Bright Chimatiro for valuable research assistance, and
Blessings Chinsinga, Daniel Young, and the National Statistics Office of Malawi for sharing data
and expertise. We also thank Francisco Garfias, Guy Grossman, David Laitin, Adrienne LeBas,
Kristin Michelitch, Salma Mousa, Jan Pierskalla, Brigitte Seim, and seminar participants at George
Washington University, The Ohio State University, Stanford University, Texas A&M University,
University of Gothenburg, University of Pittsburgh, Yale University, the 2014 American Political
Science Association conference, and the 2014 African Studies Association conference for com-
ments on earlier versions of this paper.
AbstractThis article demonstrates that ethnic segregation is a key determinant of local publicgoods provision. We argue that this results from politicians’ strategic engagement in eth-nic favoritism: only when ethnic groups are sufficiently segregated can elites efficientlytarget coethnics with local public goods. We test this expectation with fine-grained datafrom Malawi on the spatial distribution of ethnic groups, geolocated distributive goods(water wells), and the ethnic identities of political elites. We find that members of parlia-ment provide more local public goods to their electoral districts when ethnic groups aregeographically segregated, but that this increased investment is primarily targeted towardcoethnics. Thus, while segregation promotes overall public goods provision, it also leadsto greater favoritism in the distribution of these goods. Our logic and evidence providean elite-driven explanation for both the considerable variation in ethnic favoritism acrosscontexts and the under-provision of public goods in ethnically diverse settings.
The expectation that political elites seek to favor their ethnic kin has long been a staple
in the study of African politics (Bates, 1983; Joseph, 1987). A number of empirical studies show
that coethnics of African political leaders have better health and educational outcomes (Franck and
Rainer, 2012), superior infrastructure (Burgess et al., 2015), and preferential access to foreign aid
(Jablonski, 2014; Briggs, 2014). But the existing literature also demonstrates substantial variation
in the prevalence of ethnic favoritism. For example, Franck and Rainer (2012) find strong evidence
of ethnic favoritism in only six of the eighteen African countries that they study, and Kramon and
Posner (2013) demonstrate variation in ethnic favoritism both across African countries and across
types of distributive goods within countries. These findings thus raise an important puzzle that has
yet to be sufficiently addressed: Why is there ethnic favoritism in the distribution of local public
goods in some contexts but not others?
We propose that ethnic group segregation helps account for variation in ethnic favoritism.
In particular, we argue that segregation promotes greater overall investments in local public goods
and leads to more ethnic favoritism in their distribution. This is because targeting coethnics with
local public goods — which are locally non-excludable but costly to access from distant locations
— is difficult unless ethnic groups are sufficiently spatially segregated. Thus, we expect greater
investments in local public goods in segregated contexts, but also greater ethnic favoritism in the
distribution of such goods where groups are segregated. This argument should apply to local
public goods in contexts where political elites have discretion over distribution of the good, where
the good is in demand from the population, and where the provision of the good is attributable to a
particular individual leader.
We test this argument using data on Malawian members of parliament (MPs) and the pro-
vision of an important local public good within their electoral districts. Drawing on administrative
records, we collect information on the allocation of new water wells (“boreholes”) between 1998
and 2008. We focus the analysis on boreholes because in Malawi these goods are in high demand
and individual MPs have both formal and informal influence over their distribution. We use eth-
1
nicity data from 12,000 localities (roughly 2.3 square miles each) to construct an index of ethnic
group segregation within each electoral district. This index distinguishes electoral districts that
have large ethnic clusters — that is, several proximate localities dominated by one ethnicity —
from electoral districts in which coethnics are not as spatially clustered. We also determine the
share of the population within each locality that is ethnically matched with its MP. With these
measures, we evaluate how segregation affects investments in local public goods across electoral
districts as well as ethnic favoritism in the distribution of these goods within electoral districts.
Our results show that ethnic segregation is indeed a key predictor of both investments in
and allocations of local public goods. First, we show that more boreholes were built in segre-
gated districts, consistent with the expectation that being able to target coethnics with local public
goods within districts encourages politician to invest in these goods. Second, we observe more
ethnic favoritism in the construction of new boreholes when ethnic groups are segregated: using
a difference-in-difference approach, we show that coethnic localities in segregated districts were
20-25% more likely to receive a new borehole between 1998 and 2008 than coethnic localities in
less segregated districts. In a “placebo” test, we find no evidence that segregation shapes MPs’
allocation of private goods, which we would not expect to be affected by segregation. We also dis-
cuss and attempt to rule out potential alternative explanations, including collective action capacity,
MP quality, residential sorting, and plurality group favoritism.
This paper makes several contributions to research on local public goods provision, ethnic
politics, and distributive politics. First, this study focuses explicitly on the nature of the link
between ethnic segregation and local public goods provision, which is implicit in many studies of
African politics. For example, scholars have long recognized that the spatial clustering of ethnic
groups in Africa helps explain why ethnic divisions are so often salient for politics, especially
those surrounding the distribution of state resources (e.g. Bates, 1983; Kasara, 2007; Kimenyi,
2006). While our direct empirical test of this assumption is a contribution to the literature, our
more significant contribution is to make explicit theoretically why segregation should matter for
2
local public goods provision, to demonstrate that segregation can vary considerably even within
one country, and to show systematically that this variation is consequential for distributive politics.
In this way, we add to a large body of existing research that has focused on the role of ethnic
diversity in explaining variation in local public goods provision (e.g., Alesina et al., 1999; Easterly
and Levine, 1997; Habyarimana et al., 2009; Miguel and Gugerty, 2005). We show that areas
with similar levels of ethnic diversity can vary significantly in their degree of ethnic segregation
(see Figure 2), and that this variation is consequential for overall levels of local public goods
provision. Thus, while we are focused primarily on segregation rather than diversity, our theoretical
framework linking ethnic segregation to local public goods investment introduces a top-down,
elite-led explanation for why more diverse localities enjoy fewer local public goods.
Second, the quality of our data, our measure of ethnic segregation, and the subnational
nature of our analysis offer new and more rigorous evidence that segregation shapes the degree
to which political elites favor coethnic constituents in the provision of local public goods. While
numerous studies demonstrate a relationship between segregation and outcomes that may be re-
lated to differential investment in local public goods — intergroup inequality (Alesina et al., 2016;
Baldwin and Huber, 2010), voter expectations of investment (Nathan, 2016), and ethnic voting and
party organization (Alesina and Zhuravskaya, 2011; Ichino and Nathan, 2013; Ishiyama, 2012;
Velasquez, 2013) — the two studies that directly analyze segregation’s effect on ethnic favoritism
do so at the national level and with mixed results (Franck and Rainer, 2012; De Luca et al., 2015).
Our fine-grained census data allow us to use a higher quality measure of ethnic segregation than
these previous studies, and our subnational research design allows us to better isolate the impact
of ethnic segregation on distributive politics.1
1While Franck and Rainer (2012) find no evidence that ethnic favoritism by heads of state is more
pronounced in segregated countries in Africa, De Luca et al. (2015) find that it is, both within
Africa and across regions. Our approach offers several advantages compared to these two existing
studies. First, both studies only consider shared ethnicity with the head of state, and are thus forced
3
Third, we advance a growing literature linking ethnic demography to political outcomes.
We show that ethnic demography not only impacts the behavior and attitudes of voters (Ichino and
Nathan, 2013; Kasara, 2013), but also the distributive strategies of political leaders. In fact, our
paper complements Ichino and Nathan (2013), who argue that local ethnic minorities vote across
ethnic lines in anticipation of benefiting from ethnic favoritism targeted at the majority group. Our
findings confirm their assumption that leaders seek to favor their coethnics with local public goods,
but also demonstrate that this assumption only holds under certain conditions, namely when ethnic
groups are sufficiently segregated.
Finally, our findings have implications for the broader distributive politics literature, which
emphasizes why political elites often favor some groups over others (Golden and Min, 2013).
to leverage cross-country variation in segregation. In contrast, our within-country design ensures
greater homogeneity across units in the study: each MP in Malawi faces a relatively comparable
strategic environment, including the same historical context, electoral system, party system, and
institutional framework for local public goods distribution. Thus, our analysis allows us to control
for potentially important but hard to measure variables that could confound cross-national analysis.
Second, our fine grained, census-based measure of ethnic segregation in Malawi is more appropri-
ate than existing cross-national measures of ethnic segregation. For example, Franck and Rainer’s
(2012) segregation measure is based on the geographic mapping of language groups that assumes
clearly defined boundaries with no overlap (see Matuszeski and Schneider, 2006), and thus, by
design, cannot capture ethnic integration at the local level, an important source of variation in
our own data. Similarly, the measure of segregation used by De Luca et al. (2015), from Alesina
and Zhuravskaya (2011), relies on data that significantly underestimates subnational diversity (see
Gershman and Rivera, 2016). Our measure of ethnic segregation, which uses census data on the
ethnic make-up of more than 12,000 localities, and our sub-national analyses focused on electoral
districts within Malawi, offers a higher quality and more rigorous test of the relationship between
segregation and ethnic favoritism.
4
Sometimes these groups are defined ethnically, while in other contexts they follow caste, partisan,
or religious cleavages. Regardless of how groups are defined, our framework suggests that their
spatial distribution helps define the conditions under which politicians will use local public goods
to engage in favoritism.
Segregation and Local Public Goods Provision
We build on the distributive politics literature to make predictions about how ethnic segregation
shapes politicians’ incentives to provide local public goods.2 The theory should apply where the
following conditions hold. First, the political elite of focus must have some discretion over dis-
tribution of the good of focus. For example, the segregation of an MPs coethnics should not be
consequential if other actors in the political system, such as the president, have greater discretion
over the distribution of the good, or if the good is allocated by formula or by bureaucrats who are
sufficiently insulated from political interference. Second, there should be demand for the good in
the population. If a political leader has discretion over the allocation of a good, but will receive
little credit for providing it, then ethnic segregation may not play a role. Third, and relatedly, voters
should be able to attribute a good to the effort of particular leaders (Harding and Stasavage, 2014;
Harding, 2015). That is, political leaders must be able to claim credit for the provision of the good,
which allows them to secure the electoral and social benefits of targeting coethnics.
Our theory has four components: elite incentives to favor coethnics, budget constraints, the
cost structure of local public goods, and ethnic segregation.
2“Local public goods” are locally non-rivalrous and non-excludable, but costly to access from dis-
tant locations.
5
Incentives for Ethnic Favoritism
Three features of the political environment in much of Africa create incentives for ethnic fa-
voritism. The first incentive arises from differences in politicians’ ability to effectively target
groups of voters with material benefits. As Dixit and Londregan (1996, 1134) note, politicians’
greater understanding of some voters “translates into greater efficiency in the allocation of par-
ticularistic benefits.” This relative efficiency defines a “core” constituency (Cox and McCubbins,
1986). The theoretical literature highlights that politicians are likely to favor their core electoral
districts in contexts where ideological or programmatic differences between parties are small (Cox
and McCubbins, 1986; Dixit and Londregan, 1996), as is largely the case in Africa (Posner, 2005).
In much of Africa, core supporters are ethnically defined: politicians are able to allocate
distributive goods more efficiently to coethnics than to non-coethnics.3 For example, Carlson
(2015) finds that Ugandan voters disproportionately reward the provision of services by coethnic
politicians; Wantchekon (2003) finds that clientelist appeals are more effective when delivered by
coethnics; Kramon (2013) finds that vote buying in Kenya is more effective when politicians target
coethnics; and Adida et al. (2016) find that voters only reward good legislative performance of
coethnic incumbents, but do not reward good performance by non-coethnic incumbents. These
results are likely driven by several factors. Strong expectations of ethnic favoritism, distrust of
out-group politicians, or cognitive biases may cause voters to discount or ignore the provision
of resources by non-coethnic elites (Bates, 1983; Carlson, 2015; Posner, 2005). Politicians may
also be better at engaging politically useful intermediaries in their ethnic home areas (Kasara,
2007). Intermediaries can enhance the efficiency of resource distribution by providing elites with
greater knowledge of their coethnics’ preferences and by monitoring and mobilizing communities
to ensure that they support the incumbent (Nichter, 2008; Stokes et al., 2013).
Second, broader strategic considerations may also drive coethnic targeting. Theories of
3“Efficiency” refers to electoral returns received (the output) for a given input of time and resources.
6
neo-patrimonial politics highlight that there is often an ethnic calculus to coalition-building (Joseph,
1987; van de Walle, 2003). African presidents allocate cabinet positions to elites from different
ethnic groups in exchange for regime support or the delivery of ethnic voting blocs (Arriola, 2009).
These posts come with opportunities for rent-seeking and discretion over the distribution of jobs
and resources. Since cabinet positions are typically allocated to elites who can deliver the support
of their ethnic community, MPs have incentives to maintain strong support amongst their coethnics
in order to enhance their pre- and post-election bargaining position (van de Walle, 2003).
Third, there are social and psychological drivers of coethnic favoritism. Political elites
often face strong informal pressures to take care of their “own” (Lindberg, 2003). Voters generally
expect to benefit when their coethnics are in power (Posner, 2005), and elites may lose social
standing or face social sanctioning if they fail to deliver (Bates, 1983). In Ghana, for example,
Lindberg (2010, 136) finds that “everyday tools of shame, harassment, collective punishment of the
family, and loss of prestige and status” serve as informal pressures on MPs. Moreover, consistent
with social identity theory (e.g., Tajfel and Turner, 1985), elites may derive psychological benefits
from favoring in-group members (Ekeh, 1975).
For all these reasons, we anticipate that politicians will have have incentives to favor coeth-
nic citizens over non-coethnics in local public goods provision. We do not claim that politicians
never have incentives to allocate goods to non-coethnics: we recognize that voters sometimes
support non-coethnic politicians (Ichino and Nathan, 2013; Conroy-Krutz, 2012) and politicians
sometimes provide local public goods to non-coethnic voters. Our theory only requires that the
political or personal returns to coethnic provision are higher, on average, than the returns to non-
coethnic provision.
Budget Constraints
Our second component highlights politicians’ budget constraints. While incumbents often have
incentives to disproportionately serve coethnics, they can choose to do this in a variety of ways. In
7
addition, there are many other political activities they could engage in, such as legislating or raising
campaign contributions. Limited time and resources mean that they must prioritize some activities
over others. Thus, even if they have the discretion to build new local public goods within their
electoral district, they may not do so if they think other activities carry higher political returns. It is
therefore necessary to understand the conditions under which politicians are motivated to allocate
local public goods. The final two components of our theory jointly specify such conditions.
Cost Structure of Local Public Goods
The cost structure of local public goods influences when politicians will be motivated to invest
in them. Local public goods have relatively high fixed costs but relatively low marginal costs.
Compared to providing a private good like cash or an agricultural subsidy, a politician must invest
more resources to ensure that a local public goo is constructed. But once that fixed cost is paid,
additional beneficiaries come at almost no extra cost. This implies that politicians will prefer to
invest in local public goods only when they benefit a sufficient number of (electorally responsive)
residents — coethnics — in a given local community.
Ethnic Segregation
In sum, politicians often have incentives to favor their own ethnic group, must choose among many
potential strategies to do so, and will choose local public goods only if these goods will benefit a
sufficient number of coethnic residents. When these conditions are met, ethnic segregation should
impact the degree to which politicians invest in local public goods as well as where within their
electoral district they choose to allocate these goods.
The logic is demonstrated in Figure 1, which shows two hypothetical electoral districts
with identical levels of diversity, population size, and population density, but different residential
patterns of a politician’s coethnic (gray squares) and non-coethnic (black dots) constituents across
localities (solid lines). While ethnically matched localities have the same percentage of MP co-
8
ethnics in both electoral districts, coethnic localities in the more segregated district are surrounded
by other coethnic localities, but coethnic localities in the less segregated district are interspersed
with localities populated by other ethnic groups. Since the catchment area of local public goods
often crosses locality boundaries (illustrated by the transparent circle), local public goods benefit
more coethnics when coethnic localities are spatially clustered. Because more coethnics benefit
under segregation, there is a higher chance that the relatively high fixed cost of a good can be jus-
tified by MPs under segregation than under integration. This logic generates our first observable
implication:
H1: Investments in local public goods will be greater in electoral districts where members of the
MP’s ethnic group are spatially segregated from members of other ethnic groups.
[Figure 1 about here.]
If greater local public goods provision in segregated electoral districts is indeed driven by incen-
tives for ethnic favoritism, such goods should be disproportionally allocated to coethnic localities
within segregated electoral districts. This expectation is consistent with prior research on ethnic
favoritism in local public goods provision (e.g., Ichino and Nathan, 2013; Burgess et al., 2015),
but adds the novel expectation that segregation will condition the degree to which MPs engage in
ethnic favoritism.
H2: Within electoral districts, ethnic favoritism in the distribution of local public goods will
increase with ethnic segregation.
