Experienced poverty and local conflict violence

Abstract

That poorer countries face higher risk of civil war is among the most robust findings in the literature on internal conflicts. However, we lack knowledge about whether a similar correlation exists on a more local level. Research into the local poverty–conflict nexus has largely relied on objective proxies of poverty that are only loosely related to the rationale for conflict. The results have been mixed, thus highlighting the need for more effective juxtaposition of theory and data. Using a subjective measure of poverty that determines whether individuals’ basic needs are being met, this article presents new empirical evidence for existing propositions linking local poverty and conflict-based violence. The study analyzes georeferenced survey data from the pan-African Afrobarometer survey for 4008 subnational districts across 35 African states, producing results that show how areas with high levels of poverty are indeed more likely to experience conflict. However, the correlation is likely to be indirect. Interaction models demonstrate that poverty is more likely to exacerbate violence if an area’s local institutions are weak or when impoverishment overlaps with group grievances against the government. Robustness tests, using coarsened exact matching and region-level fixed effects, provide considerable empirical support for a strong relationship between poverty and conflict at the local level.

Keywords

Conflict GIS governance poverty spatial data subnational

Introduction

Civil wars are more frequent than any other type of conflict in the modern era, with the majority occurring in low-income countries (Hegre and Sambanis, 2006; Jakobsen et al., 2013). While most country-level studies find that poverty and inadequate economic development increase the risk of conflict—a relationship that appears to be causal (Braithwaite et al., 2016)—we lack consensus on the precise mechanisms driving this phenomenon (Justino, 2009). Researchers have explained a correlation between low GDP per capita and conflict using diverse hypotheses, including lowered opportunity costs for individuals to rebel (Collier et al., 2009) and responses to a state’s weak capacity (Fearon and Laitin, 2003).

However, as argued by Hegre (2016), development’s highly correlated indicators make it difficult to distinguish between the theoretical mechanisms underlying the development–conflict nexus. Moreover, previously proposed models often represent processes operating on various geographical scales at individual, group, and state levels. Few researchers have backed up theoretical expectations with data at scientifically fitting levels of analysis, consequently ignoring intra-country variations of explanatory variables and outcomes. Furthermore, aggregated measures are incapable of capturing significant variations in economic conditions (Elbers et al., 2003) and conflict intensity (Rustad et al., 2011) within countries. In addition, conflict areas are, in general, atypical of a nation as a whole (Buhaug and Lujala, 2005), which calls for a subnational level analysis.

Addressing these disconnects—and the fact that most conflict operates at a local level (Rustad et al., 2011)—a recent body of studies has focused on how subnational variations in poverty determine the locations within a country where conflicts break out (Buhaug et al., 2011; Hegre et al., 2009; Østby et al., 2009). To date, their findings are largely mixed, with no consensus yet on strength, direction, or mechanisms behind the relationship. The problem here may be the use of varying proxies for poverty that are only loosely linked to the rationale for conflict and/or insufficient attention on the local sociopolitical context.

The present study’s empirical contributions seek to help rectify the inadequate measures of poverty that have come to characterize the literature. To begin with, the article improves our understanding of whether and where a local poverty–conflict nexus exists by deploying experiential data on individuals’ actual wellbeing—which I argue is more closely connected to people’s motives and rationale for taking up arms. Second, the article examines the sociopolitical context’s conditioning effect on the poverty–conflict nexus. This is achieved by including data on individuals’ perceptions surrounding the quality of their local institutions, the presence of group grievances, and local unemployment rates. These factors, I argue, are more closely linked to reasons for fighting than are common proxies such as night-time luminosity and estimates of economic activity, both of which are often derived from dividing GDP per capita by local population counts.

Poverty—a state in which individuals’ basic needs go unmet—has been shown to motivate people to join rebellions. Humphreys and Weinstein (2008), for instance, found that poverty predicted inscription in the Revolutionary United Front during Sierra Leone’s civil war. Barrett (2011) similarly saw how promises of loot lured the poor to enlist in the 1997–1998 dispute in Nigeria’s local government area known as Toto. Combatants of the Toto conflict were also more likely to join the rebellion if they stood to gain personal protection, food, and shelter.

For the present study, I developed a dataset by aggregating survey responses from the pan-African Afrobarometer survey to subnational districts and combining the results with information on post-survey violent conflicts. The dataset consists of 4008 subnational districts, spanning 35 African countries. As most districts were only assessed once, thus restricting study of within-unit variation, survey responses were also aggregated to higher-order subnational regions, resulting in a dataset of 111 regions that were surveyed at least twice; this permitted a region-level fixed-effects model design.

Using a pooled cross-sectional dataset of districts, I found that high levels of poverty were linked to increases in local conflict-based violence. Districts with a large share of poor individuals, both in absolute terms and relative to country average, had a higher risk of conflict than more affluent areas. This relationship held in a coarsened exact matching setup, as well as in a region-level fixed effects design with repeated measurements across time. While the results reveal a local poverty–conflict link, they do not aid in uncovering underlying mechanisms.

Using interactions models, I found that poverty increased the risk of conflict, although only where local institutions are weak. The results also show that poverty-stricken areas in which individuals strongly perceive group injustice have a greater risk of conflict than similarly impoverished regions with no aggrieved population. A departure from the local individual opportunity cost explanation, local economic opportunities do not seem to condition the poverty–conflict nexus. In sum, the results suggest that while poverty is significantly connected to conflict, high-quality institutions and inclusiveness of ethnic groups can prevent violence. Although a wide range of robustness checks and alternative model specifications were implemented, including matching and fixed-effects models, the issue of endogeneity could not be ruled out; doing so would require some kind of exogenous instrument, which I have been unable to identify.

The remainder of this article elaborates on the theoretical framework linking subnational poverty to local conflict-based violence. This is followed by a discussion of existing methods for measuring local poverty and their potential shortcomings. Next presented is the study’s research design and modeling strategy, followed by a discussion of empirical results. The conclusion considers the study’s limitations and proposes avenues for future research on poverty in locations that support rebel groups.

Poverty and conflict

A direct link

A connection between low income and risk of conflict is among the most robust findings in the literature on civil wars (Hegre and Sambanis, 2006). However, there is little consensus on the mechanisms through which poverty may produce conflict. Collier and Hoeffler (1998) claimed that low per-capita income lowers the opportunity cost of rebellion because when they have less to lose from taking up arms, poorer individuals become more inclined to rebel. Fearon and Laitin (2003) observed that poorer countries experience more conflict because they are unable to monitor and control all of their territory, thereby creating pockets of hospitable conditions for insurgents; Tollefsen and Buhaug (2015) identified a similar scenario at the local level.

Other researchers have argued that poverty may generate grievances and motivate rebellion. For example, Davies (1962) posited that periods of economic growth followed by recession may boost a revolutionary spirit as individuals worry about work prospects and their investments going to waste. Similarly, Gurr (1970) noted that relative deprivation may lead to conflict as the gap widens between desired and actual levels of individual wellbeing; once a level of extreme poverty is reached, however, the likelihood of resorting to violence decreases dramatically. As Hobsbawm (1959: 79) stated: “When people are really hungry they are too busy seeking food to do much else.”

The literature is rife with case studies of poverty-motivated conflicts. De Soysa and Gleditsch (1999) argued that the conflicts in Liberia and Sierra Leone resulted from the weak states’ failure to fulfill basic needs and stimulate employment, leading aggrieved populations to battle over resources and rally against any remaining state power. In Uganda, Deininger (2003) showed that meager economic opportunities at the community level increased the likelihood of conflict. Ikejiaku (2012) noticed that participants in Niger Delta attacks were often poor, unemployed youth. Irobi (2005) also pointed to poverty and unemployment as the direct causes of conflict in South Africa.

In their study of ex-combatants, Humphreys and Weinstein (2008) noted how poor individuals were more likely to join the Revolutionary United Front in Sierra Leone. Analyzing the Rwandan genocide, Friedman (2013) observed that higher rates of local unemployment among Hutu populations were associated with higher participation rates. Similarly, Nillesen and Verwimp (2009) determined that rebel recruitment activities in Burundi primarily took place in economically deprived villages. Barrett (2011) saw how promises of loot incentivized the poor to join in the rebellion during Nigeria’s Toto conflict.

