Abstract
The vast majority of civil wars occur in economically less developed countries, as measured by GDP per capita. Two suggested explanations for this are prominent: one emphasizing that poverty facilitates rebel recruitment due to lowered economic opportunity cost of rebelling, and the other highlighting that low state reach and capacity give political and military opportunity for organizing insurgency. I argue that the latter account is more powerful. Low state reach is vital not only to rebel survival; it also enables rebels to obtain control over remote settlements, which facilitates the effective use of persuasion, coercion, organization, and economic rewards for mobilizing recruits and other resources. Although low economic opportunity costs can ease recruitment, it may not be essential if such tools are available. The argument is supported by a quantitative analysis covering 133 countries from 1989 to 2006. Countries experiencing civil war were distinguished more by low state reach (measured by road density, telephone density, and % urban of the population) than by depth of poverty (measured by the mean income of the poorest decile). Moreover, the negative association between GDP per capita and civil war risk disappeared when controlling for state reach, but remained strong controlling for poverty.
Introduction
Civil war is predominantly a “problem of the poor”, as Sambanis (2002: 216) has put it. While it has become exceedingly rare in upper middle- and high income countries, civil war still is prevalent in low-income countries. 2 Quantitative studies also find GDP per capita to be one of the strongest cross-national predictors of civil war (e.g. Hegre and Sambanis, 2006). Although this empirical association is undisputed, its underlying causes are highly contested. Two explanations are particularly prominent: the first emphasizes that poverty gives a lower economic opportunity cost of rebelling, thus making rebellion financially viable (Collier and Hoeffler, 2004). The second stresses that low state reach and capacity impede counterinsurgency and facilitate insurgent mobilization in peripheral areas (Fearon and Laitin, 2003). Little has been done so far to assess and compare the explanatory power of these accounts.
Drawing on case study research, I argue that the state reach account, which stresses politico-military opportunity, is more powerful than the economic opportunity explanation. Importantly, the latter does not take into account that the reasons for contributing resources to insurgents derive in large part from the process of insurgency, and that rebel organizations—if conditions allow—can manipulate not only economic incentives, but also security incentives and emotional and moral impetuses for participation. Although low economic opportunity costs can ease recruitment, it may not be essential if insurgents are able to gain control over remote settlements, which facilitates the effective use of coercion, organization, persuasion, and economic rewards to mobilize recruits and other resources. The opportunity for rebels to do this depends mainly on state reach and capacity, not poverty. Moreover, beyond a certain level of state reach and capacity, there is hardly any military opportunity for insurgency. 3
I perform a quantitative analysis of the two accounts using data for 133 countries in the 1989–2006 period. This is made possible by replacing the theoretically diffuse GDP per capita measure with indicators more closely linked to the theories’ core implications. The economic opportunity cost account points to poverty—and particularly the income of the poorest, who have the least to lose—as the main variable underlying the GDP–civil war association. Accordingly, I use the mean income of the poorest decile of the population based on survey data as my proxy for economic opportunity costs. The politico-military opportunity account, on the other hand, suggests that the GDP–civil war association should disappear when controlling for the reach of the state apparatus. I measure state reach by an index including road density, telephone density, and the percentage of the population living in urban areas. The results show that countries experiencing civil war were distinguished particularly by their low state reach, and not so much by their depth of poverty. Moreover, the negative association between GDP per capita and civil war risk disappeared when controlling for state reach, but remained strong when controlling for poverty. This indicates that the concentration of civil war in poorer countries owes more to political and military opportunity for organizing insurgency than to low economic opportunity cost of rebelling.
The article proceeds as follows. First, I discuss the two accounts’ suggested mechanisms and their links to structural conditions, drawing on case study based knowledge. Next, I discuss existing cross-national evidence before I derive hypotheses and identify indicators that help empirically distinguish between the two theories. I then explain operationalizations and describe the data and estimation of missing values. The empirical analysis begins with an exploration focusing on the civil war cases before turning to regression analysis to test the hypotheses. I conclude by summing up and discussing how the evidence may be interpreted.