9
Malawian Context
We test our theory using disaggregated data gathered in the ethnically diverse country of Malawi.4
While there are over twelve distinct ethnic communities, early European contact and subsequent
colonial rule reinforced three main ethno-regional identities in Malawi — the Tumbuka in the
North, the Chewa in the Center, and the Yao in the South — based on the dominant group within
each of the country’s regions (Vail and White, 1991). These ethnic divisions have been relevant
for political behavior at least since the introduction of multiparty elections in 1994 (Posner, 2004),
partly due to their regional segregation. Voting has typically fallen along ethno-regional lines,
although this pattern was weakest in the 2009 election, when the incumbent president received
widespread support across regions (Ferree and Horowitz, 2010).5
We focus on members of Malawi’s unicameral parliament, the National Assembly, who
are elected by plurality vote in 193 single-member electoral districts. Within this first-past-the-
post system, the vote share needed to secure a seat depends on the number of other candidates
contesting. In 1999, the median number of candidates per electoral district was 3, but this increased
to 6 in 2004 and 2009. As a result, the vote share among elected members of parliament (MPs)
decreased from 68% in 1999 to 49% in 2004 and 46% in 2009. Under these conditions, ethnic
favoritism by incumbent MPs is electorally viable whenever the MP’s ethnic group comprises
a plurality of the electoral district. Under the period of study considered here, Malawian MPs
4Chewa are the largest group (33%), followed by the Lomwe (18%), Yao (14%), Ngoni (12%),
Tumbuka (9%), and seven smaller groups (Government of Malawi, 2008). There is significant
variation in segregation across Malawi (Figures SI.1 and SI.6 of Supporting Information).5The 2009 respite from ethno-political voting is typically attributed to the precarious position of the
president, Bingu wa Mutharika, after defecting from the ruling United Democratic Front (UDF)
and establishing his own party, which forced him to extend state-based patronage to areas beyond
the UDF-dominated Southern Region (Ferree and Horowitz, 2010).
10
matched the plurality ethnic group in more than 70% of electoral districts. However, even when the
MP’s ethnic group does not constitute a plurality, he or she may still have non-electoral incentives
for favoring coethnics, as discussed above. This may be especially true in the Malawian context,
where reelection rates among MPs are quite low (32% in 2004 and 25% in 2009).
One way for MPs to favor coethnics within their electoral district is through the targeted
provision of local public goods. As in many African countries, Malawi has an institutional struc-
ture in which politicians exert significant leverage over the allocation of local public goods, and
MPs play a crucial role in the planning, funding, and management of such goods in their elec-
toral districts. Formal responsibility for the provision of these goods lies with District Assemblies,
which by law comprise MPs and locally-elected councilors (Chinsinga, 2005). However, local-
level elections for councilors were not held until 2000 and after their first term expired in 2005,
councilors were never again elected during the period under study. Thus, local development initia-
tives were largely left to MPs and centrally-appointed district officials (Chasukwa et al., 2014), but
MPs also heavily influence the decisions made by district officials (O’Neil et al., 2014). MPs also
exert considerable informal influence over the allocation of local public goods. As local “big men,”
they lobby for and influence development projects funded by the central government and NGOs
(Cammack et al., 2007; Chasukwa et al., 2014). As a result of this discretion, MPs have increas-
ingly focused on delivering development projects (Cammack et al., 2007; Chinsinga, 2007, 2009),
a trend that mirrors dynamics in other parts of Africa (Lindberg, 2010). Voters’ expectations that
MPs should provide public goods is reflected in public opinion data: surveys from 2003 and 2007
show that a vast majority of Malawians would prefer an MP who delivered local public goods over
one who implemented sound public policy and produced nationally beneficial legislation (Mthinda
and Khaila, 2006; Tsoka and Chinsinga, 2009).
Our main analysis focuses on MPs and the provision of new water wells — “boreholes” —
for three reasons. First, demand for boreholes is high across Malawi (DeGabriele, 2002). Almost
half of all rural Malawians had no access to a protected water source in 1998 (Government of
11
Malawi, 1998), and boreholes are overwhelmingly the main protected water source in rural Malawi
(Baumann and Danert, 2008), although boreholes are in used in urban areas as well (National
Statistical Office of Malawi and ORC Macro, 2001).6 Dionne (2012) reports that rural Malawians
and village headmen in three rural districts ranked access to clean water as their community’s
single greatest need.
Second, MPs have significant discretion over the provision of boreholes in their electoral
districts. Many reports on water access in Malawi note the pervasive influence of politics and
favoritism, especially on behalf of MPs, in the construction of new boreholes (e.g., Ferguson and
Mulwafu, 2004; WaterAid, 2008, 2010). The relatively low cost of borehole provision — roughly
$5000 (Baumann and Danert, 2008) — means that MPs can use their personal or CDF funds to
provide them on their own, giving the MP full discretion over provision. In Dowa, for example, an
MP was hailed by constituents for drilling 125 boreholes over three years using “personal money
through her development office” (The Malawi Voice, 2014). MPs also impact the placement of
government funded borehole projects by, for example, lobbying the Ministry of Irrigation and
Water Development or through their outsized influence in district development councils (O’Neil
et al., 2014). MPs also influence the placement of boreholes provided by other actors, such as
international NGOs, through informal pressure or partnerships. One Malawian MP described this
process as going “shopping for people who can assist” once she could no longer fund additional
development in her district (Gilman, 2009, p.198).
Finally, borehole provision can be directly attributed to MP effort. For example, MPs claim
and receive credit for borehole projects, even when they have not provided the direct funding: the
6According to DHS data gathered in 2000 (National Statistical Office of Malawi and ORC Macro,
2001), about 40 percent of the rural population had access to a borehole, while another 40 percent
did not have access to a clean water source. These people rely on unprotected wells and surface
water (e.g., lakes or streams). Twelve percent had access to a community stand pipe, and only
about 2 percent had access to piped water.
12
MP for Zomba-Likangala, for example, was credited with building a borehole despite the funds
being provided by an international NGO (The Nation, 2012). When MPs use their CDF funds
on water projects, constituents understand that the resources have come from the MPs personal
fund (The Malawi Voice, 2014). Constituents also hold MPs accountable for a lack of borehole
provision, and there is an example of constituents seeking to replace their MP precisely because
he failed to provide boreholes in his constituency (Nyirenda, 2014). In sum, boreholes are in
high demand in Malawi, MPs have significant discretion over their allocation, and constituents
generally attribute their provision to MP effort. These characteristics make boreholes an excellent
local public good with which to test the theory’s implication for ethnic favoritism by Malawian
MPs.
Data and Measurement
We assemble data at two different geographic levels. Our smallest units of observation are 12,380
census enumeration areas, which we call “localities.” On average, 1,000 people reside in these lo-
calities (Table SI.2 in Supporting Information (SI)). Because the localities are small — on average
6 square kilometers — the catchment area of many local public goods crosses locality boundaries.
Our theory thus predicts that the decision to provide a public good to a given locality will depend
on that locality’s ethnic connection to the political leader and on the political leader’s ethnic con-
nection to surrounding localities. Our second units of observation are 193 electoral districts, within
which localities are nested.7 On average, an electoral district includes 64 localities.
We construct three key measures. First, we extract the geographic coordinates of all bore-
holes from maps produced from the 1998 and the 2008 censuses, which were provided to us by
the National Statistics Office. By subtracting the boreholes that were already present in 1998 from
those present in 2008, we determine the location of all new boreholes built during that ten-year
7Electoral districts are nested within 28 administrative districts.
13
period (see Figure SI.2 in the SI). From this, we construct an electoral district-level count of the
number of new boreholes (on average, 39) and a dichotomous locality-level indicator for whether
or not each locality received a new borehole between 1998 and 2008 (33% did).8
Second, we assemble an original dataset on the ethnic identity of each Malawian MP who
served between 1994 and 2009 (details in SI). We combine this information with census data
on the ethnic make-up of each locality to create two measures of ethnic match between an MP
and each of the localities within his or her electoral district.9 We create two indicators of the
MP’s ethnic linkage with a locality. The first measure, Match, is equal to 1 if the MP was of the
same ethnicity as the largest group within that locality at anytime between 1999 and 2008, and
0 otherwise. By this measure, 76 percent of localities were matched at some point. The second
measure, Match Proportion, is equal to the proportion of the locality’s population from the MP’s
ethnic group.10 The average proportion of the locality’s population from the same ethnic group
as the MP was 0.59. Figure SI.3 maps the spatial variation in these ethnic match variables across
Malawi. Ethnically matched localities exist in large numbers in electoral districts at all levels of
ethnic segregation, making it possible for MPs to favor ethnically matched localities in even the
least segregated settings.
Finally, we calculate a measure of ethnic segregation for each electoral district based on
8Because the vast majority of localities (84%) received either one borehole or no new borehole, we
use a dichotomous indicator of receiving at least one new borehole.9For each locality in the 2008 census, we know the total population and the proportion of the
population belonging to each ethnic group. While it would be ideal to measure ethnicity prior
to 1998, the 1998 census did not ask about ethnicity. We discuss the possibility of residential
sorting in the Alternative Explanations section.10If the ethnicity of the MP changed between the legislative terms, we average across the proportions
for each term.
14
the ethnic demography of all localities within it.11 We employ the spatial dissimilarity index
(Reardon and O’Sullivan, 2004), a widely used measure of segregation, which ranges between
0 and 1 with higher values indicating greater segregation (details in SI). Using this index, we
measure how segregated the MP’s ethnic group is from other ethnic groups in each electoral district
across the two legislative terms in 1998-2008.12 If the ethnicity of the MP changed between
the legislative terms, we average across the two MP-specific segregation measures. We do not
measure segregation for the ten most ethnically homogenous electoral districts (ELF < 0.05):
ethnic segregation is only meaningful with at least some ethnic diversity, and a small number
of minority group members exert undue influence on segregation measures amid low diversity
(Reardon and O’Sullivan, 2004).13
To illustrate what our segregation measure captures, Figure 2 shows that two electoral dis-
tricts with similar levels of diversity (ELF scores of 0.51 and 0.65) can differ markedly in their
degree of segregation (segregation scores of 0.70 and 0.21). Figure SI.6 further emphasizes the
degree of variation in segregation at all levels of ethnic diversity. In addition to this continuous
measure of segregation, we also classify each electoral district into low, medium, or high segrega-
tion categories based on terciles of the spatial dissimilarity index: low segregation is below 0.401,
medium between 0.401 and 0.490, and high above 0.490.14 Figure SI.5 in the SI shows example
11In particular, our input is the proportion of each locality’s population that belongs to each of the
following 12 ethnic groups: Chewa, Lambya, Lomwe, Ngonde, Ngoni, Nyakusa, Nyanja, Sena,
Senga, Tonga, Tumbuka, and Yao.12This MP-specific measure of segregation is more relevant to our theory than a weighted measure
of segregation aggregated across all groups. In practice, the two measures are highly correlated
(r = 0.97).13Results are robust to including all electoral districts (see SI).14The proportion of the 1,315 localities in low segregation districts that are ethnically matched with
their MP is 0.47. The same proportion is 0.50 for the 1,542 localities in moderately segregated
15
electoral districts in each category, which have segregation indices roughly equal to the median for
each category, and Figure SI.4 of the SI shows the variation in segregation across electoral districts.
[Figure 2 about here.]
Segregation and Local Public Goods Provision Across Electoral Districts
Our theory predicts that investments in local public goods should be higher in ethnically segregated
electoral districts (H1). Figure SI.7 shows a positive bivariate relationship between the number of
boreholes built in 1998-2008 and ethnic segregation across Malawi’s electoral districts. To account
for several potential confounders, we rely on a regression framework. Because our outcome vari-
able is the count of boreholes built in an electoral district, we use a Poisson model modified to
account for overdispersion in the data (Gelman and Hill, 2006; Wooldridge, 1997).15 We model
the number of new boreholes an electoral district receives (yd) as follows:
yd ∼ overdispersed Poisson(θd, ω) , θd = exp(αa[d]+βSegd +X ′dγ
)(1)
where ω is an overdispersion parameter estimated from the data, and where d indexes electoral
districts and a administrative districts. Our main variable of interest is Seg, which measures ethnic
segregation. In Equation 1, Seg is continuous, which assumes a linear relationship (on a log-count
scale) between segregation and borehole investments. To allow for non-linearities, we also present
results from a model that includes two dummy variables indicating medium and high segregation,
leaving electoral districts with low segregation as the omitted reference category. We include in
vector Xd a set of electoral district-level covariates that are likely predictors of borehole invest-
ments. In a first model, we include controls for ethnic diversity (ELF), the (natural log of the)
districts and 0.32 for the 645 highly segregated district.15This approach allows us to relax the assumption that the conditional variance and mean are equal,
and guards against understating the standard errors.
16
proportion of the electoral district’s area that is urban,16 land area in square kilometers, and bore-
holes per 10,000 residents in 1998. Together, these variables capture indicators of collective action
capacity, as well as demand and need for boreholes. In a second model, we expand the list of
covariates to include the degree of MP electoral competition, MP coethnic share of the population,
and presidential coethnic population share, all of which help account for national-level political in-
fluences. We also further control for need using the number of NGO-funded water aid projects per
capita and all aid projects per capita using geo-coded project locations from the AidData project
(Strandow et al., 2011), and each electoral district’s accessibility with a measure of distance to the
nearest major city (Lilongwe or Blantyre). In all models we also include administrative district
fixed effects, αd[c], because important decisions, including borehole allocation, are often made at
this level.17
The results in Table 1 show that segregation is a robust positive predictor of new borehole
investments. The coefficients on segregation are positive, statistically significant, and substantively
large. Given Model 1, and holding covariates at their mean or mode, we would expect highly
segregated districts (90th percentile on our segregation index) to invest in 17 more boreholes than
less segregated districts (10th percentile), with a 95% confidence interval (CI) of 3 to 39.18 The
16The spatial location of urban areas in Malawi is captured using the Global Rural-Urban Mapping
Project (GRUMP) data (Balk et al., 2006; Global Rural-Urban Mapping Project, 2011). GRUMP
defines the extent of urban areas based on population, nightlight data, and settlement points, and
we calculated the proportion of each electoral district that is urban based on GRUMP’s mapping.17Members of parliament serve on the development committees for administrative districts, and
exercise the significant influence over the allocation of development projects through that body
(Chinsinga, 2008; Chiweza, 2010; O’Neil et al., 2014).18Throughout, we generate expected values and confidence intervals based on 10,000 simulations
that approximate the sampling distribution of the parameters in the model (King et al., 2000; Gel-
man and Hill, 2006, Ch. 7).
17
effects are comparable to the effect of ethnic diversity: highly diverse electoral districts (90th
percentile on ELF) invest in 18 fewer boreholes, on average, than low diversity electoral districts
(10th percentile), with a 95% CI of -46 to -2. Model 3 and 4, which use two dummy variables
instead of a continuous measure of segregation, confirm these results. Given Model 3, we would
expect electoral districts with medium segregation to invest in 11 more boreholes than electoral
districts with low segregation (95% CI: 3 to 24), and electoral districts with high segregation to
invest in 10 additional boreholes (95% CI: 1 to 23).
[Table 1 about here.]
Segregation and Ethnic Favoritism Within Electoral Districts
We next evaluate whether ethnic favoritism within electoral districts increases with segregation
(H2). We use a set of difference-in-differences to test this hypothesis. We examine the 3502
localities in 120 electoral districts that were not ethnically matched with their MP prior to 1999
(based on parliamentary elections in 1994). In the 1999 and 2004 elections, 55 of those 120
districts experienced a change in the ethnicity of their MP, resulting in 1599 localities becoming
ethnically matched with their MP and 1903 localities remaining unmatched. Thus we observe two
groups of localities in two time periods: Group 1 localities were not matched in the first period
(1994-1998) or in the second period (1999-2009), whereas Group 2 localities were not matched in
the first period but became matched in the second (i.e., Group 2 localities experienced a “coethnic
switch”).
The goal of our difference-in-differences approach is to estimate the effect of the coethnic
switch experienced by Group 2, using the time-trend of Group 1 as a counterfactual. This ap-
proach allows us to hold constant any time-invariant locality characteristics that influence public
goods provision, including local ethnic diversity, collective-action capacity, and locality demand
18
for public goods.19 As discussed above, we implement this approach using Match, a dummy vari-
able equal to 1 for matched localities in the second time period. (That is, Match equals 0 for Group
1 localities in both time periods and for Group 2 localities in the first time period.) We then esti-
mate the probability that a locality i has a borehole in year t ∈ {1998,2008} using the following
model:
yit = αi + γt +βMatchit + eit (2)
We include locality fixed effects, represented by αi, as well as a time period fixed effect (γt).20 The
outcome, yit , is a dummy variable indicating the presence of a borehole. We use this equation to
estimate β, which gives the change in probability of borehole provision given a coethnic switch
(relative to no coethnic switch), which we interpret as the degree of ethnic favoritism.
To estimate how ethnic favoritism is conditioned by segregation, we use two strategies.