While both country- and micro-level studies suggest a positive correlation between poverty and conflict, large-N subnational studies have yielded conflicting results for a local link. For instance, Buhaug et al. (2011) used spatial data from the PRIO-GRID (Tollefsen et al., 2012) on gridded economic activity (from the dataset in Nordhaus et al., 2006) to determine that conflict breakouts were more likely in poorer areas. However, their findings also suggest that areas with considerably higher economic activity than the country average in very poor countries have a higher risk of conflict onset. Others have used survey data on household assets as a proxy for poverty. Østby et al. (2009) found that regions with strong relative deprivation were at higher risk of civil conflict, while Hegre et al. (2009) pinpointed conflicts to Liberia’s richer areas.

Another strand of research has focused on economic shocks as catalysts for conflict. Berman and Couttenier (2015) used data on global demand for an area’s agricultural commodities to show how positive local income shocks were negatively and significantly related to subnational conflict. Using night-time luminosity, Ahrens (2015) determined that economic growth shocks did not increase risk of conflict.

As the section above shows, subnational studies have employed a wide range of proxies for poverty, resulting in mixed findings. That said, even if researchers identify an empirical link between local poverty and conflict, it is unclear if what is being captured are poverty levels among combatants or overall poverty of combat zones. Hence, there is an actor–action discrepancy, wherein rebels may draw motivation and support from one place, but target another (Hegre et al., 2009).

It is necessary to recognize these potential shortcomings, but also recall that conflicts typically affect specific areas within a country (Hallberg, 2012). Disadvantaged groups with explicit ethnic claims or secessionist aims, furthermore, tend to fight in fixed locations (Beardsley et al., 2015). Deprived and marginalized groups are often residing in peripheral areas (Cederman et al., 2013), rarely with adequate capacity for, or even interest in, launching attacks on the capital city (Buhaug, 2006). The implication is that conflict gets concentrated in poorer, less developed areas of a country, where non-access may improve rebels’ relative capacity for conflict and rebellion (Tollefsen and Buhaug, 2015). Separatists, as well as those rebel groups with governmental aspirations, may benefit from fighting in areas where they are likely to gain local support (Kilcullen, 2011). As Beardsley et al. (2015) posited, rebel groups generally prefer to keep their operations local, although that decision ultimately depends on the strength of local support, local resources, and the government’s willingness to target the insurgents. A perfect distinction between support locations and target areas seems impossible, not least since the two seem to overlap in places; still, we must acknowledge this limitation when interpreting results in this study as well as in subnational research on target locations in general.

Hypothesis 1: Areas with a large share of poor individuals face increased risk of conflict-based violence.

An indirect link

Some scholars have traced a direct relationship between poverty and conflict, although most would see the link as indirect. For this study, I employed interaction models to explore potential contextual moderators in the following scenarios: (a) when the individual opportunity cost of joining a rebel organization is low (Collier and Hoeffler, 1998); (b) when the state’s capacity is weak (Fearon and Laitin, 2003); and (c) when ethnic group grievances exist (Cederman et al., 2013). To test these mechanisms, I applied local unemployment rates, the quality of local institutions, and the intensity of local group grievances to the interaction models with poverty.

First, the individual opportunity cost explanation suggests that individuals without regular income are more likely to enlist in rebel militias and partake in combat activities (Collier and Hoeffler, 1998). Poor individuals are easier to recruit because they have less to lose when taking up arms (Collier and Hoeffler, 2004). Jakobsen et al. (2013) tested several proxy measures of poverty using factor analysis, finding that per-capita income was more closely related to a wealth/poverty distinction than grievances and state capacity. Their results show that low income increases the risk of conflict, in keeping with the individual opportunity cost explanation.

Studying the effect of unemployment on conflict, Berman et al. (2011) deployed survey data from Afghanistan, Iraq, and the Philippines and found no support for the individual opportunity cost explanation. Their results showed that unemployment in these countries was not related to increased risk of political violence. Yet employed individuals may still partake in rebellion if its anticipated earnings outweigh income from regular work (Justino et al., 2013).

According to the individual opportunity cost explanation, joining a militia or a rebellion becomes economically viable if the income from such activities exceeds the recruit’s conventional wage packet. That said, areas with high unemployment may provide more potential recruits who are willing to join rebel factions. Nillesen and Verwimp (2009) observed how recruitment activities in Burundi targeted its most deprived regions, where individuals’ threshold for enlisting was lower than in affluent areas.

Hypothesis 2: Areas with a large share of poor individuals face increased risk of conflict-based violence when unemployment is high.

Second, the likelihood of poverty sparking conflict depends on how effectively local institutions enforce rule of law, fulfill basic needs, and cooperate with local populations, thereby earning their trust. Institutions that carry out such tasks are considered high quality, more capable of controlling their territory and deterring potential challengers.

At the country level, low GDP per capita has been taken as an indicator of weak state capacity (Fearon and Laitin, 2003). When a state is weak, rebel groups have better chance to gain ground. Conversely, when a state is strong, institutions are more capable of resolving dissension before it escalates to violence, defusing uprisings and discouraging insurgents before they establish a foothold. State weaknesses breed opportunities for political and military conflicts (Fearon and Laitin, 2003). As Benson and Kugler (1998: 206) put it, “politically efficient governments are much more likely to avert internal challenges.”

In explaining which countries experience conflict, Holtermann (2012) showed that states’ capacity and reach outweigh the effect of poverty. Poorer countries are often incapable of controlling the rural peripheries of their territory (Buhaug, 2010), making rebellion in these inaccessible areas more likely (Tollefsen and Buhaug, 2015). Rebels can take advantage of power vacuums to sideline the government and establish local political and military control. This may afford them the chance to restore public services (Kalyvas, 1999) to build trust among local populations, secure a support base, and collect income tax put toward financing future operations.

At the subnational level, Wig and Tollefsen (2016) showed that areas managed by high-quality local institutions—classified as uncorrupt, law-governed, capable, trusted by the public, and efficient in performance—are less likely to experience violent conflict. To begin with, such institutions have a strong legal system and are thus better prepared to resolve local grievances and disputes before they become violent. Such institutions, aided by an upright police force, can curtail uprisings before their instigators become serious threats. Strong local governments are more likely to develop and maintain infrastructure needed to control their territory and bolster trade and economic growth. Finally, areas with strong institutions are more likely to fulfill basic needs in a competent and equitable manner. They have the capacity not only to deter potential rebels, but also to adequately provide for their citizens. In sum, impoverished areas with high-quality institutions are less likely to experience conflict than comparably poor regions with low-quality institutions.

Hypothesis 3: Areas with a large share of poor individuals face increased risk of suffering conflict-based violence when local institutions are weak.

Third, the overlap of poverty and group grievances may fuel collective motivation to engage in violent conflict. Horowitz (1985), for one, noted how groups who face discrimination are more likely to rebel than others. Gurr (1993) posited that countries inhabited by disadvantaged groups were more likely to experience conflict. Addressing the individual level, Gurr (1970) had also suggested that relative deprivation may cause conflict when the gap widens between desired and actual living conditions. Dissatisfaction with the status quo could lead citizens to attempt a government overthrow or consider secession if they believe it may improve their circumstances as individuals or a group. As Gurr (1970: 12–13) explained, “the primary causal sequence in political violence is first the development of discontent, second the politicization of that discontent, and finally its actualization in violent action against political objects and actors.”

Studies have shown that horizontal inequality between groups is a major determinant for conflict (Stewart, 2002), potentially causing group grievances and, ultimately, violence (Østby, 2008). While poverty may in itself cause discontent, its intersection with group perceptions of injustice and unfair treatment by the government can incentivize rebellion.