Theory
Economic Opportunity
The idea that low-income countries are prone to civil war because the poor have low economic opportunity cost of rebelling is often invoked (e.g. Sambanis, 2002). It is aptly formulated by Paul Collier and colleagues (Collier, 2000b; Collier and Hoeffler, 2004; Collier et al., 2009). 4 They argue that a key step in explaining civil war is to account for the formation of a rebel army large enough to put up substantial resistance against government forces. Also, they point out that most nascent rebel groups have strong financial constraints, and therefore, the ability to form an army may hinge upon the salary levels demanded by potential fighters. Poverty may thus facilitate rebellion because it makes rebel labor cheap: “[r]ecruits must be paid, and their cost may be related to the income foregone by enlisting as a rebel” (Collier and Hoeffler, 2004: 659). 5
Assuming for now that this economic recruitment mechanism is pervasive, is it true that poverty facilitates insurgency? A first challenge is that lower income levels should reduce labor costs not only for the rebel organization, but also for the government, which is competing with the rebels over skilled soldiers. However, as Collier (2000b) argues, the rebel organization is likely to be more sensitive to labor costs than the government. Not only does the government have greater financial resources; it may also conscript soldiers. Another objection is that poor societies may yield fewer assets for the rebels to loot or “tax” in a more regularized way. Although this is true, existing assets in low-income countries may be easier to tax for any local ruler—rebel or government—because capital mobility is low (Boix, 2008; Fearon, 2008). Typically, the bulk of capital lies in land, which, in contrast to financial assets or human skills, cannot easily be moved in response to tax increases.
Returning to the core mechanism, how important do economic opportunity costs seem to be for the choice of rebelling? One rough indicator is whether insurgents tend to come from poorer backgrounds than noncombatants. Surveys allowing for comparison of insurgents and noncombatants exist for too few cases to conclude on the question. However, probably the only study using such survey data finds that rebel combatants as well as government combatants in Sierra Leone were on average somewhat poorer than noncombatants (Humphreys and Weinstein, 2008). 6 This fits with the economic opportunity cost argument.
However, in civil war, economic concerns are often trumped by other concerns, as the rich case study literature suggests. In several cases, such as Zimbabwe (Kriger, 1992), Sierra Leone (Humphreys and Weinstein, 2008), and Guatemala (Stoll, 1993), many rebel participants emphasize security reasons for participating, due to threats by the government, the rebels, or both. In some cases, such as El Salvador, many highlight emotional or moral engagement (Viterna, 2006: 20–21; Wood, 2003: 231). Other reasons, like fun or adventure (Arjona and Kalyvas, 2006) or status (Kriger, 1992: 115) are also found. Surely, economic motivations figure in some rebel accounts. Arjona and Kalyvas (2006) report that one-fifth of Colombian guerillas said they joined due to the promise of material goods. Still, the overall tendency seems to be that most rebels emphasize other motivations than economic gain. 7
Collier (2000a) argues for skepticism towards rebels’ statements, since talk is cheap and rebels have incentives to construct a narrative of grievance rather than greed. Not only do rebels’ statements speak to the importance of non-economic motivations, however; so do the mobilization practices of rebel organizations. Why would so many rebel organizations prioritize methods like persuasion, coercion, organization, and network building if political beliefs and preferences, social incentives, and security incentives were not important for participation? 8 This points to a key limitation of the economic opportunity cost account: its ignorance of the politics of resource mobilization. Despite the tempting analogy, rebel organizations are not like firms recruiting employees in the market; they can use tools to shape people’s beliefs about the context of their choice, as well as mold the actual context itself, thereby affecting people’s perceived incentives in terms of profit, security, or social esteem. The availability of these tools, essential for mobilizing all types of resources, depends less on poverty than on state reach and capacity.