First, we run Equation 2 among three subsets of electoral districts based on their levels of segre-
gation. We again use the terciles of the spatial dissimilarity measure to group districts into low,
medium, and high segregation. Second, we add interactions to Equation 2, interacting indicators
for medium and high segregation with Match. We also run a model that interacts Match with a
continuous measure of segregation. We estimate linear probability models and cluster the standard
errors on localities to account for the panel structure of the data.21 We repeat all analyses with the
19For differential demand to account for the DiD results, newly matched localities would have to
experience greater increases in demand for water than localities who remained unmatched and this
differential increase in demand would have to occur only in segregated electoral districts, which
seems unlikely.20With two time periods, the latter is simply an indicator for the second time period.21We show in the SI that the results are robust to a logistic specification (Figure SI.8). Our main
analysis clusters standard errors on locality because this is the level at which “ethnic match” is
19
continuous indicator of MP-locality ethnic match, Match Proportion.
To illustrate the approach, we begin by implementing a simple non-parametric test, pre-
sented in Figure 3. We calculate the proportion of localities that have at least one borehole, by
time period (1998 versus 2008) and whether the locality experienced a coethnic switch. Thus, we
are simply comparing four means at each level of segregation. This analysis indicates little or no
ethnic favoritism at low levels of segregation: 2.3 percent of Group 1 localities and 4.5 percent of
Group 2 localities had a borehole prior to 1998, which increased to 30.1 percent (Group 1) and 35.3
percent (Group 2) by 2008. In contrast, ethnic favoritism was prevalent in moderately and highly
segregated elections districts. In these districts, Group 1 and 2 localities had similar levels of bore-
hole provision prior to 1998, but localities that experienced a coethnic switch had a much higher
chance of receiving a borehole by 2008. The difference-in-differences is 15.2 percentage points in
election districts with medium levels of segregation, and 22.1 percentage points in districts with
high levels of segregation.
[Figure 3 about here.]
The regression results in Table 2 confirm that segregation spurs ethnic favoritism after ad-
justing for confounders. When we run Equation 2 in three subsets of election districts, based on
their segregation levels, we find evidence of ethnic favoritism only in moderately and highly seg-
regated electoral districts (Models 1-3 in Panel A). This finding remains when we run models that
include all electoral districts and interact Match with two segregation dummies (Models 4-6). In
moderately segregated districts, localities that experienced a coethnic switch were 7-16 percentage
points more likely to receive a borehole than localities that were never ethnically matched, a dif-
assigned. This approach is similar to Franck and Rainer (2012), who cluster on ethnic group-
survey round, and Burgess et al. (2015), who cluster on district, the levels at which ethnic match
is assigned in their respective studies. However, in Table SI.11 of the SI, we show results using a
more conservative approach to clustering — at the electoral district-year level — and the results
are, in most cases, robust.
20
ference that can be distinguished from 0 at conventional levels of confidence. In highly segregated
districts, localities with a coethnic switch were 15-21 percentage points more likely to receive a
borehole than localities without a coethnic MP (p < 0.01). These results are robust to the inclusion
of a set of time-varying covariates, including the locality’s ethnic match with the president and
the presence of other local public goods (Model 5). The results are also robust to the inclusion of
controls that interact the time period dummy with fixed characteristics of the locality (Model 6).22
Finally, we also find similar results when we interact Match with a continuous measure of electoral
district segregation (Model 7). In short, we find robust evidence of ethnic favoritism, which is more
pronounced in moderately and highly segregated electoral districts.23
[Table 2 about here.]
Panel B of Table 2 presents results of analyses that use the proportion of the MP’s coethnics
in the locality to define an ethnic match (i.e., these analyses replace Match with Match Proportion).
We find that all indicators of Match Proportion are statistically significant but that the magnitude
of the effect substantially increases with the degree of segregation, suggesting a greater degree of
ethnic favoritism in segregated electoral districts (Models 1-3). In low segregation districts, in-
creasing the proportion of the MP’s coethnics in a locality from 0 to 0.5 (a fairly typical change)
corresponds to about a 6.5 percentage point increase in the probability that the locality receives a
borehole. In high segregation districts, this same change predicts about a 28.5 percentage point
increase in the probability that the locality receives a borehole. This pattern is confirmed when we
22We include interactions with an indicator of urban/rural, population density, ethnic diversity, land
area, distance to Lilongwe and Blantyre, and number of boreholes per capitra in 1998. The com-
plete results are presented in Table SI.4 in the SI.23We present a range of robustness tests of these results in the SI, including alternative segregation
cutpoints (Figure SI.9) and a parametric test of how ethnic favoritism varies as a continuous func-
tion of segregation (Figure SI.8). The SI also shows that the DiD results are largely stable with the
removal of urban electoral districts from the sample (Table SI.8).
21
interact Match Proportion with the two segregation dummies (Models 4-6). The interaction be-
tween Match Proportion and the continuous measure of segregation is also statistically significant
and in the expected direction (Model 7). As above, all results are robust to the inclusion of time
varying controls and time period dummy interactions with a range of locality characteristics.
In sum, we find strong evidence that the prevalence and degree of ethnic favoritism in
electoral districts is increasing with ethnic segregation.24 Coupled with our electoral district-level
results, there is substantial empirical support for our theoretical framework: segregation shapes
both MP investment in local public goods and ethnic favoritism with respect to their geographic
allocation.25
24Because the DiD analysis focuses on the set of localities that experienced a change in the Match
variable from 1998 to 2008, we conduct an additional set of cross-sectional analyses that examine
ethnic favoritism in the full set of localities (Table SI.12). While we prefer the DiD analysis,
which controls for time-invariant differences in localities’ probability of receiving a borehole and
for common time shocks across localities (at least for a given level of segregation), the cross-
sectional results are less precise but also consistent with the argument that segregation conditions
ethnic favoritism across a large number of localities.25We also examine whether our conclusions extend to two other public goods: health clinics and
schools. The results are presented in Tables SI.13–SI.18. The results are more mixed and generally
weaker than the borehole results. This is likely due to the fact that MPs have less discretion over
the provision and allocation of clinics and schools, which are constructed at far lower rates (for
example, less than 2% of localities received a new clinic between 1998 and 2008) and may be more
heavily influenced by political decisions at the national level. Consistent with this interpretation,
Kramon and Posner (2013) find evidence that coethnicity with the president in Malawi is associated
with greater education and health outcomes, but find no evidence that Malawian presidents favor
their coethnics with clean water access. For MPs, the scope conditions of our theory are better met
by the types of investment that MPs regularly make in these sectors, which fall short of providing
22
Alternative Explanations
While the empirical patterns reported above are consistent with our theory, this section discusses
and empirically assesses a number of alternative explanations.
Local Ethnic Homogeneity and Collective Action. One alternative centers on the expecta-
tion that homogenous localities are better able to collectively mobilize to locally produce public
goods (Miguel and Gugerty, 2005; Habyarimana et al., 2009). If local public goods are locally pro-
duced at a higher rate in homogenous localities and segregated electoral districts in general have
more homogenous localities than integrated ones, then segregated electoral districts could mechan-
ically have more public goods. This explanation is, however, inconsistent with our locality-level
results (Figure 3 and Table 2), which show that ethnically matched localities are the primary bene-
ficiaries of local public goods in segregated electoral districts. If locality ethnic homogeneity alone
were driving our results, we would not expect the effect of segregation to be conditional on ethnic
match with an MP.
Collective action capacity could also interact with a supply-side mechanism to produce our
locality-level results. For example, MPs may be more responsive to bottom-up pressures from
coethnic communities than from non-coethnic communities. While we agree that the ability of
communities to pressure coethnic leaders is important, this alternative explanation cannot explain
why we find weak evidence of ethnic favoritism in less segregated electoral districts, where rel-
atively homogenous localities that are coethnic with the MP do exist. It could, however, be that
segregation facilitates collective action to demand local public goods across localities, and so part
of the reason segregation is influential is because of a demand side mechanism. Although we
cannot completely rule this out, we believe this channel is likely to be less important than the sup-
an entirely new school or clinic. For example, MPs are likely to strategically target the provision
of school toilets or beds for a clinic, but we are unable to observe these types of investments with
census data.
23
ply side mechanism our theory highlights, as homogenous coethnic localities in less segregated
electoral districts should also be able to demand local public goods from the MP.
MP Quality. Another alternative explanation is that there is a correlation between segre-
gation and the quality of the MP, producing a spurious relationship between segregation and local
public goods provision. To rule out this explanation, we carry out a placebo test that examines
whether segregation also affects the provision of private goods in the form of agricultural sub-
sidies. Like local public goods, agricultural subsidies are highly valued by Malawian residents
(Harrigan, 2008) and political elites exert discretion over their distribution (Chasukwa et al., 2014;
Øygard et al., 2003; Tambulasi, 2009). But, unlike local public goods, they can be politically tar-
geted to specific individuals or households, meaning that segregation should be less consequential
for their strategic provision. In the SI, we show that the provision of these goods is not affected by
segregation (Table SI.19).26
Residential Sorting. If Malawians move in response to the provision of local public goods,
then our ability to detect the effect of ethnic demography on their provision could be threatened.
However, we anticipate that such residential sorting would lead to more diverse populations, and
thus more integration, near local public goods, as migrants move towards better served areas —
the opposite of what we observe. Furthermore, rural-rural migration in Malawi is relatively con-
strained due to the scarcity of land and customary rules governing land tenure (Chirwa, 2008;
26In addition, while we find evidence of ethnic favoritism in the distribution of these goods, the de-
gree of ethnic favoritism is greatest in the most integrated electoral districts and decreasing with
segregation (Table SI.20). This finding is at odds with recent research showing no evidence of eth-
nic favoritism in the distribution of agricultural subsidies withinin Malawi (Dionne and Horowitz,
2016). However, the reported null effect was for presidential coethnics, suggesting that MP coeth-
nicity is more important for favoritism in the provision of private goods than presidential coethnic-
ity. This second finding allays a separate potential concern: that some residents are better able to
get a coethnic leader elected and be more effective in lobbying for public resources.
24
Kishindo, 2004).27 What rural-rural migration does exist is unlikely to shift the ethnic landscape
because both marriage and accessing communally held land typically occur within ethnic commu-
nities.28 Rural-rural migration across ethnic communities is typically limited to laborers on large
tobacco or tea estates (Potts, 2006), areas which are likely to have more, not less, local public goods
provision. Taken together, these patterns of migration suggest that residential sorting is unlikely to
account for our results.
Plurality Group Favoritism. Finally, we interpret our results as evidence of in-group fa-
voritism. It is possible, however, that MPs are instead targeting benefits to the largest ethnic group
in an electoral district, whether it is their own group or not, in order to maximize their electoral
coalition. With few districts in which the MP is not a member of the ethnic plurality, we cannot
distinguish plurality group favoritism from coethnic favoritism. We note, however, that the inter-
pretation we have offered is plausible in light of the existing evidence that politicians in much of
Africa have incentives to favor their own ethnic group. Further, this alternative interpretation does
not undermine our general argument: regardless of the group that the political elite is seeking to
favor, our logic suggests that the segregation of that group shapes how it is favored.
Conclusion
This paper advances a theory about how ethnic segregation shapes elite strategies for engaging in
ethnic favoritism. We show that more boreholes – an important local public good in the Malawian
context – are allocated to electoral districts where ethnic groups are spatially segregated, and that
27Census data shows that only 10% of rural Malawians reside outside their district of birth. This
figure is based on individual-level information about districts of birth and residence for a random
10% sample (n = 1,282,335) of the 2008 census data (Minnesota Population Center, 2014).28Customary and cultural barriers limit access to land outside one’s ethnic community (Potts, 2006),
and most marriages are formed within 5 miles of one’s home village (reported in Englund, 2002).
25
ethnic favoritism in borehole provision is more common in segregated contexts. These patterns are
consistent with our claim that ethnic segregation conditions how elites invest in and allocate local
public goods.
Our theory and results make several contributions to the study of ethnic politics in Africa
and distributive politics more broadly. First, they underscore the importance of ethnic segrega-
tion in understanding distributive politics in diverse contexts. In particular, our within-country
research design and high quality data on ethnic group distributions provide compelling evidence
that segregation indeed affects ethnic favoritism, despite mixed results from studies evaluating this
relationship cross-nationally (Franck and Rainer, 2012; De Luca et al., 2015).
Our framework also helps make sense of outstanding puzzles in the literature on ethnic
politics in Africa. For example, while ethnic favoritism is pervasive in some contexts, it is not
universal (Franck and Rainer, 2012). Nor is there ethnic favoritism in the allocation of all distribu-
tive goods in a given context (Kramon and Posner, 2013). Our theory contributes by specifying
the conditions under which ethnic favoritism should manifest in local public goods provision, as
the geographic reach of different types of goods will define the scale at which ethnic composition
matters. Our theory also has implications for the question of why local ethnic diversity is often
associated with low public goods provision. While past explanations focus on local collective ac-
tion (Alesina et al., 1999; Habyarimana et al., 2009; Miguel and Gugerty, 2005), our framework
suggests that political leaders under-invest in public goods in highly diverse local areas because
such goods are too difficult to target to their coethnic supporters. Thus, distributive politics may
help to account for the under-provision of public goods in ethnically diverse areas.
Our study also contributes to recent work on ethnic geography and vote choice. While
we do not observe vote choice in Malawi, our theory implicitly generates expectations about the
relationship between ethnic segregation and ethnic-based voting. Past research has found that the
geographic concentration of ethnic groups is positively associated with ethnic bloc voting and the
existence of ethnic parties (Alesina and Zhuravskaya, 2011; Ishiyama, 2012; Velasquez, 2013),
26
our theory suggests that geographically segregated groups will tend to vote ethnically because they
anticipate that local public goods will be targeted to their area. Consistent with this expectation,
Nathan (2016) finds that variation in ethnic segregation across urban neighborhoods in Ghana pre-
dicts ethnic voting, which he attributes to the (unobserved and untested) expectation that politicians
provide different types of goods to localities with different ethnic geographies. In rural Ghana,
Ichino and Nathan (2013) find that citizens who make up a local ethnic minority are willing to
vote for a non-coethnic presidential candidate, and argue that this is because they expect to ben-
efit from the local public goods targeted toward the ethnic majority. Our study is consistent with
such voter expectations, but also implies that local ethnic minorities should be most likely to vote
across ethnic lines in contexts of high ethnic segregation. Future research should directly assess
the relationship between segregation and vote choice, as well as evaluate the electoral returns to
strategically targeting local public goods provision.
Finally, while we test the theory in Malawi, we expect the argument to generalize to other
contexts for two reasons. First, Malawi is similar to many other countries in that political elites
have incentives to favor some groups over others. Research on distributive politics shows this
to be the case in a range of socio-economic and institutional contexts: in Australia, a wealthy
democracy with single-member districts (Denemark, 2000); in Sweden, a wealthy democracy with
proportional representation (Dahlberg and Johansson, 2002); in India, a developing democracy
with single-member districts (Min, 2015); in Benin, a developing democracy with proportional
representation (Kramon and Posner, 2013); and in Egypt, an electoral authoritarian regime (Blay-
des, 2010). Second, because our theory emphasizes the importance of segregation in shaping the
type of goods used to favor one group over others, the theory can be applied to the study of fa-
voritism in contexts where elites have discretion over different types of goods (private and public).
In urban Ghana, for example, Nathan (2016) finds that voters expect elites to distribute different
types of goods to neighborhoods with different ethnic demographies, which is consistent with our
framework. Research from Latin America documents that governments often invest in a differ-
27
ent mix of public and private goods in different local political contexts (Albertus, 2012; Magaloni
et al., 2007), patterns that our logic may help to explain. Thus, while more research is required, we
anticipate that segregation may shape distributive politics in contexts with different institutional
configurations, degrees of urbanization, and levels of economic development. In short, our cen-
tral finding — that ethnic segregation conditions the strategies that incumbents use to favor their
coethnics — has implications for the study of distributive politics beyond Malawi, and beyond
Sub-Saharan Africa. Wherever political elites have incentives to favor certain groups of voters
over others, the spatial distribution of these groups is likely to shape the distributive strategies they
adopt.
28
References
Adida, C., Gottlieb, J., Kramon, E., and McClendon, G. (2016). How coethnicity moderates theimpact of information on voting behavior: Experimental evidence from Benin. Presented at theAnnual Meeting of the American Political Science Association, Philadelphia, PA.
Albertus, M. (2012). Vote buying with multiple distributive goods. Comparative Political Studies,46(9):1082–1111.
Alesina, A., Baqir, R., and Easterly, W. (1999). Public goods and ethnic divisions. QuarterlyJournal of Economics, 114(4):1243–1284.
Alesina, A., Michalopoulos, S., and Papaioannou, E. (2016). Ethnic inequality. Journal of PoliticalEconomy, 124(2):428–488.