Various researchers have suggested that horizontal social and economic inequality may increase the risk of conflict (Østby, 2008; Stewart, 2002). Analyzing data on household assets from the Demographic Health Survey, Østby (2008) showed that horizontal social inequality positively correlates with conflict outbreak. Asset inequality between regions has also been shown to increase the risk of civil strife (Deininger, 2003). Similarly, Østby et al. (2009) found that areas with significant relative deprivation have a higher risk of civil war. Others have noted that political inequality, manifested as political exclusion, increases the likelihood of conflict at both group (Cederman et al., 2013) and subnational levels (Tollefsen and Buhaug, 2015). Cederman et al. (2011) observed that both affluent and disadvantaged groups engage in violence more often than those whose wealth approximates the national average. This finding was reiterated by Cederman et al. (2015), using novel data triangulation. In sum, impoverished groups with no grievances against the government have fewer motives to rebel than comparably poor segments who perceive themselves as being unfairly treated by the state.

Hypothesis 4: Areas with a large share of poor individuals face increased risk of conflict-based violence when local group grievances are present.

Measuring subnational poverty

It is inherently difficult to conceptualize and operationalize poverty. This is especially obvious at the subnational level, for which data are scarce, particularly in developing countries. To answer crucial questions—whether a country’s poorer areas are more likely to experience conflict, for example—subnational data must be reliable and adequately disaggregated, thus enabling proper identification of “the poor.” Yet the means by which we calculate poverty, in absolute or relative terms or using objective or subjective measures, have a significant impact on who is defined as poor (Grusky and Kanbur, 2006).

At the national level, developed countries typically provide reliable accounts of gross domestic income. In most African countries, however, economic figures are limited and poor in quality (Jerven, 2013); subnational data is rarely available and, if figures exist, they are aggregated to higher-order administrative levels. To overcome such limitations, researchers have turned to proxies of subnational poverty such as night-time light emissions (NOAA, 2014), spatially gridded economic estimates (Nordhaus, 2006), and administrative-level infant mortality rate (IMR) (Storeygard et al., 2008).¹ While the proxies provide researchers with useful variations in economic conditions within countries, they have serious limitations.

First, night-time luminosity data are unable to capture variations in poverty levels in areas without light emissions, such as rural, densely populated regions—which are widespread across Africa. This is an unfortunate shortcoming, as conflict tends to cluster in these less accessible areas (Tollefsen and Buhaug, 2015). Weidmann and Schutte (2017) demonstrated that night-time data accurately predict wealth in terms of household assets, and while this enhances the credibility of this proxy for poverty, the fact remains that many areas in Africa have no night-time luminosity. The challenge thus is accurately predicting local poverty levels in rural areas with no light as well as in urban areas where night-time luminosity is highly saturated, thereby blurring the distinction between poorer and richer areas.

Second, gridded data on the geographical intensity of economic activity, the gross cell product (GCP), capture activity per unit rather than per capita; they do not capture personal intensity of economic activity (Nordhaus et al., 2006: 5). In addition, the measure is not context-sensitive, as it calculates an area’s economic activity by dividing the country’s GDP for the population in a 1 × 1-degree grid cell.²

Finally, local IMR data (Storeygard et al., 2008) are developed using the lowest subnational-level data available. For most African countries, this equates to regional- or, in some cases, country-level data, resulting in little to no variation. To illustrate, in this study’s dataset, Nigeria, has only three distinct values, Uganda has five, the Democratic Republic of Congo has one, and Kenya has five.

Recent innovations in estimating subnational poverty have involved a combination of satellite imagery and survey data. A novel contribution was offered by Jean et al. (2016), who employed machine learning to extract socioeconomic data from high-resolution daytime satellite imagery. Weidmann and Schutte (2016) used night-time luminosity to predict local wealth. Both teams deployed data on household assets derived from the Demographic Health Survey. While this means of measuring local poverty is innovative, it cannot fully capture expenditure (Howe et al., 2009). Another study showed how, in some settings, asset-based indices erroneously place households in the poorest group (Filmer and Scott, 2012).

This study therefore conceptualizes and operationalizes poverty in terms of basic needs. This approach seeks to get around some of the aforementioned limitations and the lack of local expenditure measures, but also to adopt a measure that is more closely related to the theoretical rationale for conflict. The study employs the Lived Poverty Index (LPI), which consists of five variables in the Afrobarometer measuring the frequency of respondents’ lack of food, water, medicines, fuel, and cash income (Dulani et al., 2013). These five questions are asked across all survey rounds as well as countries in the Afrobarometer survey, enabling a comparable measure of poverty at both country and subnational levels, as well as across time.

Why use an experiential measure of poverty? While such measures may raise concerns over endogeneity, they offer a more direct assessment of the wellbeing of local populations. To reiterate, poverty—unmet basic needs, specifically—has been shown to sway individuals to enlist in rebellions and combat. Thus, an experiential subjective measure that taps into people’s wellbeing captures poverty more directly and connects more closely to the rationale for taking up arms.

The multidimensional LPI measures of experienced poverty are rarely recorded in other country- or subnational-level studies. Contrary to objective proxies of poverty, which make assumptions about local populations’ wellbeing, the LPI measure redirects to what Sen (1999) defined as being at the core of the concept of poverty. While absolute definitions of poverty point to the minimum income needed to fulfill fundamental needs, an experiential measure dodges the shortage of subnational expenditure data and captures to what degree the needs are actually met. Using subjective assessments of poverty also overcomes challenges from setting a predetermined poverty threshold (Ravallion and Lokshin, 2001); such assessments can better capture in-kind benefits from subsidized healthcare, education, sanitation, housing, and household scale economies (Posel and Rogan, 2016).

Subnational experience-based poverty measures enable a context-sensitive intermediate approach between objective–quantitative and subjective–qualitative schools of thought. This approach can better capture heterogeneity in poverty, wherein people’s wellbeing—regardless of income or expenditure—is allowed to vary (Pradhan and Ravallion 2000). Several studies have explored the correlation between objective and subjective measures of poverty. While considerable agreement seems to exist, there are also notable differences. Pradhan and Ravallion (2000) observed that subjective measures reveal even larger differences between rural and urban poverty than do objective measures. Comparing subjective and money-metric approaches to measure poverty in South Africa, Posel and Rogan (2016) found that despite considerable overlap, major discrepancies appeared between the two for larger households that have small proportions of young children and that can access water and electricity as well as for households with farming land. The discrepancies may indicate that expenditure measures of poverty are not properly capturing home production, which is particularly significant in rural areas in Africa (International Fund for Agricultural Development, 2001).

This study uses Afrobarometer data on individuals’ shortages of basic needs fulfillment for the 12 months preceding the interview date. The sample is a nationally representative cross-section of all citizens of voting age in each of the 35 countries that Afrobarometer surveys. Its design takes a clustered, stratified, and multistage probability approach. Surveyors begin by selecting subnational units by regional stratification, which reduces the likelihood of excluding people living in regions with predominant ethnic or language groups. Next, they select primary sampling units (PSUs) within each region. Starting from an initial sampling location within the PSU, the surveyors then randomly choose informants (Afrobarometer, 2016).

The LPI is the mean response to five Afrobarometer questions, asking respondents: over the past year, how often, if ever, have you or your family gone without enough: Q8A food to eat; Q8B clean water for home use; Q8C medicines or medical treatment; Q8D enough fuel to cook your food; and Q8E a cash income? Response options included never; just once or twice; several times; many times; and always. Values were recoded 0–4 and averaged for each individual and then for respondents within the same geographical unit.³

To assess how the LPI measure compares with previously used proxies of local economic conditions, data on IMR from Storeygard et al. (2008), gross cell economic activity (Nordhaus, 2006), and night-time luminosity (NOAA, 2014) were aggregated to the district units. Table 1 shows that directions of the relationships are as expected, but also reveals only a moderate correlation between the LPI and prior measures. Districts with high levels of experienced poverty have a higher IMR, lower economic activity, and less night-time luminosity. The weak correlation suggests that these measurements do not adequately proxy people’s experienced poverty. In addition, GCP and night-time luminosity are highly correlated with population density, whereas the LPI is not. This low correlation suggests that experienced poverty captures variations in wellbeing that are not identified by the objective proxies. Of the 4008 districts containing LPI measures, almost 20% had no night-time luminosity.⁴

Table 1.

Rank correlations at the district level between the Lived Poverty Index (LPI) and previously used subnational proxies

	LPI	IMR	GCP	Night-time luminosity	Population density
LPI	1.00	0.34	−0.30	−0.28	−0.16
IMR	0.34	1.00	−0.41	−0.25	−0.12
GCP	−0.30	−0.41	1.00	0.50	0.70
Night-time Luminosity	−0.28	−0.25	0.50	1.00	0.47
Population Density	−0.16	−0.12	0.70	0.47	1.00

IMR, Infant mortality rate; GCP, gross cell product.

Figure 1 depicts northwestern Uganda’s rural areas. The left panel’s dark pixels represent the two sole areas exhibiting night-time luminosity: Arua in the west and Gulu in the east. While night-time luminosity data pick up emissions from urban clusters, they fail to identify variations in poverty outside of and within urban areas (Jean et al., 2016). The right panel shows the mean levels of experienced poverty at the district level. If aggregated to the district level, all districts outside the lit areas would be assigned a 0, making it unfeasible to assess wellbeing in these areas.

Figure 1.

Night-time luminosity and Lived Poverty Index (LPI) in Uganda. Left panel shows night-time luminosity in northwestern Uganda as dark pixels, which represent Arua in the west and Gulu in the east. Right panel shows experienced poverty at the district level for the same area. Darker-shaded districts represent higher levels of experienced poverty. Points show primary sampling units (PSUs) surveyed (Round 3, 2005).

Data and research design

To test the four hypotheses, I aggregated survey data from three consecutive survey rounds of the Afrobarometer to subnational units by intersecting respondents’ village locations with polygons representing subnational districts (level 2) and regions (level 1).⁵ Regions are larger, although the number of respondents encompassed is higher, both spatially and temporally. Using regions also increases the number of units with data at multiple points in time, making a fixed-effects approach more feasible. It also enables testing whether the results are sensitive to changing the unit of analysis, which could raise concerns over a modifiable areal unit problem. The location of respondents is provided by Knutsen et al. (2017), who used a partial string matching method to identify coordinate pairs for each respondent’s place name. Respondents whose towns or villages were not georeferenced were excluded from the analysis.⁶

Averaging respondents within similar subnational units resulted in data for 4008 districts in 35 African countries, with 1094 districts in round 3 (surveyed in 2005–2006), 1137 in round 4 (surveyed in 2008), and 1777 in round 5 (surveyed in 2012). For the region fixed-effects model, I excluded regions observed only once and those without variation in the outcome variable. This produced 311 observations for 111 unique regions.

Moderator variables

To deepen understanding of the mechanisms behind the poverty–conflict nexus, Justino (2009) proposed the use of interaction models. Thus, in this study I employ interaction models to test three conditional hypotheses (H2–H4), interacting subnational poverty with local unemployment rates, local institutional quality, and the presence of local grievances.

To test the individual opportunity cost explanation, I included a variable indicating the share of currently unemployed Afrobarometer respondents in each district. The information was derived from a question asking: “Do you have a job that pays a cash income?” Respondents who answered “No” were coded as unemployed; those stating they had full- or part-time work were coded as employed. The median unemployment rate for districts is 67.5%; 4.6% of the districts have unemployment rates below 20%; 28% of the districts have unemployment rates above 80%. One potential limitation is that the variable does not capture informal labor; the Online Appendix (section 7) includes a model controlling for the share of respondents working in the informal sector.

Following Wig and Tollefsen (2016), I developed a proxy of local state capacity by constructing an index of variables capturing respondents’ perceptions concerning corruption among local government councilors, police, and tax officials; trust in their local government council and courts; and whether respondents approved or disapproved of their local councilors’ performance. Wig and Tollefsen (2016) used all these variables except how the local government performed at creating jobs and maintaining infrastructure. I, however, include these two variables to tap into the capacity and quality of local institutions. To reduce the number of variables, which could introduce multicollinearity, and to test whether the variables jointly represent some common latent concept, I used exploratory factor analysis.⁷ The direction of each variable was recoded, with 1 representing low quality and 4 high quality. The resulting measure was calculated by taking the mean of index components; it ranges from 1.094 to 3.75, with a median quality of 2.4.⁸

To account for the presence and magnitude of group grievances, I calculated how many district residents perceived their group as unfairly treated by the government. The variable measures the share of respondents perceiving their group as always being unfairly treated. The resulting variable indicates that in 12% of the districts more than half of the population feel they are always unfairly treated.⁹

Dependent variable

To measure conflict intensity in each unit, I used geographic information system (GIS) technology to calculate the number of conflict events occurring within three years after the survey.¹⁰ The dependent variable comes from the Uppsala Conflict Data Program Georeferenced Event Dataset (UCDP GED, version 2.0; Sundberg and Melander, 2013), which provides information on the location of conflict events. An “event” is defined as an “incident where armed force was by an organized actor against another organized actor, or against civilians, resulting in at least 1 direct death at a specific location and a specific date.” For inclusion in the UCDP GED, an event must belong to a conflict causing over 25 annual battle-related deaths following the PRIO/Uppsala Armed Conflict Dataset definition (Gleditsch et al., 2002). Events are coded using spatial and temporal information derived from news sources. UCDP GED includes state-based, non-state, and one-sided event types. The dependent variable includes all three types of events.¹¹Figure 2 shows conflict events across districts in Uganda, superimposed on aggregated district data from Afrobarometer (round 3) on experienced poverty.

Figure 2.

Mean Lived Poverty Index (round 3, 2005) in Uganda overlain with conflict events (included in the UCDP GED) from three subsequent years, shown as dots. Darker shading indicates higher poverty levels.

To account for the effect of previous conflict, I included a half-life parameter, whereby the effect of previous conflict decreases exponentially with years since the last conflict. Areas to have experienced conflict in the past year obtain the maximum value of 1; that is halved after two peaceful years and infinitesimally towards 1989, when UCDP GED coding began. The half-life parameter is constructed as 2^{−(years since conflict/2)}. For observations without previous conflict, the variable is zero.

To account for the diffusion effect of conflict from contiguous districts j of unit i in year t − 1, I included a binary spatial lag. I assigned a value of 0 if no neighbors experienced conflict and 1 if conflict occurred in any contiguous unit in t − 1 (Ward and Gleditsch, 2008).¹²

Other control variables

To account for factors that may affect variations in poverty and conflict, I included a number of covariates. High-population countries are associated with increased risk of armed conflict (Bruckner, 2010), a correlation explained by the existence of large recruitment pools, affluent urban areas, and the strategic importance of locations. Moreover, subnational conflict is more likely to occur in areas with a large population outside the capital city (Raleigh and Hegre, 2009). I therefore included the log of population size in the unit, as calculated from the Gridded Population of the World (CIESIN, 2005).¹³ The district level correlation with mean LPI is weak, as shown in Table 1. This differs from expenditure measures that are highly correlated with population size.

Capital cities are often considered the culmination of state capacity and wealth. First, it is widely acknowledged that living standards in more rural areas in poorer countries lag behind urban areas (for an excellent overview of urban bias in development, see Sahn and Stifel, 2003). Second, the local loss of strength gradient suggests that conflict is only localized near the capital if the rebels are strong vis-à-vis the government. Weaker rebel groups only gain footing on the periphery (Buhaug, 2010). I thus include a variable indicating the (logged) distance from the centroid of each district to the capital city. Similarly, other research suggests that border areas of state peripheries provide safe havens for rebel groups (Salehyan, 2007), increasing the likelihood of conflict there (Buhaug and Rød, 2006). In contrast, Tollefsen and Buhaug (2015) found positive though non-significant support for conflict being located in border areas. Historically, border areas have suffered from marginalization and persistent poverty (Goodhand, 2003). To account for the strategic importance of peripheral locations and their marginalized position, I included the distance from the centroid of each unit to the nearest international border (logged). I also calculated the size of units from polygons in the Global Administrative Areas dataset. Across models, I included dummy variables for each survey round to account for both round- and time-specific heterogeneity.