Alesina, A. and Zhuravskaya, E. (2011). Segregation and the quality of government in a crosssection of countries. American Economic Review, 101(5):1872–1911.
Arriola, L. (2009). Patronage and political stability in Africa. Comparative Political Studies,42(10):13–39.
Baldwin, K. and Huber, J. D. (2010). Economic versus cultural differences: Forms of ethnicdiversity and public goods provision. American Political Science Review, 104(4):644–662.
Balk, D. L., Deichmann, U., Yetman, G., Pozzi, F., Hay, S. I., and Nelson, A. (2006). Determin-ing global population distribution: Methods, applications and data. Advances in Parasitology,62:119–156.
Bates, R. (1983). Modernization, ethnic competition, and the rationality of politics in contempo-rary Africa. In Rothchild, D. and Olorunsola, V. A., editors, State versus Ethnic Claims: AfricanPolicy Dilemmas, pages 152–71. Boulder, CO: Westview Press.
Baumann, E. and Danert, K. (2008). Operation and maintanance of rural water supplies in Malawi.Swiss Resource Centre and Consultances for Development Report.
Blaydes, L. (2010). Elections and Distributive Politics in Mubarak’s Egypt. New York: CambridgeUniversity Press.
Briggs, R. C. (2014). Aiding and abetting: Project aid and ethnic politics in Kenya. World Devel-opment, 64:194–205.
Burgess, R., Jedwab, R., Miguel, E., Morjaria, A., and Padro i Miquel, G. (2015). The value ofdemocracy: Evidence from road building in Kenya. American Economic Review, 105(6):1817–51.
29
Cammack, D., Golooba-Mutebi, F., Kanyongolo, F., and O’Neil, T. (2007). Neopatrimonial poli-tics, decentralisation and local government: Uganda and Malawi in 2006. In Good Governance,Aid Modalities and Poverty Reduction, Research project (RP-05-GG):. Advisory Board for IrishAid.
Carlson, E. (2015). Ethnic voting and accountability in Africa: A choice experiment in Uganda.World Politics, pages 1–33.
Chasukwa, M., Chiweza, A. L., and Chikapa-Jamali, M. (2014). Public participation in localcouncils in Malawi in the absence of local elected representatives: Political elitism or pluralism?Journal of Asian and African Studies, 49(6):705–720.
Chinsinga, B. (2005). District assemblies in a fix: The perils of self-seeking tendencies in decen-tralisation policy reforms in Malawi. Development Southern Africa, 22(4):529–548.
Chinsinga, B. (2007). Democracy, Decentralisation, and Poverty Reduction in Malawi. RüdigerKöppe Verlag, Cologne, Germany.
Chinsinga, B. (2008). Decentralisation and poverty reduction: Malawi – a critical appraisal. InCrawford, G. and Hartmann, C., editors, Decentralisation in Africa: A Pathway Out of Povertyand Conflict?, pages 73–106. Amsterdam University Press, Amsterdam, Netherlands.
Chinsinga, B. (2009). The interface between local level politics, constitutionalism and state for-mation in Malawi through the lens of Constituency Development Fund (CDF). Paper presentedat “Democratization in Africa: Retrospective and Future Prospects,” Leeds.
Chirwa, E. (2008). Land tenure, farm investments, and food production in Malawi. IPPG PaperNo. 18.
Chiweza, A. L. (2010). Public-sector reforms and decentralisation of public services: Lessonsfrom Malawi (1994-2006). In Tambulasi, R. I., editor, Reforming the Malawian Public Sector:Retrospectives and Prospectives, pages 29–55. Council for the Development of Social ScienceResearch in Africa (CODESRIA).
Conroy-Krutz, J. (2012). Information and ethnic politics in Africa. British Journal of PoliticalScience, 43(345-73).
Cox, G. W. and McCubbins, M. D. (1986). Electoral politics as a redistributive game. Journal ofPolitics, 48(2):370–389.
Dahlberg, M. and Johansson, E. (2002). On the vote purchasing behavior of incumbent govern-ments. American Political Science Review, 96(1):27–40.
De Luca, G., Hodler, R., Raschky, P. A., and Valsecchi, M. (2015). Ethnic favoritism: An axiomof politics. Working Paper.
DeGabriele, J. (2002). Improving community based management of boreholes: A case study fromMalawi. Land Tenure Center, University of Wisconsin-Madison.
30
Denemark, D. (2000). Partisan pork barrel in parliamentary systems: Australian constituency levelgrants. Journal of Politics, 62(3):896–915.
Dionne, K. Y. (2012). Local demand for a global intervention: Policy priorities in the time ofAIDS. World Development, 40(12):2468–2477.
Dionne, K. Y. and Horowitz, J. (2016). The political effects of agricultural subsidies in Africa:Evidence from Malawi. World Development, 87:215–226.
Dixit, A. and Londregan, J. (1996). The determinants of success of special interests in redistributivepolitics. Journal of Politics, 58(4):1132–1155.
Easterly, W. and Levine, R. (1997). Africa’s growth tragedy: Policies and ethnic divisions. Quar-terly Journal of Economics, 112(4):1203–1250.
Ekeh, P. P. (1975). Colonialism and the two publics in Africa: A theoretical statement. Compara-tive Studies in Society and History, 17(01):91–112.
Englund, H. (2002). The village in the city, the city in the village: Migrants in Lilongwe. Journalof Southern African Studies, 28(1):137–154.
Ferguson, A. and Mulwafu, W. (2004). Decentralization, participation and access to water re-sources in Malawi. Madison, Wisconsin: BASIS Collaborative Research Support Program Re-port.
Ferree, K. and Horowitz, J. (2010). Ties that bind? The rise and decline of ethno-regional parti-sanship in Malawi, 1994–2009. Democratization, 17(3):534–563.
Franck, R. and Rainer, I. (2012). Does the leader’s ethnicity matter? Ethnic favoritism, educationand health in Sub-Saharan Africa. American Political Science Review, 106(2):294–325.
Gelman, A. and Hill, J. (2006). Data Analysis Using Regression and Multilevel/HierarchicalModels. Cambridge University Press, Cambridge, UK.
Gershman, B. and Rivera, D. (2016). Subnational diversity in sub-Saharan Africa: Insights from anew dataset. Paper presented at the Workshop on Ethnicity and Diversity, Carlos III-Juan MarchInstitute.
Gilman, L. (2009). The Dance of Politics: Gender, Perfromance, and Democratization in Malawi.Temple University Pres, Philadelphia, PA.
Global Rural-Urban Mapping Project (2011).
Golden, M. and Min, B. (2013). Distributive politics around the world. Annual Review of PoliticalScience, 16:73–99.
Government of Malawi (1998). Population and housing census. National Statistical Office, Zomba,Malawi.
31
Government of Malawi (2008). Population and housing census. National Statistical Office, Zomba,Malawi.
Habyarimana, J. P., Humphreys, M., Posner, D. N., and Weinstein, J. (2009). Coethnicity: Diversityand the Dilemmas of Collective Action. Russell Sage Foundation Publications, New York.
Harding, R. (2015). Attribution and accountability: Voting for roads in Ghana. World Politics,67(4):656–689.
Harding, R. and Stasavage, D. (2014). What democracy does (and doesn’t) do for basic services:School fees, school inputs, and african elections. Journal of Politics, 76(1):229–245.
Harrigan, J. (2008). Food insecurity, poverty and the Malawian Starter Pack: Fresh start or falsestart? Food Policy, 33(3):237–249.
Ichino, N. and Nathan, N. (2013). Local ethnic geography and instrumental ethnic voting: Voterbehavior across rural Ghana. American Political Science Review, 107(2):344–61.
Ishiyama, J. (2012). Explaining ethnic bloc voting in Africa. Democratization, 19(4):761–788.
Jablonski, R. (2014). How aid targets votes: The effect of electoral strategies on the distribution offoreign aid. World Politics, 66(2):293–330.
Joseph, R. (1987). Democracy and Prebendalism in Nigeria. Cambridge University Press, NewYork, NY.
Kasara, K. (2007). Tax me if you can: Ethnic geography, democracy, and the taxation of agriculturein Africa. American Political Science Review, 101(01):159–172.
Kasara, K. (2013). Separate and suspicious: Local social and political context and ethnic tolerancein Kenya. Journal of Politics, 75(04):921–936.
Kimenyi, M. S. (2006). Ethnicity, governance and the provision of public goods. Journal ofAfrican Economies, 15:62–99.
King, G., Tomz, M., and Wittenberg, J. (2000). Making the most of statistical analyses: Improvinginterpretation and presentation. American Journal of Political Science, 44(2):347–361.
Kishindo, P. (2004). Customary land tenure and the new land policy in Malawi. Journal of Con-temporary African Studies, 22(2):213–225.
Kramon, E. (2013). Vote Buying and Accountability in Democratic Africa. PhD Dissertation,University of California, Los Angeles.
Kramon, E. and Posner, D. N. (2013). Who benefits from distributive politics? How the outcomesone studies affect the answers one gets. Perspectives on Politics, 11(2).
Lindberg, S. I. (2003). It’s our time to “chop": Do elections in Africa feed neo-patrimonialismrather than counter-act it? Democratization, 10(2):121–40.
32
Lindberg, S. I. (2010). What accountability pressures do MPs in africa face and how do theyrespond? Evidence from Ghana. Journal of Modern African Studies, 48(01):117–142.
Magaloni, B., Diaz-Cayeros, A., and Estévez, F. (2007). Clientelism and portfolio diversification:A model of electoral investment with applications to Mexico. In Kitschelt, H. and Wilkinson,S. I., editors, Patrons, Clients, and Policies: Patterns of Democratic Accountability and PoliticalCompetition, pages 182–205. New York: Cambridge University Press.
Matuszeski, J. and Schneider, F. (2006). Patterns of ethnic group segregation and civil conflict.Working Paper, Harvard University.
Miguel, E. and Gugerty, M. K. (2005). Ethnic diversity, social sanctions, and public goods inKenya. Journal of Public Economics, 89(11-12):2325–2368.
Min, B. (2015). Power and the Vote: Elections and Electricity in the Developing World. NewYork: Cambridge University Press.
Minnesota Population Center (2014). Integrated Public Use Microdata Series, International: Ver-sion 6.3. Minneapolis: University of Minnesota.
Mthinda, C. and Khaila, S. (2006). Responsiveness and accountability in Malawi. Afrobarometer:Briefing Paper No. 31.
Nathan, N. (2016). Local ethnic geography, expectations of favoritism, and voting in urban Ghana.Comparative Political Studies, 49(14):1896–1929.
National Statistical Office of Malawi and ORC Macro (2001). Malawi Demographic and HealthSurvey 2000. Zomba, Malawi and Calverton, Maryland, USA: National Statistical Office andORC Macro.
Nichter, S. (2008). Vote buying or turnout buying? Machine politics and the secret ballot. AmericanPolitical Science Review, 102(01):19–31.
Nyirenda, M. (2014). Constituents demand recall of MCP MP: Pay tribute to Kamanya. The NyasaTimes, September 3.
O’Neil, T., Cammack, D., Kanyongolo, E., Mkandawire, M. W., Mwalyambwire, T., Welham, B.,and Wild, L. (2014). Fragmented governance and local service delivery in Malawi. London:Overseas Development Institute Report.
Øygard, R., Garcia, R., Guttormsen, A., Kachule, R., Mwanaumo, A., Mwanawina, I., Sjaastad,E., Wik, M., Bonnerjee, A., Braithwaite, J., Carvalho, S., Ezemenari, K., Graham, C., andThomson, A. (2003). The maze of maize: Improving input and output market access for poorsmallholders in the Southern African region. Department of Economics and Resource Manage-ment, Agricultural University of Norway Report No. 26.
Posner, D. N. (2004). Measuring ethnic fractionalization in Africa. American Journal of PoliticalScience, 48(4):849–863.
33
Posner, D. N. (2005). Institutions and Ethnic Politics in Africa. New York: Cambridge UniversityPress.
Potts, D. (2006). Rural mobility as a response to land shortages: The case of Malawi. Population,Space and Place, 12:291–311.
Reardon, S. F. and O’Sullivan, D. (2004). Measures of spatial segregation. Sociological Method-ology, 34(1):121–162.
Stokes, S., Dunning, T., Nazareno, M., and Brusco, V. (2013). Brokers, Voters, and Clientelism:The Puzzle of Distributive Politics. Cambridge University Press, Cambridge, UK.
Strandow, D., Findley, M., Nielson, D., and Powell, J. (2011). The UCDP-AidData codebookon geo-referencing foreign aid. Uppsala Conflict Data Program. Uppsala, Sweden: UppsalaUniversity.
Tajfel, H. and Turner, J. C. (1985). The social identity theory of intergroup behavior. In Worchel,S. and Austin, W. G., editors, Psychology of Intergroup Relations. Chicago, Ill.: Nelson Hall.
Tambulasi, R. I. (2009). The public sector corruption and organized crime nexus: The case of thefertiliser subsidy programme in Malawi. African Security Review, 18(4):19–31.
The Malawi Voice (2014). Dr. Jean Kalilani rating high in Dowa. November 14.
The Nation (2012). MP commissions K1.6m borehole in Zomba. July 11.
Tsoka, M. and Chinsinga, B. (2009). Afrobarometer survey Malawi round 4 summary results.Centre for Social Research: Chancellor College, Zomba.
Vail, L. and White, L. (1991). Tribalism in the political history of Malawi. In Vail, L., editor,The Creation of Tribalism in Southern Africa, pages 151–192. University of California Press,Berkeley.
van de Walle, N. (2003). Presidentialism and clientelism in Africa’s emerging party systems.Journal of Modern African Studies, 41(2):297–321.
Velasquez, M. (2013). From subject to citizen: Evidence of geographic variation in ethnic polit-ical mobilization. Paper presented at the Working Group in African Political Economy, Mas-sachusetts Institute of Technology.
Wantchekon, L. (2003). Clientelism and voting behavior: Evidence from a field experiment inBenin. World Politics, 55(3):399–422.
WaterAid (2008). Reaching out to the excluded exclusion study on water, sanitation and hygienedelivery in Malawi. WaterAid Report.
WaterAid (2010). Wateraid Malawi: Country strategy 2011-2015. WaiterAid Report.
34
Wooldridge, J. M. (1997). Quasi-likelihood methods for count data. In Pesaran, M. H. and Schmidt,P., editors, Handbook of Applied Econometrics, volume 2, pages 352–406. Blackwell, Oxford,UK.
35
Table 1: Segregation and Borehole Investments across Electoral Districts
Dependent variable:
Number of New Boreholes
(1) (2) (3) (4)
Segregation (continuous) 1.86∗∗ 1.76∗∗
(0.72) (0.76)Dummy for Medium Segregation 0.43∗∗∗ 0.50∗∗∗
(0.15) (0.16)Dummy for High Segregation 0.38∗∗ 0.38∗∗
(0.17) (0.19)Ethnic Diversity (ELF) −0.89∗∗ −0.63 −0.86∗∗ −0.45
(0.39) (0.61) (0.38) (0.60)Population Density (ln) 0.52∗∗∗ 0.41∗ 0.38∗∗ 0.25
(0.20) (0.21) (0.18) (0.20)Urban Proportion (ln) −0.05 −0.02 −0.04 −0.02
(0.04) (0.04) (0.04) (0.04)Land Area (square KM) (ln) 0.69∗∗∗ 0.63∗∗∗ 0.66∗∗∗ 0.56∗∗∗
(0.15) (0.15) (0.14) (0.15)Boreholes per 10,000 residents in 1998 0.24∗∗ 0.23∗∗ 0.24∗∗∗ 0.22∗∗
(0.09) (0.10) (0.09) (0.09)Electoral Competitiveness 0.005 0.01
(0.005) (0.005)MP Coethnic Population Share 0.02 0.16
(0.42) (0.42)President Coethnic Population Share 0.78 0.50
(1.14) (1.13)Distance to Nearest City (ln) 0.31 −0.03
(0.83) (0.82)Water Aid Projects per 10,000 residents 0.31 0.07
(0.47) (0.49)All Aid Projects per 10,000 residents −0.07 −0.05
(0.05) (0.05)Constant −0.62 −2.25 −0.19 0.05
(0.82) (4.84) (0.80) (4.82)
Admin. District Fixed Effects X X X XObservations 183 182 183 182
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
36
Table 2: Segregation and Ethnic Favoritism in the Provision of Boreholes
Low Med. High All All All All(1) (2) (3) (4) (5) (6) (7)
A. Match with MP: Largest Ethnic Group in Locality
Match with MP 0.03 0.16∗∗∗ 0.21∗∗∗ 0.05 0.03 0.05∗ −0.02(0.04) (0.03) (0.06) (0.03) (0.03) (0.03) (0.07)
Match x Med. Seg. 0.08∗∗ 0.09∗∗ 0.07∗
(0.04) (0.04) (0.04)Match x High Seg. 0.18∗∗∗ 0.18∗∗∗ 0.15∗∗∗
(0.06) (0.06) (0.06)Match x Continuous Seg. 0.29∗
(0.16)
B. Match with MP: Proportion Coethnic
Match with MP 0.13∗∗ 0.20∗∗∗ 0.57∗∗∗ 0.12∗∗ 0.09∗ 0.12∗∗ −0.20(0.06) (0.05) (0.15) (0.05) (0.05) (0.05) (0.19)
Match x Med. Seg. 0.06 0.09 0.10(0.06) (0.06) (0.06)
Match x High Seg. 0.49∗∗∗ 0.48∗∗∗ 0.47∗∗∗
(0.14) (0.14) (0.14)Match x Continuous Seg. 0.96∗∗
(0.47)
Locality Fixed Effects X X X X X X XTime Period Fixed Effects X X X X X X XTime Varying Controls X X XFixed Controls x Time Period X XNumber of Electoral Districts 42 43 35 120 120 120 120Number of Localities 1315 1542 645 3502 3502 3502 3502Number of Observations 2630 3084 1290 7004 7004 7004 7004
Notes: The table shows difference-in-differences estimates of a locality-MP ethnic match, for different levels ofsegregation. The dependent variable is an indicator for whether a locality has a borehole. Columns (1)-(3) include asubset of electoral districts based on their segregation levels (low, medium, and high, respectively), while (4)-(7)include electoral districts across all levels of segregation. Panel A uses the locality’s largest ethnic group to defineMP ethnic match (Match), while Panel B uses the MP’s share of coethnics (Match Proportion). Models with timevarying controls include controls for ethnic match with the president and the presence of a health clinic or school inthe locality. Models with fixed controls interacted with the time period dummy include time period interactions withan indicator of whether the locality is urban, population density (logged), ethno-linguistic fractionalization, area(square KM), distance to Lilongwe, distance to Blantyre, and the number of boreholes per capita in 1998. The tableshows that ethnic favoritism is more prevalent when the MP’s electoral district is more ethnically segregated. ∗p<0.1;∗∗p<0.05; ∗∗∗p<0.01.