Modeling strategy

The basic count model for non-negative integer responses is the Poisson regression model. However, the Poisson model assumes equidispersion, where the mean equals the variance. Implicitly, this applies to assumptions that the various counts are independent of each other. If overdispersion goes unaccounted for, standard errors will be biased, increasing the risk of type 1 errors (Hilbe, 2011).

An alternative modeling strategy is the traditional parameterization of the negative binomial (NB henceforth) model (typically denoted NB2), derived from a Poisson-gamma mixture distribution. This includes an extra parameter α to accommodate for overdispersion. A test of the dependent variable confirms the existence of overdispersion.¹⁴ It is also highly likely that one conflict event increases the likelihood of more events, violating the count independence assumption where one event increases the likelihood of another. Thus, the NB model is used in the district-level analyses to account for distinctive features of event counts.

While the NB model accounts for overdispersion in cross-sectional data, the independence assumption is typically violated in longitudinal studies, when units are repeatedly observed over time. Conflict events measured in unit i at time t are likely to be correlated with observations of i at t − 1. The NB model can be extended to accommodate for separate fixed effects (FE henceforth) for each distinct panel in the data. Using FE models on data of units observed over time makes it possible to partial out stable characteristics of these units (Allison, 2009). This provides a better estimate of the effect of Xit on Yit, reducing potential omitted variable bias. Fixed effects negative binomial (FENB henceforth) models may be estimated, either conditionally, as proposed by Hausman et al. (1984), or unconditionally by including a dummy variable for each panel (Hilbe, 2011).

When the number of panels is large, the unconditional FENB model is susceptible to biased estimates owing to the incidental parameters (IP henceforth) problem. Also, it can be computationally inefficient if the number of dummy variables to be estimated is large. Hence, the Hausman et al. (1984) model was proposed to ameliorate the IP problem and to increase computational efficiency. However, as both Allison and Waterman (2002) and Hilbe (2011) highlight, the conditional FENB model is not a true fixed-effects model; it fails to account for time-invariant predictors, by allowing for individual-specific variation in the dispersion parameter, rather than in the conditional mean. Guimaraes (2008) and Greene (2007) reached similar conclusions. Using simulations, Allison and Waterman (2002) did not find clear evidence for the IP problem in the unconditional FENB model. Yet as Hilbe (2011) argued, the unconditional FENB model should be reported with bootstrapped standard errors. Thus, I analyzed the panel data using the unconditional FENB model. Across pooled-cross sectional models 1–7, robust (HC3) standard errors are reported.¹⁵ The Online Appendix (section 16) presents results using standard errors clustered at the country level with no significant differences. For the unconditional FENB models, bootstrapped standard errors are reported to explore the variability in estimates.¹⁶

Results

The district level

Table 2 presents results from the negative binomial regression model using the district pooled cross-sectional data. Model 1 reports baseline estimates of the LPI treatment variable without control variables to reveal any potential post-treatment bias, while Model 2 shows the coefficients of the control variables only. Model 1 suggests a positive and significant relationship with post-survey conflict. For Model 3, including relevant control variables, the results imply a positive relationship between perceived poverty and post-survey conflict, even when controlling for the spatial and temporal proximity of conflict. Conflicts tend to be located in poorer areas of countries. However, the prevalence of poverty is likely to vary between countries. Model 4 introduces a measure of LPI relative to the national average, and shows that when the gap widens between a district’s poverty levels and the national average, conflict becomes more likely. This indicates that conflict violence is more likely to be located in relatively poorer areas of countries. Model 5 employs an alternative measure of the share of respondents in a district living without fulfillment of basic needs.¹⁷ The results show that as the share of people with unmet basic needs increases, the risk of conflict increases. For comparison, Model 6 suggests that the risk decreases significantly when the majority of a population has never experienced unmet basic needs.

Table 2.

Negative binomial regression at district level using robust standard errors

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Experienced Poverty	0.809^*** (0.179)		1.365^***(0.207)				0.647***(0.175)
Relative Experienced Poverty				1.449^***(0.348)
High Experienced Poverty					2.323^***(0.619)
Low Experienced Poverty						−2.292^***(0.617)
Conflict History		5.805^***	4.775^***	4.716^***	4.862^***	4.859^***
		(0.704)	(0.736)	(0.753)	(0.797)	(0.803)
Conflict Spatial Lag (t−1)		0.850	1.414^***	1.418^***	1.566^***	1.561^***
		(1.271)	(0.286)	(0.281)	(0.300)	(0.301)
Population (ln)		0.427	0.354^***	0.457^***	0.382^***	0.385^***
		(0.362)	(0.120)	(0.148)	(0.127)	(0.127)
Distance to Capital (ln)		0.013	0.539^*	0.448	0.530	0.533
		(0.227)	(0.319)	(0.379)	(0.350)	(0.353)
Distance to Border (ln)		0.172	0.208^**	0.059	0.095	0.090
		(0.293)	(0.092)	(0.088)	(0.094)	(0.094)
Area in square km (ln)		−0.181	−0.209^**	−0.143	−0.168^*	−0.168^*
		(0.161)	(0.092)	(0.110)	(0.101)	(0.102)
Round 4 Dummy	0.028	−0.276	0.0004	−0.195	−0.071	−0.071
	(0.235)	(0.398)	(0.345)	(0.353)	(0.361)	(0.363)
Round 5 Dummy	0.265	−0.518	0.112	−0.088	0.088	0.093
	(0.289)	(0.548)	(0.365)	(0.377)	(0.404)	(0.408)
BIC	2850.096	2723.397	2667.549	2683.628	2680.019	2680.721	3433.413
Observations	4008	3952	3952	3952	3952	3952	1548
Theta	0.029^*** (0.002)	0.042^***(0.004)	0.050^***(0.005)	0.048^***(0.005)	0.049^***(0.005)	0.049^***(0.005)	29,770.540

Figure 3 visualizes the poverty term from Model 2.¹⁸ The solid line represents the expected number of conflict events across values of experienced poverty, with dotted lines showing 95% confidence intervals. Across panels, covariates are set to their medians. The left panel represents a district without previous or neighboring conflict, while in the center panel, conflict history is set to 5 years, and in the right panel, to two years. Across the three settings, poverty is associated with increased risk of conflict. Also, shorter timespans since past conflict fortify this effect. The plots also indicate that affluent districts have a negligible risk of conflict, regardless of conflict history.

Figure 3.

Expected post-survey conflict and experienced poverty across three settings: no previous conflict (left), conflict 5 years ago (center) and conflict 2 years prior (right).

In substantive terms, the visualizations show that a district has an expected conflict count of 0.02 if its citizens’ basic needs never go unmet and if there is no history of conflict or conflict in neighboring districts. When basic needs always go unmet, however, the expected count increases to 3.8. Furthermore, an impoverished district where basic needs are always lacking, with conflict 5 years before the survey, has an expected conflict count of 6.8. Still, an area that has no experienced poverty (0) although it had conflict 5 years prior still has an almost negligible expected count of 0.05. Hence, richer districts appear more immune to repeat conflict, even with a recent history of armed conflict. The results hold when replacing the negative binomial model with a standard ordinary least squares estimator (see the Online Appendix section 9).

While the results presented so far suggest that poverty increases the risk of conflict violence, the possibility that the result is endogenous—where poverty is not a cause, but rather a result of conflict—cannot be ruled out. Complete causal identification would require an exogenous instrument of poverty. In the absence of such an instrument, matching is used to mimic a randomized experiment where treated and control groups are similar on all observed and unobserved background characteristics (Stuart, 2010). I use coarsened exact matching to match districts that are as similar as possible on a series of observed background covariates (Iacus et al., 2009)—most importantly, similar levels of past conflict.¹⁹

Next, I analyzed the matched dataset using regression to explore variations in poverty and conflict violence within matched subclasses of districts. While critics warn it does not solve the ignorability assumption (Miller, 2015), matching does improve control of comparison groups and ensures as much similarity as possible between treatment and control groups (Ho et al., 2007), reducing concerns of omitted variable bias. Model 7 introduces a dichotomized version of the experienced poverty variable, where districts above mean LPI are coded 1 (treated) and those below mean LPI are coded 0 (untreated).²⁰ The result provides considerable support for a positive effect of poverty on the intensity of conflict violence when limiting the analysis to only districts that are comparable on average.

Additional tests are provided in the Online Appendix, showing a highly robust relationship. In particular, results do not change when excluding most violent districts (see Online Appendix section 11). Addressing another concern that districts with few respondents reduce internal representativeness, I also show how results do not change in alternative models where districts with a low number of respondents are excluded (see Online Appendix section 12). Last, a hurdle regression is employed, showing that poverty is more closely related to conflict intensity than conflict onset (see Online Appendix section 10).

Region-level robustness test

To further test the robustness of the estimates, I employed the unconditional FENB model, as described in the ‘Modeling strategy’ section, to make use of variation within repeated observations of the same regions across time. Model 8 in Table 3 confirms the previous findings at the district level, suggesting that poverty increases the risk of local conflict. The results hold across all three models, revealing a robust relationship when including unit dummies to account for fixed effects.²¹

Table 3.

Fixed effects negative binomial models, region level

	Dependent variable
	Number of conflict events
	(8)	(9)	(10)
Experienced Poverty	0.922^**	0.922^**	0.920^**
	(0.389)	(0.400)	(0.424)
Conflict History		−0.389	−0.440
		(0.552)	(0.613)
Conflict Spatial Lag (t−1)			0.064(0.305)
Round 4 Dummy	−0.038	0.002	−0.002
	(0.238)	(0.257)	(0.250)
Round 5 Dummy	−0.259	−0.171	−0.167
	(0.261)	(0.305)	(0.307)
BIC	1962.822	1967.757	1973.44
Observations	311	311	311
Overdispersion (θ)	0.918^*** (0.110)	0.919^*** (0.110)	0.920^*** (0.110)

Bootstrapped standard errors in parentheses (1000 replications): ^*p < 0.1; ^**p < 0.05; ^***p < 0.01.

A conditional relationship?

While the results presented so far point to a positive relationship between poverty and conflict, the literature has described the relationship as indirect. Exploring the possibility of a conditional relationship, I used interaction models to test whether poverty is conditional of H2: local income opportunities; H3: local state capacity; and H4: local group grievances.

Figure 4 shows the results of three interaction models where each moderator variable was interacted with levels of experienced poverty.²² In the left panel, Model 10²³ illustrates how levels of unemployment do not seem to moderate the relationship between poverty and conflict. In fact, there is no significant difference in the conditional effect of levels of unemployment on the poverty–conflict nexus. Contrary to the opportunity cost explanation, the results suggest that impoverished areas with high unemployment are no more likely to experience conflict than poor areas where the majority of residents are employed. Other research might shed light on this null finding. In the micro-level study by Berman et al. (2011), survey data from Iraq, Afghanistan, and the Philippines revealed that no areas with high unemployment rates were more likely to experience conflict. The scholars attributed the null finding to reduced cost of counterinsurgency information in areas with high unemployment rates, which could provide cheap information for government intelligence agencies. In addition, areas with high conflict propensity might see increased security measures, reducing conflict propensity by simultaneously hurting the local economy and cutting employment opportunities for citizens. Thus, the results presented here align closely with those of Berman et al. (2011). Another explanation for the null finding is that poorer areas with high unemployment rates see less conflict because greed-motivated rebel groups view these regions as having less strategic value.

Figure 4.

Expected number of conflict events when experienced poverty is interacted with unemployment (left panel); institutional quality (center panel); and share of population with group grievances (right panel).

In keeping with hypothesis 3, Model 11 (depicted in Figure 4’s center panel) indicates that high-quality local institutions have a strong pacifying effect on the poverty–conflict relationship. Poverty is associated with more conflict when local institutions are considered weak, while areas with high-quality institutions seem to mitigate poverty’s effect on the risk of conflict—hence the result that strong institutions are capable of ameliorating the effects of poverty on conflict. These results provide considerable support for the findings by Wig and Tollefsen (2016) of the powerful pacifying role played by high-quality local institutions.

Finally, Model 12 (Figure 4’s right panel) shows that conflict is more likely in areas with a large impoverished population which perceives their group as being unfairly treated. This is in keeping with H4. However, districts with a large impoverished population holding no grievances against the government have an insignificant risk of conflict. Thus, being poor will not necessarily increase inclination to rebel as long as most individuals deem their government to be fair. However, when poor individuals experience collective grievances and unfair treatment by incumbents, motivation to instigate violence increases. An additional operationalization of grievances is available in the Online Appendix (section 8), where the perceptual measure is replaced with a variable indicating if the district is located in the territorial area of a politically excluded group; the results do not change. This may be due to the very low number of districts affected by political discrimination. Moreover, grievances are also tied to factors besides political exclusion. They may be better captured using a subjective perceptual measure of sense of injustice, such as unfair treatment, lack of rights, economic and social marginalization, and other aspects not captured by political exclusion per se.

Limitations

Although survey data significantly enriches information availability at the subnational level, it also has weaknesses. While Afrobarometer employs regional stratification and random sampling within selected regions to avoid undersampling particular regions or groups, the survey is not entirely representative at the subnational level. While the internal validity of the institutional quality is considered to be quite good,²⁴ interpretations of the results must take this limitation into account. As a robustness test, districts with few respondents were excluded.²⁵ As the Afrobarometer covers 35 of the 54 countries in Africa, we need to be cautious about external generalization outside of the countries included in the study.

To reduce the potential that districts with extreme numbers of conflict events were driving the result, most violent districts were excluded; still, the results remained. I also employed a hurdle model to explore whether an empirical difference existed in the relationships between poverty and conflict onset and poverty and conflict intensity; the finding was that while poverty is related to conflict onset, it seems to be more closely linked to its intensity (see Online Appendix sections 10 and 11).

Perceptions of grievances are an appropriate way of measuring dissent among local populations. However, using perceptions to measure institutional quality might introduce biases. For instance, in settings where institutions perform well, higher expectations might lead to measurement bias. Perceptions of institutions might depend on group belonging dynamics, where members of certain communities have systematic negative or positive biases in their perceptions of government institutions. Nevertheless, perception-based indicators of local institutional quality are likely to reflect actual quality and, to date, present the best subnational indicator of the quality of local institutions in developing countries.

Still, another shortcoming is the lack of an exogenous instrument to identify the causal effect of poverty on conflict. This study employed a broad range of tests, including matching and a fixed-effects design to reduce endogeneity bias without altering the results. The models also tested the relationship using two different units of analysis, inspiring confidence that the results were not susceptible to a modifiable areal unit problem that could arise if results are sensitive to scaling issues (Fotheringham and Wong, 1991).

Conclusion

Past research on the local link between poverty and conflict has produced mixed results. These results have stemmed mainly from proxies of poverty that are either too aggregated or only loosely connected to the rationale for conflict. To overcome these limitations, this study employed a novel measure of subjective poverty: the fulfillment of people’s basic needs. I showed how this measure is more closely tied to what incentivizes violence and drives people to take up arms against their governments.

The results show that subnational areas with high levels of poverty face an increased risk of conflict violence. Areas where respondents have access to food, water, medicine or medical aid, fuel, and a cash income are more likely to experience conflict. Reducing endogeneity concerns, matching and fixed-effect models provided considerable support for a link between poverty and conflict.

The results presented here are steps towards resolving previously inconclusive findings. This study, moreover, took a stride by testing conditional mechanisms of the poverty–conflict nexus at the local level. To my knowledge, no previous study has employed interaction models with local-level contextual data to explore in which settings poverty may cause conflict.