37
Figure 1: Two Hypothetical Electoral Districts with Different Levels of Segregation
●●
●●
●
●
●
●● ●
●●
●
●●
●
● ●
●
●
●●●
●
●● ●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●
●● ●
●
●●●
●●
●●
● ●
●
●
●
●●
●
●●●
●
●
●
●
●
●●
●●● ●●●
●
●●
●
●
●
●●
●
●
●
●●
●●
●● ●
● ●●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
● ●
●
●
●
●
●●
●
●
●
● ●
●●
●●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
● ●●●
●
●
● ●
●●
●
●
●
●●
●●
●
●
●
●●
●● ●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●●
● ●●
●
●●●
●
●
●
●
●●● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●●
● ●
●●
●
●
●
●
●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●●
●
●●
● ●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●●
●
●
●
●●
● ●●●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
● ● ●
●
●
●
●
●●
●●
●
●●●
●
●●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
● ●
●
●●
●●
●
●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●
●
●●
●
●
●
● ●
●●
● ● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●●●
●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
● ●
●
●●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
● ●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●● ●
●
● ●●
●●
●
●
●
●●
●
●●
●
●●
●●
●
● ●
●●
●
●
●
● ●●
●
●
●
●
●
●●
●
●●
●
●
●●
●●
●
●
●
●●
●
●●
●●
● ●●
●
●
●
●●
●
●●
●
●●
●●
●
●●
●●●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●●
●●
●
●●
●
●
●
● ●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●●
●●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●●
● ●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●●
●●
● ●
●
●●●
●
●
●
●
●
(a)
● ●
●
●
●●● ●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●●●
●
●
●
●
● ●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●●●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●
●● ●●
●
●
●●
●
●
●●
●
●
●●
● ●
●●
●
●
●
●
● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
● ●
●
● ●
●
●
●
●
●
●●●
●
●
●
●●
●
● ●
●
●
●
●
●●
●
●●
●● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
● ●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
● ●
● ●
●●
●
●
●
●●
●
●●
●
● ●
●
●
●
●
●●
●●●
●
●
●
●● ●
●
●
● ●●●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
● ●
●●
●
● ●
●
●
●
●
●
●
●●●
●●●
●
●●
●
●
●
●●
●●●
●●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
● ●
●
●●
●
● ●
●●●
●●
●
●
●
●●
●
●
●●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
● ●
● ●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●●
●● ●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●
●
●
●● ●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●●●
●●●
● ●
●
●
●
●●
●
●
●● ●
●
●
●●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●●
●
●
● ●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●● ●
● ●●
●
●
●
●
●
● ●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●
●
(b)
Note: Assume that (a) and (b) represent two electoral districts, each of which is divided into 16 localities. Eachlocality is populated by a politician’s coethnics (gray squares) or non-coethnics (black dots). The diamond shows thelocation of a local public good, and the transparent circle represents its catchment area. Ethnic diversity, populationsize, and population density — three predictors of local public goods investment — are held constant across theelectoral districts. Coethnic localities in both districts have the same proportion of coethnics. The figure illustratesthat more coethnics benefit from the local public good in (a) — the more segregated district — than in (b). This figurerelates to Figure 1 in Ichino and Nathan (2013), which shows how local ethnic geography influences voters. Therelationship of these two figures highlights complementarities between our arguments.
38
Figure 2: Ethnic Segregation in Two Electoral Districts
Note: This figure provides an example of two electoral districts with similar levels of diversity but differentsegregation scores. The spatial dissimilarity score for the MP’s ethnic group is 0.70 in Phalombe North and 0.21 inMachinga South. Each dot represents one individual (shaded according to ethnic match with the MP).
39
Figure 3: Ethnic Favoritism Is More likely in Segregated Electoral Districts
●
●
0.0
0.1
0.2
0.3
0.4
0.5
Low Segregation
Pr(
Loca
lity
Has
Bor
ehol
e)
1998 2008
●
Coethnic after 1998Never Coethnic
●
●
0.0
0.1
0.2
0.3
0.4
0.5
Medium Segregation
1998 2008
●
●
0.0
0.1
0.2
0.3
0.4
0.5
High Segregation
1998 2008
Note: Analysis includes 3502 localities nested in 120 electoral districts. All of these localities were not coethnic withtheir MP in 1998. 1599 became coethnic with their MP in either the 1999 or the 2004 parliamentary elections; theseare denoted with a triangle. The 1903 localities denoted with a circle were never coethnic with their MP in the studyperiod.
40
Supporting Information:Segregation, Ethnic Favoritism, and the Strategic Targeting ofLocal Public Goods
Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Ethnicity Data for Members of Parliament . . . . . . . . . . . . . . . . . . . . . . . . . 10
Measure of Ethnic Group Segregation . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Electoral District Raw Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Complete DiD Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Robustness Tests: Including All Electoral Districts . . . . . . . . . . . . . . . . . . . . 17
Robustness Tests: Including Rural Electoral Districts Only . . . . . . . . . . . . . . . . 20
Robustness Tests: Excluding Districts in Machinga and Mangochi . . . . . . . . . . . . 23
DiD Robustness Test: Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
DiD Robustness Test: Clustering by Electoral District-Year . . . . . . . . . . . . . . . . 28
DiD Robustness Test: Randomly Generated Segregation Cutoffs . . . . . . . . . . . . . 30
Cross-Sectional Ethnic Favoritism Analyses . . . . . . . . . . . . . . . . . . . . . . . . 32
Other Public Goods: Health Clinics and Schools . . . . . . . . . . . . . . . . . . . . . . 34
Placebo Test: Segregation and the Provision of Private Goods . . . . . . . . . . . . . . . 42
1
Descriptive Statistics
Figure SI.1: The Spatial Distribution of Malawi’s Major Ethnic Groups
Note: Each dot in the figure represents 100 individuals, and has been color coded according to ethnicity. The grayborders delineate Malawi’s 193 electoral districts.
2
Figu
reSI
.2:N
ewPu
blic
Goo
ds,1
998-
2008
3
Figure SI.3: Locality-MP Ethnic Match
4
Figure SI.4: Ethnic Segregation of Electoral Districts
5
Figure SI.5: Ethnic Segregation Categories
Note: The figure shows three electoral districts with low (left), medium (middle), and high (right) segregation scores.Each dot represents one individual, shaded according to ethnic match with the MP (black = coethnic; gray =non-coethnic). The segregation scores in the three electoral districts are 0.21, 0.43, and 0.70, respectively.
6
Table SI.1: Summary Statistics Across Electoral Districts
Mean SD Min Max N
Demographics
MP Ethnic Group Segregation 0.43 0.13 0.04 0.80 193
Ethnic Diversity (ELF) 0.41 0.24 0.01 0.84 193
Population Density (in 10,000s/sqkm) 0.41 1.00 0.01 9.10 193
Urban Proportion 0.09 0.25 0.00 1.00 193
Geography
Land Area, Sq Km 389.08 256.20 7.23 1203.07 193
Distance to Nearest City 110.91 109.88 0.00 496.45 193
Representation
MP Coethnic Population Share 0.61 0.31 0.01 0.99 191
President Coethnic Population Share 0.15 0.17 0.00 0.49 193
Electoral Competitiveness (Percentage Point Margin of Victor) 34.44 17.92 1.79 85.29 193
Public Goods
Boreholes per 10,000 residents in 1998 0.57 0.82 0.00 5.20 193
New Borehole Indicator, 1998-2008 0.82 0.38 0.00 1.00 193
No. of New Boreholes, 1998-2008 39.00 40.64 0.00 207.00 193
Schools per 10,000 residents in 1998 3.34 2.16 0.00 13.30 193
New School Indicator, 1998-2008 0.69 0.46 0.00 1.00 193
No. of New Schools, 1998-2008 5.64 9.03 0.00 68.00 193
Clinics per 10,000 residents in 1998 0.57 0.60 0.00 3.95 193
New Clinic Indicator, 1998-2008 0.44 0.50 0.00 1.00 193
No. of New Clinics, 1998-2008 1.16 1.93 0.00 15.00 193
All Aid Projects per 10,000 residents 1.52 2.65 0.00 22.75 193
Water Aid Projects per 10,000 residents 0.08 0.28 0.00 2.76 193
Education Aid Projects per 10,000 residents 0.09 0.23 0.00 1.39 193
Health Aid Projects per 10,000 residents 0.16 0.29 0.00 1.84 193
Private Goods
Proportion Receiving Fertilizer Subsidy, 2004 0.60 0.18 0.05 1.00 159
7
Table SI.2: Summary Statistics Across Localities
Mean SD Min Max N
Demographics
Ethnic Diversity (ELF) 0.32 0.26 0.00 0.87 12380
Ethnic Group Majority Present 0.80 0.40 0.00 1.00 12380
Proportion of Largest Ethnic Group 0.74 0.24 0.00 1.00 12380
Population (in 1,000s) 1.04 0.55 0.00 8.29 12380
Population Density (in 1,000s/sqkm) 1.12 4.14 0.00 90.95 12380
Urban 0.14 0.35 0.00 1.00 12380
Geography
Land Area, Sq Km 6.07 7.51 0.02 275.30 12380
Distance to Nearest City 9873.90 3881.42 84.09 19788.67 12380
Representation
MP Ethnic Match Ever, 1999-2009 0.76 0.43 0.00 1.00 11983
MP Ethnic Match Share Avg. 0.59 0.36 0.00 1.00 12292
Public Goods
No. of Boreholes, 1998 0.05 0.32 0.00 6.00 12380
New Borehole Indicator, 1998-2008 0.33 0.47 0.00 1.00 12380
No. of New Boreholes, 1998-2008 0.61 1.10 0.00 13.00 12380
No. of Schools, 1998 0.33 0.59 0.00 6.00 12380
New School Indicator, 1998-2008 0.08 0.27 0.00 1.00 12380
No. of New Schools, 1998-2008 0.09 0.35 0.00 4.00 12380
No. of Clinics, 1998 0.05 0.24 0.00 4.00 12380
New Clinic Indicator, 1998-2008 0.02 0.13 0.00 1.00 12380
No. of New Clinics, 1998-2008 0.02 0.15 0.00 4.00 12380
Private Goods
Proportion Receiving Fertilizer Subsidy, 2004 0.60 0.23 0.00 1.00 469
8
Figure SI.6: Relationship Between Diversity and Segregation Across Electoral Districts
0
.2
.4
.6
.8
Ethn
ic S
egre
gatio
n
0 .2 .4 .6 .8Ethnic Diversity
Urban ConstituenciesRural Constituencies
Note: This figure shows the relationship between an electoral district’s degree of ethnic diversity, measured using thestandard ethnolinguistic fractionalization index, and the degree of ethnic group segregation, measured using theaverage MP-specific ethnic group spatial dissimilarity index. The correlation coefficient is -0.43 across all 193districts, but only -0.28 among rural districts. This negative relationship is driven by the fact that segregation isincreasingly difficult as diversity increases. Despite the negative correlation, there is considerable variation insegregation at all levels of diversity.
9
Ethnicity Data for Members of Parliament
We compiled new data on the ethnic identity of each Malawian MP who served between 1994
and 2009. To assemble this dataset, we first gathered the names of all MPs from official election
returns (Government of Malawi, 1994, 1999, 2004). Two Malawian research assistants then coded
the ethnic identity of each MP with the assistance of staff at the Malawi Electoral Commission,
Administrative District Offices, and local elites within each electoral district. The inter-rater relia-
bility score across the two coders was 0.66. Where codings differ, we use the coding with the best
documented sources.
10
Measure of Ethnic Group Segregation
Our fine-grained ethnicity data enable us to improve upon past efforts to examine the impact of
segregation on ethnic favoritism. In particular, we improve upon the analysis in Franck and Rainer
(2012), which finds no evidence that ethnic favoritism by African presidents is conditioned by
country-level segregation. Franck and Rainer’s measure of segregation comes from Matuszeski and
Schneider (2006), who developed it from the spatial distribution of language groups provided in
the Global Mapping International (GMI) dataset. GMI partitions the globe into mutually exclusive
language-group polygons such that each area of the world has only one language group whose
boundaries are defined and non-overlapping. Thus, the data cannot capture ethnic integration that
occurs from members of more than one group living in the same local area. As our data show,
however, there exists a great deal of local ethnic heterogeneity. Additionally, because Matuszeski
and Schneider measure the segregation of language groups relative to a spatial grid within each
country, levels of segregation on this measure are driven almost entirely by the number of ethnic
group borders in a country: where there are more borders — because of more groups or because
large groups reside in segregated pockets — the country is scored as less segregated. As a result,
the measure is likely to generate misleading codings of segregation. Our disaggregated data thus
allow for a more appropriate test of segregation’s impact.
Using these fine-grained data, we rely on the spatial dissimilarity index to measure how
geographically clustered different ethnic groups are relative to what an even geographic distribution
of the ethnic groups would look like. This and its non-spatial counterpart are widely used and
accepted measures of ethnic and racial segregation (e.g., Cutler et al., 2012). The non-spatial
version of the dissimilarity index captures the deviation between locality and electoral district
ethnic group proportions. In the case of complete integration, all localities would have the same
ethnic group proportions as the whole district. The spatial version is similar but also accounts
for the ethnic make-up of neighboring localities. It measures the deviation between the ethnic
composition of the district and individuals’ “local environment,” where the local environment can
11
consist of (parts of) several neighboring localities. We implement this measure in R, using the
function spseg in package seg.
In this section, we describe the theory behind the spatial dissimilarity measure in fur-
ther detail.1 We are interested in measuring the spatial distribution of two mutually exclusive
groups: coethnics of the MP and non-coethnics of the MP. Index these two groups g ∈ {c,nc}
(c = coethnics; nc = non-coethnics). Further, let p index geographical locations within the MP’s
jurisdiction, which is denoted J, and let q index points located some distance from p. Param-
eters super-positioned with ˜ describe the local spatial environment of a point rather than the
point itself. Table SI.3 describes each component of the spatial dissimilarity measure. In the
table, “population density” means the population count per unit area (e.g., 10 m2) at location p;
φp =∫
q∈J exp(−2||p−q||)dq; and ||p−q|| represents the euclidean distance between p and q. The
spatial dissimilarity index D̃ is then:
D̃ = ∑g
∫p∈J
τp
2NI|πg
p−πg|d p
Note that this measure will approach 0, indicating minimum segregation, when the group
proportions at the local environment (πgp) are similar to the overall ethnic composition of the juris-
diction (πg). Further, exp(−2||p− q||) is one of many potential non-negative functions we could
have chosen to define the proximity of p and q. In our case, the farther away p and q are, the less
will q influence the local environment at p. This is the default option in seg, the R package we use
to implement this measure.
1See Reardon and O’Sullivan (2004) for a detailed discussion of the nature and validity of this and
other spatial segregation measures.