Interaction models show that local unemployment rates do not condition the poverty–conflict nexus, contrary to the individual opportunity cost explanation. Poorer areas with widespread unemployment do not seem to have a higher risk of local conflict more than those with ample employment opportunities. Conflict, meanwhile, is less likely in impoverished areas with strong, high-quality local institutions. However, impoverished areas governed by weak, low-quality local institutions have greater conflict propensity. The results suggest that high-quality institutions have a strong pacifying effect. Also, group grievances have a conditioning effect on the poverty–conflict nexus; impoverished areas where the majority population is aggrieved and perceive their group as being unfairly treated by the government have a greater risk of conflict than poor regions with a contented population.

The results presented here highlight the crucial role of institutions and sense of inclusion in reducing the effect of poverty on conflict. Strong institutions can deter local challengers in their ability to resolve the roots of group grievances and hostilities between groups and to meet basic needs across population groups. Furthermore, ethnic inclusivity reduces the effect of poverty on the risk of local conflict violence. Thus, local conflict violence is a function of both opportunity and motivation, based on explanations of conflict: poverty presents rebels with recruitment opportunities, particularly in the absence of strong, high-quality local institutions. The results also point to motivation as an important factor in elevating the risk of conflict. While this study does not pit the two mechanisms against each other, the results indicate that they act complementarily at the local level when it comes to increasing the risk of local conflict violence.

Although matching and fixed-effect models might ameliorate concerns about it, endogeneity cannot be fully ruled out. To do so would warrant an instrument variable approach. However, identifying an entirely exogenous instrument of local poverty is complex. Still, the models presented here provide second-best alternatives for inferring from observational data. Moreover, internal and external validity and possible biases introduced with perceptual data might affect the robustness of the results. Using subjective indicators of poverty and local context-specific characteristics offers an innovative way to study the subnational relationship between poverty and conflict, as well as the conditioning effect of local socioeconomic, political, and ethnic characteristics in developing countries with limited data availability.

This article outlined a subnational framework on how to study the relationship between local socioeconomic conditions in target locations using geocoded survey data. Recent datasets, such as the ACD2EPR (Wucherpfennig et al., 2012), provide information about ties between rebel groups and ethnic groups. By linking this to ethnic settlement data, such as the SIDE (von Schweinitz and Hunziker, 2018), and poverty measures from the Afrobarometer, we can analyze poverty levels within rebel groups’ ethnic homelands. Future micro-level studies could benefit from these comprehensive data sources to explore the ethnic background of rebel groups and where these rebel groups belong. This presents an opportunity to explore poverty in rebel group support locations vs their target areas, and in so doing, to distinguish whether richer groups target poorer areas and/or poorer groups target richer areas.

Footnotes

Acknowledgements

I would like to thank the editor and three anonymous reviewers, as well as the participants of the 14th Jan Tinbergen European Peace Science Conference in The Hague for their comments on earlier drafts.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Notes

References

Afrobarometer (2016) Sampling principles. Available at: http://www.afrobarometer.org/surveys-and-methods/sampling-principles (accessed 12 September 2017).

Ahrens

(2015) Civil conflicts, economic shocks and night-time lights. Peace Economics, Peace Science and Public Policy21(4): 433–444.

Allison

(2009) Fixed Effects Regression Models, Vol. 160. London: SAGE.

Allison

Waterman

(2002) Fixed-effects negative binomial regression models. Sociological Methodology32(1): 247–265.

Barrett

(2011). Interviews with killers: Six types of combatants and their motivations for joining deadly groups. Studies in Conflict & Terrorism34(10): 749–764.

Beardsley

Gleditsch

(2015) Roving bandits? The geographical evolution of African armed conflicts. International Studies Quarterly59(3): 503–516.

Benson

Kugler

(1998) Power parity, democracy, and the severity of internal violence. Journal of Conflict Resolution42(2): 196–209.

Berman

Couttenier

(2015) External shocks, internal shots: The geography of civil conflicts. Review of Economics and Statistics97(4): 758–776.

Berman

Callen

Felter

Shapiro

(2011) Do working men rebel? Insurgency and unemployment in Afghanistan, Iraq, and the Philippines. Journal of Conflict Resolution55(4): 496–528.

10.

Braithwaite

Dasandi

Hudson

(2016) Does poverty cause conflict? Isolating the causal origins of the conflict trap. Conflict Management and Peace Science33(1): 45–66.

11.

Bruckner

(2010) Population size and civil conflict risk: Is there a causal link?The Economic Journal120(544): 535–550.

12.

Buhaug

(2006) Relative capability and rebel objective in civil war. Journal of Peace Research43(6): 691–708.

13.

Buhaug

(2010) Dude, where’s my conflict? LSG, relative strength, and the location of civil war. Conflict Management and Peace Science27(2): 107–128.

14.

Buhaug

Lujala

(2005) Accounting for scale: Measuring geography in quantitative studies of civil war. Political Geography24(4): 399–418.

15.

Buhaug

Rød

(2006) Local determinants of African civil wars, 1970–2001. Political Geography25(3): 315–335.

16.

Buhaug

Gleditsch

Holtermann

Tollefsen

Østby

(2011) It’s the local economy, stupid! Geographic wealth dispersion and conflict outbreak location. Journal of Conflict Resolution55(5): 814–840.

17.

Cederman

Weidmann

Gleditsch

(2011) Horizontal inequalities and ethnonationalist civil war: A global comparison. American Political Science Review105(3): 478–495.

18.

Cederman

Gleditsch

Buhaug

(2013) Inequality, Grievances and Civil War. Cambridge: Cambridge University Press.

19.

Cederman

Weidmann

Bormann

(2015) Triangulating horizontal inequality toward improved conflict analysis. Journal of Peace Research52(6): 806–821.

20.

CIESIN (2005). Gridded population of the world, version 3 (gpwv3): Population count grid, future estimates. Technical report, NASA Socioeconomic Data and Applications Center (SEDAC).

21.

Collier

Hoeffler

(1998) On the economic causes of civil war. Oxford Economic Papers50(4): 563–573.

22.

Collier

Hoeffler

(2004) Greed and grievance in civil war. Oxford Economic Papers56(4): 563–595.

23.

Collier

Hoeffler

Rohner

(2009) Beyond greed and grievance: Feasibility and civil war. Oxford Economic Papers61: 1–27.

24.

Davies

(1962) Towards a theory of revolution. American Sociological Review27(1): 5–19.

25.

Deininger

(2003) Causes and consequences of civil strife: Micro-level evidence from Uganda. Oxford Economic Papers55(4): 579–606.

26.

de Soysa

Gleditsch

(1999) To Cultivate Peace: Agriculture in a World of Conflict. Environmental Change and Security Project Report Issue 5.

27.

Dulani

Mattes

Logan

(2013) After a decade of growth in Africa, little change in poverty at the grassroots. Afrobarometer Policy Brief1.

28.

Elbers

Lanjouw

(2003) Micro-level estimation of poverty and inequality. Econometrica71(1): 355–364.

29.

Fearon

Laitin

(2003) Ethnicity, insurgency, and civil war. American Political Science Review97(1): 75–90.

30.

Filmer

Scott

(2012) Assessing asset indices. Demography49(1): 359–392.

31.

Fotheringham

Wong

(1991) The modifiable areal unit problem in multivariate statistical analysis. Environment and Planning A23(7): 1025–1044.

32.

Friedman

(2013) Local economic conditions and participation in the Rwandan genocide. HiCN Working Papers 160, Households in Conflict Network.

33.

Gleditsch

Wallensteen

Eriksson

Sollenberg

Strand

(2002) Armed conflict 1946–2001: A new dataset. Journal of Peace Research39(5): 615–637.

34.

Goodhand

(2003) Enduring disorder and persistent poverty: A review of the linkages between war and chronic poverty. World Development31(3): 629-646.

35.

Greene

(2007) Functional Form and Heterogeneity in Models for Count Data. Breda: Now Publishers.

36.

Grusky

Kanbur

(2006) Introduction: The conceptual foundations of poverty and inequality measurement. In: Grusky

Kanbur

(eds), Poverty and Inequality. Stanford, CA: Stanford University Press.

37.

Guimaraes

(2008) The fixed effects negative binomial model revisited. Economics Letters99(1): 63–66.

38.

Gurr

(1970) Why Men Rebel. Princeton, NJ: Princeton University Press.

39.

Gurr

(1993) Minorities at Risk: A Global View of Ethnopolitical Conflicts. Washington, DC: USIP Press Books.

40.

Hallberg

(2012) PRIO conflict site 1989–2008: A geo-referenced dataset on armed conflict. Conflict Management and Peace Science29(2): 219–232.

41.

Hausman

Hall

Griliches

(1984) Econometric models for count data with an application to the patents—R & D relationship. Econometrica52(4): 909–938.

42.

Hegre

(2016) Civil conflict and development. In: van de Walle

Lancaster

(eds), Handbook on the Politics of Development. Oxford: Oxford University Press.

43.

Hegre

Sambanis

(2006) Sensitivity analysis of empirical results on civil war onset. Journal of Conflict Resolution50(4): 508–535.

44.

Hegre

Østby

Raleigh

(2009) Poverty and civil war events: A disaggregated study of Liberia. Journal of Conflict Resolution53(4): 598–623.

45.

Hilbe

(2011) Negative Binomial Regression. Cambridge: Cambridge University Press.

46.

Imai

King

Stuart

(2007) Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference. Political Analysis15(3): 199–236.

47.

Hobsbawm

(1959) Primitive Rebels. Manchester: Manchester University Press.

48.

Holtermann

(2012) Explaining the development–civil war relationship. Conflict Management and Peace Science29(1): 56–78.

49.

Horowitz

(1985) Ethnic Groups in Conflict. Berkeley, CA: University of California Press.

50.

Howe

Hargreaves

Gabrysch

Huttly

(2009) Is the wealth index a proxy for consumption expenditure? A systematic review. Journal of Epidemiology and Community health63(11): 871–877.

51.

Humphreys

Weinstein

(2008) Who fights? The determinants of participation in civil war. American Journal of Political Science52(2): 436–455.

52.

Iacus

King

Porro

(2009) CEM: Software for coarsened exact matching. Journal of Statistical Software30(13): 1–27.

53.

Ikejiaku

(2012) Poverty-–conflict nexus: The contentious issue revisited. European Journal of Sustainable Development1(2): 127–150.

54.

International Fund for Agricultural Development (2001) Rural Poverty Report 2001: The Challenge of Ending Rural Poverty. Oxford: Oxford University Press.

55.

Irobi

(2005) Ethnic conflict management in Africa: A comparative case study of Nigeria and South Africa. Development Policy Management Network Bulletin13(3): 1–23.

56.

Jakobsen

De Soysa

Jakobsen

(2013) Why do poor countries suffer costly conflict? Unpacking per capita income and the onset of civil war. Conflict Management and Peace Science30(2): 140–160.

57.

Jean

Burke

Xie

Davis

Lobell

Ermon

(2016) Combining satellite imagery and machine learning to predict poverty. Science353(6301): 790–794.

58.

Jerven

(2013) Poor Numbers: How we are Misled by African Development Statistics and What to do About it. Ithaca. NY: Cornell University Press.

59.

Justino

(2009) Poverty and violent conflict: A micro-level perspective on the causes and duration of warfare. Journal of Peace Research46: 315–333.

60.

Justino

Bruck

Verwimp

(2013) A Micro-level Perspective on the Dynamics of Conflict, Violence, and Development. Oxford: Oxford University Press.

61.

Kalyvas

(1999) Wanton and senseless? The logic of massacres in algeria. Rationality and Society11(3): 243–285.

62.

Kilcullen

(2011) The Accidental Guerrilla: Fighting Small Wars in the Midst of a Big One. Oxford: Oxford University Press.

63.

Knutsen

Kotsadam

Olsen

Wig

(2017) Mining and local corruption in Africa. American Journal of Political Science61(2): 320–334.

64.

Miller

(2015) The uses and abuses of matching in political science. Working Paper. George Washington University.

65.

Nillesen

Verwimp

(2009). Grievance, commodity prices and rainfall: A village level analysis of rebel recruitment in Burundi. HiCN Working Papers 58, Households in Conflict Network.

66.

NOAA (2014) DMSP-OLS nighttime lights time series, version 4. Available at: https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html (accessed 20 February 2017).

67.

Nordhaus

(2006) Geography and macroeconomics: New data and new findings. Proceedings of the National Academy of Sciences of the United States of America103(10): 3510–3517.

68.

Nordhaus

Azam

Corderi

Hood

Victor

Mohammed

Miltner

Weiss

(2006) The G-ECON Database on Gridded Output: Methods and Data. New Haven, CT: Yale University.

69.

Østby

(2008) Polarization, horizontal inequalities and violent civil conflict. Journal of Peace Research45(2): 143–162.

70.

Østby

Nordås

Rød

(2009) Regional inequalities and civil conflict in sub-Saharan Africa, 1986–2004. International Studies Quarterly53(2): 301–324.

71.

Posel

Rogan

(2016) Measured as poor versus feeling poor: Comparing money-metric and subjective poverty rates in South Africa. Journal of Human Development and Capabilities17(1): 55–73.

72.

Pradhan

Ravallion

(2000) Measuring poverty using qualitative perceptions of consumption adequacy. The Review of Economics and Statistics82(3): 462–471.

73.

Raleigh

Hegre

(2009) Population, size, and civil war. A geographically disaggregated analysis. Political Geography28(4): 224–238.

74.

Ravallion

Lokshin

(2001) Identifying welfare effects from subjective questions. Economica68(271): 335–357.

75.

Rustad

Buhaug

Falch

Gates

(2011) All conflict is local: Modeling subnational variation in civil conflict risk. Conflict Management and Peace Science28(1): 15–40.

76.

Sahn

Stifel

(2003) Urban–rural inequality in living standards in Africa. Journal of African Economies12(4): 564-597.

77.

Salehyan

(2007) Transnational rebels—Neighboring states as sanctuary for rebel groups. World Politics59: 217–242.

78.

Sen

(1999) Development as Freedom. New York: Anchor Books.

79.

Stewart

(2002) Root causes of violent conflict in developing countries. British Medical Journal324(7333): 342–345.

80.

Storeygard

AD.

Levy

Deane

(2008) The global distribution of infant mortality: A subnational spatial view. Population, Space and Place14(3): 209–229.

81.

Stuart

(2010) Matching methods for causal inference: A review and a look forward. Statistical Science: A Review Journal of the Institute of Mathematical Statistics25(1): 1–21.

82.

Sundberg

Melander

(2013) Introducing the UCDP georeferenced event dataset. Journal of Peace Research50(4): 523–532.

83.

Tollefsen

Buhaug

(2015) Insurgency and inaccessibility. International Studies Review17(1): 6–25.

84.

Tollefsen

Strand

Buhaug

(2012) PRIO-GRID: A unified spatial data structure. Journal of Peace Research49(2): 363–374.

85.

von Schweinitz

Hunziker

(2018) New spatial data on ethnicity: Introducing SIDE. Journal of Peace Research, in press.

86.

Ward

Gleditsch

(2008) Spatial Regression Models. Thousand Oaks, CA: Sage.

87.

Weidmann

Gleditsch

(2010) Mapping and measuring country shapes. The R Journal2(1): 18–24.

88.

Weidmann

Schutte

(2017) Using night light emissions for the prediction of local wealth. Journal of Peace Research, 54(2): 125–140.

89.

Wig

Tollefsen

(2016) Local institutional quality and conflict violence in Africa. Political Geography53: 30–42.

90.

Wucherpfennig

Metternich

Cederman

Gleditsch

(2012) Ethnicity, the state, and the duration of civil war. World Politics64(1): 79–115.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

98.64 MB

0.00 MB

5.33 MB