12
Table SI.3: Key Components of the Spatial Dissimilarity Index
Symbol Concept Equivalent expression
N Total population in jurisdiction Jτp Population density at pτ
gp Population density of g at p
τ̃p Population density in local environment1φp
∫q∈J
τq exp(−2||p−q||)dq
τ̃gp Population density of g in local environment
1φp
∫q∈J
τgq exp(−2||p−q||)dq
πg Proportion of g of total population
π̃gp Proportion of g in local environment
τ̃gp
τ̃pI Overall diversity of J 2πcπnc
13
Electoral District Raw Data
Figure SI.7 plots the number of boreholes built in 1998-2008 against ethnic segregation across
Malawi’s electoral districts. Consistent with our hypothesis, new borehole investments and ethnic
segregation are positively correlated.2
Figure SI.7: Segregated Electoral Districts Invested in More Boreholes than Integrated Districtsin 1998-2008
●
●
●
●
● ●●●
●
● ●●
●
●
●
●
●
● ●
● ●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●●
●
●
●
● ●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●●
●●● ●●
●
●●
●● ●
●
●
● ●
●●
●
●
●●
●●
●
●●
●
●
●
●●● ●● ●
●
●
●
●
●
●
●
●●
●●
●
●● ●● ●
●● ●
●
●
●
●
●
●● ●
●●
●● ●
●
0
50
100
150
200
0.2 0.4 0.6 0.8
Segregation
Num
ber
of N
ew B
oreh
oles
Population(1,000s)
●
●
●
●
50
100
150
200
Note: This figure shows the relationship between segregation and borehole investments across 183 electoral districtsin Malawi. Ten districts are not shown because they are very homogeneous (ELF scores below 0.05). Point size isproportional to districts’ population size. The lines are population-weighted loess smoothers; the solid line uses all183 observations while the dashed line excludes 8 districts with very high (2 standard deviations above the mean)borehole investments.
2The three districts with the highest levels of segregation have extremely small populations and did
receive new boreholes, pulling the loess curve down at very high levels of segregation.
14
Complete DiD Results
Table SI.4 presents the complete DiD results. The Table corresponds to Table 2, displaying the
coefficients on each of the control variables.
15
Table SI.4: Segregation and Ethnic Favoritism in the Provision of Boreholes
Dependent variable:
Indicator for whether locality has a boreholeLow Med. High All All All All
(1) (2) (3) (4) (5) (6) (7)
Match with MP 0.03 0.16∗∗∗ 0.21∗∗∗ 0.05 0.03 0.05∗ −0.02(0.04) (0.03) (0.06) (0.03) (0.03) (0.03) (0.07)
Match x Med. Seg. 0.08∗∗ 0.09∗∗ 0.07∗
(0.04) (0.04) (0.04)Match x High Seg. 0.18∗∗∗ 0.18∗∗∗ 0.15∗∗∗
(0.06) (0.06) (0.06)Match x Continuous Seg. 0.29∗
(0.16)Presence of Clinic 0.04 0.04 0.05
(0.07) (0.07) (0.07)Presence of School 0.19∗∗∗ 0.20∗∗∗ 0.20∗∗∗
(0.03) (0.03) (0.03)Presidential Ethnic Match 0.09∗∗∗ 0.10∗∗∗ 0.10∗∗∗
(0.02) (0.03) (0.03)Time 0.28∗∗∗ 0.23∗∗∗ 0.28∗∗∗ 0.26∗∗∗ 0.23∗∗∗ 0.22∗∗∗ 0.22∗∗∗
(0.02) (0.02) (0.03) (0.01) (0.02) (0.05) (0.05)Time x Urban −0.21∗∗∗ −0.21∗∗∗
(0.04) (0.04)Time x Pop. Density (log) −0.01 −0.01
(0.01) (0.01)Time x ELF 0.13∗∗ 0.13∗∗
(0.05) (0.05)Time x Land Area (sqkm) −0.01∗∗ −0.005∗∗
(0.00) (0.00)Time x Distance to Lilongwe −0.00 −0.00
(0.00) (0.00)Time x Distance to Blantyre 0.00 0.00
(0.00) (0.00)Time x Boreholes per Capita, −0.14∗∗∗ −0.15∗∗∗
1999 (0.04) (0.03)
Locality Fixed Effects X X X X X X XTime Period Fixed Effects X X X X X X XTime Varying Controls X X XFixed Controls x Time Period X XNumber of Electoral Districts 42 43 35 120 120 120 120Number of Localities 1315 1542 645 3502 3502 3502 3502Number of Observations 2630 3084 1290 7004 7004 7004 7004
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
16
Robustness Tests: Including All Electoral Districts
In this section we show that our analyses of H1 and H2 are robust to including all of Malawi’s
193 electoral districts. Recall that in the main analysis, we exclude highly homogenous electoral
districts — those with ELF scores of less than 0.05 — because measures of segregation does not
produce meaningful estimates without a minimum level of diversity. Tables SI.5 and SI.6 show,
however, that our results are not sensitive to including all electoral districts in the analysis.
17
Table SI.5: Segregation and Borehole Investments across Electoral Districts (Including AllElectoral Districts)
Dependent variable:
Number of New Boreholes
(1) (2) (3) (4)
Segregation (continuous) 1.32∗∗ 1.15(0.66) (0.69)
Dummy for Medium Segregation 0.38∗∗∗ 0.42∗∗∗
(0.14) (0.16)Dummy for High Segregation 0.28∗ 0.26
(0.16) (0.17)Ethnic Diversity (ELF) −0.88∗∗ −0.80 −0.84∗∗ −0.57
(0.37) (0.60) (0.37) (0.59)Population Density (ln) 0.51∗∗∗ 0.43∗∗ 0.43∗∗ 0.36∗
(0.18) (0.20) (0.17) (0.19)Urban Proportion (ln) −0.06 −0.04 −0.06 −0.04
(0.04) (0.04) (0.04) (0.04)Land Area (square KM) (ln) 0.69∗∗∗ 0.66∗∗∗ 0.68∗∗∗ 0.63∗∗∗
(0.14) (0.14) (0.13) (0.14)Boreholes per 10,000 residents in 1998 0.27∗∗∗ 0.26∗∗∗ 0.28∗∗∗ 0.26∗∗∗
(0.09) (0.09) (0.08) (0.09)Electoral Competitiveness 0.005 0.01
(0.004) (0.004)MP Coethnic Population Share −0.12 0.04
(0.41) (0.41)President Coethnic Population Share 0.54 0.30
(1.12) (1.11)Distance to Nearest City (ln) 0.02 −0.17
(0.80) (0.80)Water Aid Projects per 10,000 residents 0.31 0.11
(0.46) (0.47)All Aid Projects per 10,000 residents −0.07 −0.05
(0.05) (0.05)Constant −0.54 −0.40 −0.32 0.72
(0.77) (4.63) (0.76) (4.63)
Admin. District Fixed Effects X X X XObservations 193 191 193 191
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
18
Table SI.6: Segregation and Ethnic Favoritism in the Provision of Boreholes (Including AllElectoral Districts)
Low Med. High All All All All(1) (2) (3) (4) (5) (6) (7)
A. Match with MP: Largest Ethnic Group in Locality
Match with MP 0.03 0.16∗∗∗ 0.21∗∗∗ 0.05 0.03 0.05∗ −0.02(0.04) (0.03) (0.06) (0.03) (0.03) (0.03) (0.07)
Match x Med. Seg. 0.08∗∗ 0.09∗∗ 0.07∗
(0.04) (0.04) (0.04)Match x High Seg. 0.18∗∗∗ 0.18∗∗∗ 0.15∗∗∗
(0.06) (0.06) (0.06)Match x Continuous Seg. 0.29∗
(0.16)
B. Match with MP: Proportion Coethnic
Match with MP 0.13∗∗ 0.20∗∗∗ 0.57∗∗∗ 0.12∗∗ 0.09∗ 0.12∗∗ −0.20(0.06) (0.05) (0.15) (0.05) (0.05) (0.05) (0.19)
Match x Med. Seg. 0.06 0.09 0.10(0.06) (0.06) (0.06)
Match x High Seg. 0.49∗∗∗ 0.48∗∗∗ 0.47∗∗∗
(0.14) (0.14) (0.14)Match x Continuous Seg. 0.96∗∗
(0.47)
Locality Fixed Effects X X X X X X XTime Period Fixed Effects X X X X X X XTime Varying Controls X X XFixed Controls x Time Period X XNumber of Electoral Districts 42 43 35 120 120 120 120Number of Localities 1315 1542 645 3502 3502 3502 3502Number of Observations 2630 3084 1290 7004 7004 7004 7004
Notes: The table shows difference-in-differences estimates of a locality-MP ethnic match, for different levels ofsegregation. The dependent variable is an indicator for whether a locality has a borehole. Columns (1)-(3) include asubset of electoral districts based on their segregation levels (low, medium, and high, respectively), while (4)-(7)include electoral districts across all levels of segregation. Panel A uses the locality’s largest ethnic group to defineMP ethnic match (Match), while Panel B uses the MP’s share of coethnics (Match Proportion). Models with timevarying controls include controls for ethnic match with the president and the presence of a health clinic or school inthe locality. Models with fixed controls interacted with the time period dummy include time period interactions withan indicator of whether the locality is urban, population density (logged), ethno-linguistic fractionalization, area(square KM), distance to Lilongwe, distance to Blantyre, and the number of boreholes per capita in 1998. The tableshows that ethnic favoritism is more prevalent when the MP’s electoral district is more ethnically segregated. ∗p<0.1;∗∗p<0.05; ∗∗∗p<0.01.
19
Robustness Tests: Including Rural Electoral Districts Only
In this section we show that our analyses of H1 and H2 are robust to dropping urban electoral dis-
tricts. Doing so allows us to more effectively control for the demand for boreholes, as the demand
for clean water is much lower in urban areas where there is greater access. To this end we drop
all electoral districts with an urban population of over 30 percent from the sample (although the
results are robust to other cutoffs as well). Using this cutoff, the following 14 electoral districts are
dropped from the sample: Blantyre Bangwe, Blantyre City Central, Blantyre City East, Blantyre
City South, Blantyre City South East, Blantyre City West, Blantyre Kabula, Blantyre Malabada,
Lilongwe City Central, Lilongwe City South East, Lilongwe City South West, Lilongwe City West,
Mzimba Mzuzu City, and Zomba Central. Tables SI.7 and SI.8 show that our results are robust to
the removal of these electoral districts.
20
Table SI.7: Segregation and Borehole Investments across Electoral Districts (Rural ElectoralDistricts Only)
Dependent variable:
Number of New Boreholes
(1) (2) (3) (4)
Segregation (continuous) 0.87 0.74(0.77) (0.81)
Dummy for Medium Segregation 0.36∗∗ 0.38∗∗
(0.14) (0.16)Dummy for High Segregation 0.26 0.26
(0.17) (0.19)Ethnic Diversity (ELF) −0.58 −0.79 −0.57 −0.61
(0.40) (0.61) (0.39) (0.59)Population Density (ln) 1.07∗∗∗ 0.94∗∗∗ 0.99∗∗∗ 0.81∗∗∗
(0.25) (0.28) (0.24) (0.27)Urban Proportion (ln) −0.08∗∗ −0.07∗ −0.08∗∗ −0.06
(0.04) (0.04) (0.04) (0.04)Land Area (square KM) (ln) 0.71∗∗∗ 0.68∗∗∗ 0.65∗∗∗ 0.59∗∗∗
(0.15) (0.15) (0.15) (0.15)Boreholes per 10,000 residents in 1998 0.11 0.10 0.10 0.09
(0.10) (0.10) (0.10) (0.10)Electoral Competitiveness 0.001 0.003
(0.005) (0.005)MP Coethnic Population Share −0.37 −0.19
(0.43) (0.42)President Coethnic Population Share 1.17 0.95
(1.18) (1.18)Distance to Nearest City (ln) 0.25 −0.05
(0.83) (0.82)Water Aid Projects per 10,000 residents 0.22 0.02
(0.48) (0.49)All Aid Projects per 10,000 residents −0.05 −0.04
(0.05) (0.05)Constant 0.55 −0.45 0.91 1.41
(0.86) (4.88) (0.83) (4.82)
Admin. District Fixed Effects X X X XObservations 169 169 169 169
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
21
Table SI.8: Segregation and Ethnic Favoritism in the Provision of Boreholes (Rural ElectoralDistricts Only)
Low Med. High All All All All(1) (2) (3) (4) (5) (6) (7)
A. Match with MP: Largest Ethnic Group in Locality
Match with MP −0.06 0.16∗∗∗ 0.21∗∗∗ 0.05 0.02 0.04 0.05(0.04) (0.03) (0.06) (0.03) (0.03) (0.03) (0.12)
Match x Med. Seg. 0.06 0.07∗ 0.07∗
(0.04) (0.04) (0.04)Match x High Seg. 0.15∗∗∗ 0.15∗∗∗ 0.13∗∗
(0.06) (0.06) (0.06)Match x Continuous Seg. 0.09
(0.28)
B. Match with MP: Proportion Coethnic
Match with MP 0.02 0.20∗∗∗ 0.57∗∗∗ 0.11∗∗ 0.05 0.12∗∗ −0.06(0.06) (0.05) (0.15) (0.05) (0.05) (0.05) (0.24)
Match x Med. Seg. 0.05 0.09 0.10(0.06) (0.06) (0.06)
Match x High Seg. 0.45∗∗∗ 0.43∗∗∗ 0.44∗∗∗
(0.14) (0.14) (0.14)Match x Continuous Seg. 0.59
(0.58)
Locality Fixed Effects X X X X X X XTime Period Fixed Effects X X X X X X XTime Varying Controls X X XFixed Controls x Time Period X XNumber of Electoral Districts 34 43 35 112 112 112 112Number of Localities 987 1542 645 3174 3174 3174 3174Number of Observations 1974 3084 1290 6348 6348 6348 6348
Notes: The table shows difference-in-differences estimates of a locality-MP ethnic match, for different levels ofsegregation, including only rural electoral districts. The dependent variable is an indicator for whether a locality has aborehole. Columns (1)-(3) include a subset of electoral districts based on their segregation levels (low, medium, andhigh, respectively), while (4)-(7) include electoral districts across all levels of segregation. Panel A uses the locality’slargest ethnic group to define MP ethnic match (Match), while Panel B uses the MP’s share of coethnics (MatchProportion). Models with time varying controls include controls for ethnic match with the president and the presenceof a health clinic or school in the locality. Models with fixed controls interacted with the time period dummy includetime period interactions with an indicator of whether the locality is urban, population density (logged),ethno-linguistic fractionalization, area (square KM), distance to Lilongwe, distance to Blantyre, and the number ofboreholes per capita in 1998. The table shows that ethnic favoritism is more prevalent when the MP’s electoraldistrict is more ethnically segregated. ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01.
22
Robustness Tests: Excluding Districts in Machinga and Mangochi
In this section we show that our analyses of H1 and H2 are robust to dropping electoral districts in
the Machinga and Mangochi regions. We do so because Machinga and Mangochi were affected by
a relatively large rural resettlement program that the government of Malawi established in 2004.
The program resettled households from Thyolo and Mulanje districts to Machinga and Mangochi,
potentially altering ethnic demographics in the receiving districts. See Chinsinga (2011) for details.
Tables SI.9 and SI.10 show that our results are robust to the removal of these electoral districts.
23
Table SI.9: Segregation and Borehole Investments across Electoral Districts (Excluding ElectoralDistricts in Machinga and Mangochi)
Dependent variable:
Number of New Boreholes
(1) (2) (3) (4)
Segregation (continuous) 1.91∗∗ 1.75∗∗
(0.78) (0.84)Dummy for Medium Segregation 0.39∗∗ 0.45∗∗
(0.16) (0.19)Dummy for High Segregation 0.45∗∗ 0.49∗∗
(0.19) (0.21)Ethnic Diversity (ELF) −0.75 −0.29 −0.74 −0.10
(0.46) (0.68) (0.46) (0.66)Population Density (ln) 0.51∗∗ 0.32 0.39∗ 0.19
(0.21) (0.23) (0.20) (0.22)Urban Proportion (ln) −0.04 −0.01 −0.03 0.001
(0.04) (0.04) (0.04) (0.04)Land Area (square KM) (ln) 0.75∗∗∗ 0.66∗∗∗ 0.74∗∗∗ 0.61∗∗∗
(0.16) (0.16) (0.16) (0.16)Boreholes per 10,000 residents in 1998 0.24∗∗ 0.23∗∗ 0.24∗∗ 0.24∗∗
(0.10) (0.11) (0.10) (0.10)Electoral Competitiveness 0.01∗ 0.01∗∗
(0.01) (0.01)MP Coethnic Population Share 0.05 0.22
(0.45) (0.44)President Coethnic Population Share 1.02 0.99
(1.27) (1.32)Distance to Nearest City (ln) 0.73 0.29
(0.99) (0.98)Water Aid Projects per 10,000 residents 0.56 0.34
(0.52) (0.54)All Aid Projects per 10,000 residents −0.12∗ −0.09
(0.06) (0.06)Constant −1.10 −5.08 −0.68 −2.30
(0.89) (5.74) (0.88) (5.69)
Admin. District Fixed Effects X X X XObservations 164 163 164 163
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
24
Table SI.10: Segregation and Ethnic Favoritism in the Provision of Boreholes (ExcludingElectoral Districts in Machinga and Mangochi)
Low Med. High All All All All(1) (2) (3) (4) (5) (6) (7)
A. Match with MP: Largest Ethnic Group in Locality
Match with MP 0.03 0.09∗∗ 0.22∗∗∗ 0.04 0.03 0.04 −0.02(0.04) (0.04) (0.06) (0.03) (0.03) (0.03) (0.07)
Match x Med. Seg. 0.02 0.03 0.005(0.04) (0.04) (0.04)
Match x High Seg. 0.21∗∗∗ 0.21∗∗∗ 0.18∗∗∗
(0.06) (0.06) (0.06)Match x Continuous Seg. 0.23
(0.16)
B. Match with MP: Proportion Coethnic
Match with MP 0.14∗∗ 0.12∗∗ 0.55∗∗∗ 0.13∗∗ 0.09∗ 0.11∗∗ −0.13(0.06) (0.06) (0.16) (0.05) (0.05) (0.05) (0.20)
Match x Med. Seg. −0.03 −0.01 −0.01(0.07) (0.07) (0.06)
Match x High Seg. 0.52∗∗∗ 0.51∗∗∗ 0.50∗∗∗
(0.14) (0.14) (0.14)Match x Continuous Seg. 0.65
(0.48)
Locality Fixed Effects X X X X X X XTime Period Fixed Effects X X X X X X XTime Varying Controls X X XFixed Controls x Time Period X XNumber of Electoral Districts 38 37 30 105 105 105 105Number of Localities 1244 1274 603 3121 3121 3121 3121Number of Observations 2488 2548 1206 6242 6242 6242 6242
Notes: The table shows difference-in-differences estimates of a locality-MP ethnic match, for different levels ofsegregation. The dependent variable is an indicator for whether a locality has a borehole. Columns (1)-(3) include asubset of electoral districts based on their segregation levels (low, medium, and high, respectively), while (4)-(7)include electoral districts across all levels of segregation. Panel A uses the locality’s largest ethnic group to defineMP ethnic match (Match), while Panel B uses the MP’s share of coethnics (Match Proportion). Models with timevarying controls include controls for ethnic match with the president and the presence of a health clinic or school inthe locality. Models with fixed controls interacted with the time period dummy include time period interactions withan indicator of whether the locality is urban, population density (logged), ethno-linguistic fractionalization, area(square KM), distance to Lilongwe, distance to Blantyre, and the number of boreholes per capita in 1998. The tableshows that ethnic favoritism is more prevalent when the MP’s electoral district is more ethnically segregated. ∗p<0.1;∗∗p<0.05; ∗∗∗p<0.01.
25
DiD Robustness Test: Logit Model
The difference-in-difference (DiD) setup we use in the paper shows that segregation shapes the de-
gree to which politicians engage in ethnic favoritism. These results are based on linear probability
models, and rely on three cutoffs or a continuous (linear) measure of segregation. Here, we specify
a logit model that does not depend on cutoffs and allows for non-linearities. The model, which
predicts the presence of a borehole (Y = 1; 0 otherwise), looks as follows:
Pr(Yidgt = 1) = Λ{
α+G′γ+P′δ+D′β}
for G =
gg
gg ·Sd
gg ·S2d
gg ·S3d
P =
pt
pt ·Sd
pt ·S2d
pt ·S3d
D =
dgt
dgt ·Sd
dgt ·S2d
dgt ·S3d
where i indexes locality, d electoral district, g treatment group, t time period, and Λ{·} is the
CDF of the logistic distribution. The variable gg equals 1 for treated localities (those that became
matched in the second period) and 0 otherwise, pt equals 1 in the second period and 0 otherwise,
dgt equals 1 for treated localities in the second period and 0 otherwise, and Sd is a measure of seg-
regation. This setup allows the DiD estimate to vary with segregation to a third-degree polynomial
(captured by the vector β). We use a third-degree polynomial because we find significant evi-
dence that fit is improved as compared to a second-degree polynomial or including no polynomial.
Figure SI.8 shows the result, and aligns well with the results reported in the paper.
26
Figure SI.8: DiD Estimated Using Logit and a Flexible Function of Segregation
0.0
0.2
0.4
0.6
0.0 0.2 0.4 0.6
Spatial Dissimilarity Index
Pr(
New
Bor
ehol
e), M
atch
v. N
o M
atch
Note: The y-axis is a measure of ethnic favoritism based on a difference-in-differences setup. For example, a 0.4score on the y-axis indicates that the share of newly matched localities that received a new borehole in 1998-2008was 40 percentage points higher than expected given the share of unmatched localities that received a new borehole.The figure provides parametric evidence that the DiD results we report in the paper are not sensitive to a particulardefinition of low, medium, and high segregation.
27
DiD Robustness Test: Clustering by Electoral District-Year
Our main analysis clusters the standard errors on locality because this is the level at which “ethnic
match” is assigned. This approach is similar to Franck and Rainer (2012), who cluster on ethnic
group (survey round), and Burgess et al. (2015), who cluster on district, the levels at which ethnic
match is assigned in their respective studies. While we prefer this approach, the table below
presents the DiD results with standard errors clustered on electoral district – time period (since
allocation decisions are made for both time periods). Although this more conservative approach to
clustering substantially reduces our statistical power, the results are largely robust.
28
Table SI.11: Segregation and Ethnic Favoritism in the Provision of Boreholes (Standard ErrorsClustered by Electoral District-Year)
Low Med. High All All All All(1) (2) (3) (4) (5) (6) (7)
A. Match with MP: Largest Ethnic Group in Locality
Match with MP 0.03 0.16∗ 0.21∗∗ 0.05 0.03 0.05 −0.02(0.10) (0.09) (0.08) (0.09) (0.08) (0.07) (0.11)
Match x Med. Seg. 0.08 0.09 0.07(0.10) (0.08) (0.08)
Match x High Seg. 0.18∗∗ 0.18∗∗ 0.15∗
(0.09) (0.08) (0.08)Match x Continuous Seg. 0.29
(0.24)
B. Match with MP: Proportion Coethnic
Match with MP 0.13 0.20∗∗ 0.57∗∗∗ 0.12 0.09 0.12 −0.20(0.18) (0.10) (0.22) (0.18) (0.15) (0.15) (0.34)
Match x Med. Seg. 0.06 0.09 0.10(0.18) (0.16) (0.15)
Match x High Seg. 0.49∗∗ 0.48∗∗∗ 0.47∗∗∗
(0.20) (0.18) (0.16)Match x Continuous Seg. 0.96
(0.75)
Locality Fixed Effects X X X X X X XTime Period Fixed Effects X X X X X X XTime Varying Controls X X XFixed Controls x Time Period X XNumber of Electoral Districts 42 43 35 120 120 120 120Number of Localities 1315 1542 645 3502 3502 3502 3502Number of Observations 2630 3084 1290 7004 7004 7004 7004
Notes: The table shows difference-in-differences estimates of a locality-MP ethnic match, for different levels ofsegregation. The dependent variable is an indicator for whether a locality has a borehole. Columns (1)-(3) include asubset of electoral districts based on their segregation levels (low, medium, and high, respectively), while (4)-(7)include electoral districts across all levels of segregation. Panel A uses the locality’s largest ethnic group to defineMP ethnic match (Match), while Panel B uses the MP’s share of coethnics (Match Proportion). Models with timevarying controls include controls for ethnic match with the president and the presence of a health clinic or school inthe locality. Models with fixed controls interacted with the time period dummy include time period interactions withan indicator of whether the locality is urban, population density (logged), ethno-linguistic fractionalization, area(square KM), distance to Lilongwe, distance to Blantyre, and the number of boreholes per capita in 1998. Standarderrors are clustered by electoral district-year. ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01.
29
DiD Robustness Test: Randomly Generated Segregation Cutoffs
As an additional robustness check of our DiD results, we randomly vary the segregation cutoffs
used to define low, medium, and high segregation. Figure SI.9 shows 15 replications of Figure 3,
but decomposes the four means for each level of segregation into one summary measure, the DiD
(capturing ethnic favoritism). It then plots the DiD estimate for low, medium, and high segrega-
tion. Each subplot employs a different set of mutually exclusive cutoffs for segregation, randomly
generated subject to the following constraints: the medium category lower cutoff has to fall in the
interval [0.25, 0.4]; the high category lower cutoff is then set 0.12 points higher than the medium
cutoff. This approach ensures that at least 10% of the data are included in each category.
30
Figure SI.9: DiD Estimates For Different Segregation Cutoffs
●
●●
0.0425−0.257
0.257−0.377
0.377−0.795
●
●
●
0.0425−0.266
0.266−0.386
0.386−0.795
●
●
●0.0425−0.299
0.299−0.419
0.419−0.795
●●
●
0.0425−0.33
0.33−0.45
0.45−0.795
●●
●0.0425−0.368
0.368−0.488
0.488−0.795
●
●
●
0.0425−0.263
0.263−0.383
0.383−0.795
● ●
●
0.0425−0.2690.269−0.389
0.389−0.795
●
●
●0.0425−0.321 0.321−0.441
0.441−0.795
●
●
●
0.0425−0.3410.341−0.461
0.461−0.795
●●
●0.0425−0.369
0.369−0.489
0.489−0.795
●
●
●
0.0425−0.265
0.265−0.385
0.385−0.795
●●
●0.0425−0.297
0.297−0.417
0.417−0.795
●
●
●0.0425−0.324 0.324−0.444
0.444−0.795
●●
●0.0425−0.368
0.368−0.488
0.488−0.795
●
●
●
0.0425−0.396
0.396−0.516
0.516−0.795
low medium high low medium high low medium high
low medium high low medium high low medium high
low medium high low medium high low medium high
low medium high low medium high low medium high
low medium high low medium high low medium high
0.0
0.1
0.2
0.0
0.1
0.2
−0.1
0.0
0.1
0.2
0.0
0.1
0.2
0.3
0.0
0.1
0.2
0.3
0.4
0.5
0.0
0.1
0.2
0.0
0.1
−0.1
0.0
0.1
0.2
0.00
0.05
0.10
0.15
0.20
0.25
0.0
0.1
0.2
0.3
0.0
0.1
0.2
0.0
0.1
0.2
−0.1
0.0
0.1
0.2
0.00
0.05
0.10
0.15
0.20
0.25
0.0
0.1
0.2
0.3
Segregation Category
DiD
Est
imat
e
Notes: 15 replications of our DiD results using randomly chosen cutoffs for segregation, subject to some constraints(see text on previous page). The label above each estimate gives the range our segregation measure included in theestimate.
31
Cross-Sectional Ethnic Favoritism Analyses
To show that our DiD results likely extend to the entire population of localities, this section presents
a cross-sectional analysis of ethnic favoritism within electoral districts. We estimate logistic re-
gression models in which the dependent variable indicates whether the locality received a borehole
during the period from 1998 to 2008. The main explanatory variables of interest are Match, which
indicates whether the plurality ethnic group in the locality was coethnic with the MP during the
time period, and Match Proportion, which indicates the proportion of the population in each local-
ity that was coethnic with the MP. We interact these coethnicity measures with the continuous and
categorical measure of electoral-district segregation. Each models includes electoral district fixed
effects and controls for ELF, population size, population density (log), land area (log), number of
boreholes per capita in 1998, whether or not the locality contains an urban area, distance from
Lilongwe, and distance from Blantye. The results are presented in SI.12.
32
Table SI.12: Segregation and Ethnic Favoritism in the Provision of Boreholes
Dependent variable:
New Borehole
(1) (2) (3) (4)
Ethnic Match with MP −0.03 −0.01(0.34) (0.14)
Percent Coethnic with MP −0.37 −0.02(0.62) (0.31)
Ethnic Match x Continuous Segregation 0.60(0.76)
Ethnic Match x Med. Segregation 0.59∗∗∗
(0.20)Ethnic Match x High Segregation 0.04
(0.22)Percent Coethnic with MP x Continuous Segregation 1.82
(1.29)Percent Coethnic with MP x Med. Segregation 0.78∗∗
(0.37)Percent Coethnic with MP x High Segregation 0.47
(0.38)Constant −0.08 0.14 −0.06 −0.03
(0.34) (0.37) (0.34) (0.35)
Constituency Fixed Effects X X X XControl Variables X X X X
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
33
Other Public Goods: Health Clinics and Schools
In this section, we attempt to replicate our findings for the provision of health clinics and schools. Tables SI.13 and
SI.14 report associations between segregation and total investments in clinics and schools, respectively. For clinics,
the association is always positive but statistically insignificant, while we find that more new schools are built in highly
segregated electoral districts. Within electoral districts, we find mixed and generally weak evidence of differential
targeting of clinics (Tables SI.15 and SI.16) and schools (Tables SI.17 and SI.18) by electoral district segregation. One
potential reason for the weaker results is that MPs generally have less discretion over the provision and allocation of
clinics and schools, which are constructed at far lower rates (for example, less than 2% of localities received a new
clinic between 1998 and 2008) and may be more heavily influenced by political decisions at the national level.
34
Table SI.13: Segregation and Clinic Investments across Electoral Districts
Dependent variable:
Number of New Clinics
(1) (2) (3) (4)
Segregation (continuous) 0.27 0.26(1.09) (1.15)
Dummy for Medium Segregation 0.16 0.11(0.26) (0.26)
Dummy for High Segregation 0.38 0.43(0.30) (0.32)
Ethnic Diversity (ELF) 1.72∗∗∗ 0.92 1.72∗∗∗ 0.89(0.60) (0.94) (0.60) (0.96)
Population Density (ln) 0.37 0.40 0.36 0.37(0.32) (0.33) (0.29) (0.32)
Urban Proportion (ln) −0.08 −0.13∗ −0.06 −0.11(0.06) (0.07) (0.06) (0.07)
Land Area (square KM) (ln) 0.52∗∗ 0.46∗ 0.49∗∗ 0.44∗
(0.25) (0.25) (0.25) (0.25)Boreholes per 10,000 residents in 1998 −0.49∗ −0.47∗ −0.49∗ −0.51∗
(0.27) (0.28) (0.27) (0.29)Electoral Competitiveness 0.01 0.004
(0.01) (0.01)MP Coethnic Population Share −1.09∗ −1.09
(0.65) (0.67)President Coethnic Population Share 0.45 1.15
(1.94) (2.00)Distance to Nearest City (ln) 0.80 0.75
(1.29) (1.29)Water Aid Projects per 10,000 residents 0.53 0.54
(0.81) (0.81)All Aid Projects per 10,000 residents 0.02 0.02
(0.07) (0.07)Constant −3.23∗∗ −6.77 −3.02∗∗ −6.31
(1.40) (7.57) (1.38) (7.50)
Admin. District Fixed Effects X X X XObservations 183 182 183 182
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
35
Table SI.14: Segregation and School Investments across Electoral Districts
Dependent variable:
Number of New Schools
(1) (2) (3) (4)
Segregation (continuous) −0.59 −0.16(0.88) (0.91)
Dummy for Medium Segregation 0.20 0.20(0.21) (0.21)
Dummy for High Segregation 0.27 0.42∗
(0.24) (0.25)Ethnic Diversity (ELF) 1.35∗∗∗ 0.56 1.47∗∗∗ 0.62
(0.48) (0.71) (0.47) (0.71)Population Density (ln) 0.73∗∗∗ 0.67∗∗ 0.84∗∗∗ 0.73∗∗∗
(0.26) (0.27) (0.23) (0.25)Urban Proportion (ln) −0.12∗∗ −0.10∗ −0.10∗ −0.07
(0.05) (0.06) (0.05) (0.06)Land Area (square KM) (ln) 0.89∗∗∗ 0.87∗∗∗ 0.90∗∗∗ 0.88∗∗∗
(0.19) (0.18) (0.19) (0.19)Boreholes per 10,000 residents in 1998 −0.43∗∗∗ −0.44∗∗∗ −0.43∗∗∗ −0.45∗∗∗
(0.06) (0.06) (0.06) (0.06)Electoral Competitiveness −0.01∗ −0.01∗∗
(0.01) (0.01)MP Coethnic Population Share −0.31 −0.28
(0.52) (0.51)President Coethnic Population Share 0.62 1.10
(1.44) (1.46)Distance to Nearest City (ln) −1.18 −1.30
(0.99) (0.99)Water Aid Projects per 10,000 residents −0.21 −0.24
(0.59) (0.61)All Aid Projects per 10,000 residents −0.02 −0.01
(0.06) (0.06)Constant −2.56∗∗ 4.85 −2.69∗∗ 5.42
(1.07) (5.75) (1.05) (5.74)
Admin. District Fixed Effects X X X XObservations 183 182 183 182
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
36
Table SI.15: Segregation and Ethnic Favoritism in the Provision of Clinics (Binary Match)
Dependent variable:
Indicator for presence of a clinicLow Med. High All All All
(1) (2) (3) (4) (5) (6)
Ethnic Match with MP −0.01 0.03∗∗ 0.03 0.01 0.004 −0.01(0.01) (0.01) (0.02) (0.01) (0.01) (0.02)
Ethnic Match x Med. Segregation 0.01 0.01(0.01) (0.01)
Ethnic Match x High Segregation 0.005 0.01(0.02) (0.02)
Ethnic Match x Continuous Segregation 0.04(0.06)
Locality Fixed Effects X X X X X XTime Period Fixed Effects X X X X X XTime Varying Controls X XNumber of Electoral Districts 42 43 35 120 120 120Number of Localities 2630 3084 1290 7004 7004 7004Adjusted R2 0.81 0.74 0.85 0.79 0.79 0.79
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
37
Table SI.16: Segregation and Ethnic Favoritism in the Provision of Clinics (Proportion Coethnic)
Dependent variable:
Indicator for presence of a clinicLow Med. High All All All
(1) (2) (3) (4) (5) (6)
Ethnic Match with MP −0.001 0.01 0.07 0.01 0.01 0.01(0.02) (0.01) (0.05) (0.01) (0.01) (0.05)
Ethnic Match x Med. Segregation −0.002 −0.01(0.02) (0.02)
Ethnic Match x High Segregation 0.02 0.02(0.04) (0.04)
Ethnic Match x Continuous Segregation −0.02(0.12)
Locality Fixed Effects X X X X X XTime Period Fixed Effects X X X X X XTime Varying Controls X XNumber of Electoral Districts 42 43 35 120 120 120Number of Localities 2630 3084 1290 7004 7004 7004Adjusted R2 0.81 0.74 0.85 0.79 0.79 0.79
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
38
Table SI.17: Segregation and Ethnic Favoritism in the Provision of Schools (Binary Match)
Dependent variable:
Indicator for presence of a schoolLow Med. High All All All
(1) (2) (3) (4) (5) (6)
Ethnic Match with MP 0.03 0.14∗∗∗ 0.03 0.08∗∗∗ 0.07∗∗∗ 0.12∗∗∗
(0.02) (0.02) (0.03) (0.02) (0.02) (0.04)Ethnic Match x Med. Segregation 0.01 0.01
(0.02) (0.02)Ethnic Match x High Segregation −0.03 −0.05∗∗
(0.02) (0.03)Ethnic Match x Continuous Segregation −0.14
(0.09)
Locality Fixed Effects X X X X X XTime Period Fixed Effects X X X X X XTime Varying Controls X XNumber of Electoral Districts 42 43 35 120 120 120Number of Localities 2630 3084 1290 7004 7004 7004Adjusted R2 0.83 0.75 0.73 0.77 0.78 0.78
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
39
Table SI.18: Segregation and Ethnic Favoritism in the Provision of Schools (ProportionCoethnic)
Dependent variable:
Indicator for presence of a schoolLow Med. High All All All
(1) (2) (3) (4) (5) (6)
Ethnic Match with MP −0.01 0.17∗∗∗ 0.04 0.05∗∗ 0.03 0.15∗
(0.03) (0.03) (0.08) (0.02) (0.02) (0.09)Ethnic Match x Med. Segregation 0.05∗ 0.06∗
(0.03) (0.03)Ethnic Match x High Segregation −0.02 −0.07
(0.06) (0.06)Ethnic Match x Continuous Segregation −0.21
(0.22)
Locality Fixed Effects X X X X X XTime Period Fixed Effects X X X X X XTime Varying Controls X XNumber of Electoral Districts 42 43 35 120 120 120Number of Localities 2630 3084 1290 7004 7004 7004Adjusted R2 0.83 0.75 0.73 0.77 0.78 0.78
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
40
Figure SI.10: DiDs for Clinics (Upper Panel) and Schools (Lower Panel)
●
●
0.00
0.02
0.04
0.06
0.08
0.10
Low Segregation
Pr(
Loca
lity
Has
Clin
ic)
1998 2008
●
Coethnic after 1998Never Coethnic
● ●
0.00
0.02
0.04
0.06
0.08
0.10
Medium Segregation
1998 2008
●
●
0.00
0.02
0.04
0.06
0.08
0.10
High Segregation
1998 2008
●
●
0.0
0.1
0.2
0.3
0.4
Low Segregation
Pr(
Loca
lity
Has
Sch
ool)
1998 2008
●
●
0.0
0.1
0.2
0.3
0.4
Medium Segregation
1998 2008
● ●
0.0
0.1
0.2
0.3
0.4
High Segregation
1998 2008
Note: 3502 localities (enumeration areas) located in 120 electoral districts are included in the analyses. All of theselocalities were not coethnic with their MP in 1998. 1599 localities became coethnic with their MP in either the 1999or the 2004 parliamentary elections; these are denoted with a triangle. The 1903 localities denoted with a circle werenever coethnic with their MP in the study period.
41
Placebo Test: Segregation and the Provision of Private Goods
This section examines whether segregation also affects private goods transfers across and within constituencies. These
analyses serve primarily as a placebo test for our public goods analyses. In particular, some unobserved characteristic
of constituencies may be correlated with both the degree of segregation and the quality of the MP, producing a spurious
relationship between ethnic segregation and the quantity of new local public goods. Similarly, some localities may be
better able to get a coethnic leader elected and more effective in lobbying for new public investments. However, if this
were the case, we would expect segregation to be positively associated with investment in, and the ethnic targeting of,
all types of distributive goods. These analyses thus help us to rule out that selection effects are driving the patterns
we observe in the local public goods data. In addition to serving as a placebo test, these analyses provide a limited
test of an additional observable implication of our theory. Because incumbents should exert less effort to invest in
public goods in less segregated constituencies, they should be more willing to serve their coethnics in other ways —
for instance, by providing private goods.
We study the largest and most politically salient form of private transfer from the Malawian government
to citizens: coupons given to individual households to subsidize the cost of fertilizer and other agricultural inputs
through the Targeted Input Program (TIP) introduced in 2000 (see Harrigan, 2008, for an overview of TIP and earlier
programs).3 The vast majority of Malawians are subsistence farmers growing maize for household consumption. In
recent years, population pressures and reduced soil quality have resulted in declining productivity and increased food
insecurity. In response to declining agricultural productivity and increased food insecurity, the Government of Malawi
instituted a number of programs, culminating in the Targeted Input Program (TIP). TIP was introduced in 2000 and
scaled up after the 2002 famine.4 The program provided seeds, fertilizer, and other agricultural inputs to households,
especially the most vulnerable households. Between 2000 and 2004, an estimated 7 million beneficiaries received
inputs through TIP (Harrigan, 2008).
While TIP was designed to be programmatic (Chinsinga, 2005), in practice political elites exercised con-
siderable discretion over the distribution of subsidies (Chasukwa et al., 2014; Øygard et al., 2003; Tambulasi, 2009)
and evidence suggests that political elites utilized that discretion politically (Mason and Ricker-Gilbert, 2013; Ricker-
3TIP was eventually replaced by a larger scale subsidy program in 2005, after our data were col-
lected.4TIP was eventually replaced by a larger scale subsidy program in 2005, after our data were col-
lected.
42
Gilbert and Jayne, 2011).5
Ideally, we would analyze data on the total number of subsidies distributed within each constituency, and
the geographic location and ethnicity of each recipient within constituencies. Unfortunately, such data do not exist.
Instead, we utilize a nationally representative survey data gathered as part of Malawi’s second Integrated Household
Survey (IHS2), which records whether a household received a TIP coupon during the three years prior to the survey.
The survey was designed by the Government of Malawi’s National Statistics Office and was implemented with support
from the World Bank and the International Food Policy Research Institute (IFPRI). Data were collected on 11,279
Malawians residing in 560 randomly selected Enumeration Areas between March 2004 and March 2005. Fully 53%
of the sample received a TIP transfer.
The sampling procedure of the IHS2 is as follows. The sample includes all three regions: north, center,
and south. The country was first stratified into urban and rural strata. Urban areas include the four major urban
centers: Lilongwe, Blantyre, Mzuzu, and the Municipality of Zomba. The rural strata were further broken down into
27 additional strata corresponding to Malawi’s 27 administrative districts. One district, Likoma, was excluded because
it is an island and difficult to travel to. The sampling was therefore stratified into 30 strata: 26 districts and four urban
areas.
In the first stage of the sampling procedure, EAs were randomly selected from within each strata. The number
of EAs selected was proportional to the total size of the strata: 12 EAs from those with 0 to 75,000 households; 24
EAs from those with 75,000 to 125,000 households; 36 EAs from those with 125,000 to 175,000 households; and 48
EAs from those with 175,000 to 225,000 households. In the second stage, 20 households were selected at random
from within each of the EAs chosen in the first stage. Figure SI.11 maps the EAs for which we have data. Because of
the random sampling of EAs, only 148 of the 193 constituencies had sufficient data to include in our analysis.6
5Dionne and Horowitz (2016) find no evidence of ethnic targeting in the distribution of TIP’s succes-
sor program, Malawi’s Agricultural Input Subsidy Program, in three Malawian districts between
2008 and 2009. However, they only evaluate whether coethnics of the president or member of the
largest three groups were favored, and do not consider the effect of sharing an ethnicity with one’s
MP.6The IHS2 sample was generated by random selection of EAs, and (by chance) did not include data
from 45 electoral constituencies. Figure SI.11 of the online appendix shows the distribution of
sampled EAs.
43
Figure SI.11: EAs Included in the Malawi Integrated Household Survey (IHS2) Sample
44
We first analyze the relationship between constituency-level segregation and overall investments in TIP trans-
fers. The dependent variable is the proportion of households in the constituency that received a transfer. Because this
measure is based solely on a sample, it is measured with error.7 We deal with this partially in our analyses by weight-
ing each constituency by the inverse of the standard error of the constituency level estimate (following Saxonhouse,
1976). Since the dependent variable at the constituency level is the proportion of households receiving a TIP transfer,
we implement a fractional logistic regression model (Papke and Wooldridge, 1996). Because TIP was designed to
benefit the poor and ultra-poor, we control for the proportion of the sample within each constituency that is classified
as such by the IHS2, in addition to controls for ethnic diversity, urban center, and population.
Results are presented in Table SI.19, which shows no significant relationship between segregation and the
provision of these private goods.
Table SI.19: Segregation and Private Goods Provision
Dependent variable:
Proportion Receiving Agricultural Subsity
(1) (2) (3) (4)
Segregation (continuous) 1.96 1.92(2.21) (2.27)
Dummy for Medium Segregation 0.09 0.14(0.47) (0.49)
Dummy for High Segregation 0.41 0.40(0.56) (0.59)
Admin. District Fixed Effects Fixed Effects X X X XControl Variables X X X XObservations 166 165 166 165
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
We also examine ethnic favoritism in TIP allocations within constituencies. To do so, we create a dichotomous
ethnic match variable that takes a value of 1 if the household head in the IHS2 survey shares an ethnicity with the MP
of the constituency in which the household is located, and 0 otherwise. However, because IHS2 does not ask about
7On average, each constituency sample includes 132 individuals, ranging from 39 to 344 con-
stituents.
45
ethnicity explicitly, we use language as a proxy for ethnicity.8 Because the data cover TIP receipts between 2001-
2004, we only code individual survey respondents’ ethnic linkage to the MP that was elected in their constituency in
1999. We interact this individual-level indicator of ethnic match with constituency-level segregation measures. Since
our dependent variable is a binary indicator of receipt of TIP at the individual level, we use a logistic regression, with
standard errors clustered by electoral constituency. We include dummies for whether a household is considered poor
or ultra-poor, with non-poor households as the omitted category, as well as constituency fixed effects.
Table SI.20 presents the results. We find evidence that there is more ethnic favoritism in the allocation of
private goods in integrated electoral districts. This is precisely the opposite pattern than the one we uncover with local
public goods, where ethnic favoritism is increasing in segregation.
Table SI.20: Segregation and Ethnic Favoritism in the Provision of Private Goods
Dependent variable:
Receipt of Agricultural Subsidy
(1) (2) (3) (4) (5)
Ethnic Match with MP 0.13∗∗∗ 0.05 0.09∗∗ 0.13∗∗∗ 0.08(0.03) (0.05) (0.04) (0.03) (0.06)
Ethnic Match x Med. Segregation −0.08(0.06)
Ethnic Match x High Segregation −0.04(0.05)
Ethnic Match x Continuous Segregation 0.01(0.13)
Constituency Fixed Effects X X X X XIndividual Level Controls X X X X XAdjusted R2 0.32 0.10 0.10 0.19 0.19
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
In summary, unlike with public goods, private goods provision is not affected by segregation. Moreover,
8We code respondent ethnicity as the ethnic group associated with the respondent’s home language.
In Malawi, language uniquely identifies some but not all ethnic groups. Afrobarometer survey data
from 2005, which asks about both ethnic identity and home language, suggest that language is an
appropriate indicator of ethnicity for around 75% of the population.
46
ethnic favoritism in the allocation of private goods is greatest in integrated electoral districts and decreasing with
segregation. Thus, the results from this section help to allay concerns that omitted variable bias is driving the results
of our public goods analyses.
47
References: Supporting Information
Burgess, R., Jedwab, R., Miguel, E., Morjaria, A., and Padro i Miquel, G. (2015). The value of democracy: Evidence
from road building in Kenya. American Economic Review, 105(6):1817–51.
Chasukwa, M., Chiweza, A. L., and Chikapa-Jamali, M. (2014). Public participation in local councils in Malawi in
the absence of local elected representatives: Political elitism or pluralism? Journal of Asian and African Studies,
49(6):705–720.
Chinsinga, B. (2005). Practical and policy dilemmas of targeting free inputs. In Levy, S., editor, Starter Packs:
a Strategy to Fight Hunger in Developing Countries? Lessons from the Malawi Experience 1998–2003, pages
141–153. CABI Publishing, Wallingford, UK.
Chinsinga, B. (2011). The politics of land reforms in Malawi: The case of the Community Based Rural Land Devel-
opment Programme (CBRLDP). Journal of International Development, 23(3):380–393.
Cutler, D. M., L., G. E., and Vigdor, J. L. (2012). The rise and decline of the American ghetto. Journal of Political
Economy, 107(3):455–506.
Dionne, K. Y. and Horowitz, J. (2016). The political effects of agricultural subsidies in Africa: Evidence from Malawi.
World Development, 87:215–226.
Franck, R. and Rainer, I. (2012). Does the leader’s ethnicity matter? Ethnic favoritism, education and health in
Sub-Saharan Africa. American Political Science Review, 106(2):294–325.
Government of Malawi (1994). 1994 presidential and parliamentary general election results. The Malawi Government
Gazette Extraordinary, XXXI(40):267–282.
Government of Malawi (1999). Parliamentary general election results. Available from the Sustainable Development
Network Programme website at http://www.sdnp.org.mw.
Government of Malawi (2004). Parliamentary general election results. Available from the Sustainable Development
Network Programme website at http://www.sdnp.org.mw.
Harrigan, J. (2008). Food insecurity, poverty and the Malawian Starter Pack: Fresh start or false start? Food Policy,
33(3):237–249.
48
Mason, N. and Ricker-Gilbert, J. (2013). Disrupting demand for commercial seed: Input subsidies in Malawi and
Zambia. World Development, 45:75–91.
Matuszeski, J. and Schneider, F. (2006). Patterns of ethnic group segregation and civil conflict. Working Paper,
Harvard University.
Øygard, R., Garcia, R., Guttormsen, A., Kachule, R., Mwanaumo, A., Mwanawina, I., Sjaastad, E., Wik, M., Bonner-
jee, A., Braithwaite, J., Carvalho, S., Ezemenari, K., Graham, C., and Thomson, A. (2003). The maze of maize:
Improving input and output market access for poor smallholders in the Southern African region. Department of
Economics and Resource Management, Agricultural University of Norway Report No. 26.
Papke, L. E. and Wooldridge, J. M. (1996). Econometric methods for fractional response variables with an application
to 401(k) plan participation rates. Journal of Applied Econometrics, 11:619–32.
Reardon, S. F. and O’Sullivan, D. (2004). Measures of spatial segregation. Sociological Methodology, 34(1):121–162.
Ricker-Gilbert, J. and Jayne, T. S. (2011). What are the enduring effects of fertilizer subsidy programs on recipient
farm households? Evidence from Malawi. Michigan State University’s Department of Agricultural, Food, and
Resource Economics Staff Paper No. 2011-09.
Saxonhouse, G. R. (1976). Estimated parameters as dependent variables. American Economic Review, 66(1):178–183.
Tambulasi, R. I. (2009). The public sector corruption and organized crime nexus: The case of the fertiliser subsidy
programme in Malawi. African Security Review, 18(4):19–31.
49