Politico-Military Opportunity
The other prominent account of the poverty–civil war association emphasizes that low state reach and capacity create political opportunity for insurgency. This theory is more process-oriented: it explains the poverty–civil war association mainly by the opportunity for dissident organizations to effectively apply mobilizational tools and survive government repression, rather than by any pre-war characteristic of the population, like poverty. In Fearon and Laitin’s (2003: 76) words, “[w]here states are relatively weak and capricious, both fears and opportunities encourage the rise of would-be rulers who supply a rough local justice while arrogating the power to ‘tax’ for themselves and, often, for a larger cause.” They emphasize how rebel organizations—under the right circumstances—can mobilize resources by shaping the beliefs and incentives of local populations, and that weak state governments often apply inept and indiscriminate counterinsurgent measures that further fuel the insurgency.
The typical features of poorer countries giving politico-military opportunity for insurgency are not only weak military and police capacity (Herbst, 2004), but more generally a low penetration of state authority throughout the polity—a lack of infrastructural power, to use Mann’s (1993) term. Shortage of roads and communication infrastructure and a scattered, village-based settlement pattern make information as well as state control more costly in low-income countries. This gives opportunity for building up an armed oppositional organization in the periphery, even where the state is militarily superior (c.f. Buhaug, 2010). 9 Without knowing where the rebels are, the state cannot make full use of its initial military advantage (Kalyvas, 2006: 174; Leites and Wolf, 1970: 33). This is amply shown in the counterinsurgency literature, which points to information-gathering abilities as a key determinant of counterinsurgent effectiveness (Lyall and Wilson, 2009; Lyall, 2010).
The infrastructural weakness of the state is not only important for rebel survival; it also provides a chance for rebels to establish military and political control over peripheral settlements, which gives them powerful tools for mobilizing resources. An organization’s use of coercion, provision of collective goods, organization, and indoctrination all become more effective in the context of control.
First, certain threats are only credible if the threatening organization has more local power than its rival. This is perhaps particularly important for obtaining information and hindering information flowing to the rival. Informing the locally weaker organization involves a great risk of sanction (Kalyvas, 2006). Moreover, coercion may be cost effective for mobilizing resources that require virtually universal compliance, like control of information: it is cheaper to punish a few defectors than to pay a majority of compliers (Oliver, 1980: 1363–64). Control also improves insurgents’ opportunities for rewarding cooperation. In particular, it allows insurgents to provide a variety of collective goods in return for the local population’s cooperation, including protection against government violence (Mason, 2004: 168), material benefits like collective land (Mason, 1998: 222), or the breaking of peasant chains to powerful landowners (Popkin, 1988: 10). As Skocpol (1982: 366) argues, “[i]t is hard to imagine the successful institutionalization of such social exchange between peasants and revolutionaries except in places and times unusually free from counterrevolutionary state repression.” Finally, control facilitates organization and persuasion by making it possible to enlist the local population in insurgent-affiliated organizations, like youth groups, farmer cooperatives, unions, and militias (Tse-Tung, 1967: 105–8). These organizations provide arenas for indoctrination, encourage part-time labor efforts, and enhance intelligence collection. Moreover, they may facilitate collective action by placing insurgent leaders in the center of networks and create stronger ties through more regular face-to-face interactions (Petersen, 2001: 61–75).
In sum, the politico-military opportunity account is able to capture the multiple and varied motivations for contributing to insurgency and how these motivations can be shaped by the belligerent organizations. Although it is only a partial explanation of any civil war, this account plausibly provides core insights into why poor countries are more prone to civil war than wealthier countries. I next assess the degree to which the two accounts find support in cross-national quantitative evidence.
Empirical Analysis
Existing Quantitative Evidence
The two most cited cross-national studies on the causes of civil war offer little evidence to support their interpretation of GDP per capita as a proxy for either economic opportunity costs (Collier and Hoeffler, 2004) or state capacity and reach (Fearon and Laitin, 2003). Although Collier et al. (2009) acknowledge that both economic opportunity and state capacity are important, they do not test the relative importance of the variables empirically. A few recent quantitative studies of civil war onset have applied more fine-grained measures of state capacity and accessibility. Fjelde and de Soysa (2009) test three measures capturing different dimensions of state capacity: Relative political capacity (actual revenue extraction relative to expected extraction given the country’s economy), Government expenditure/GDP, and Contract intensive money (the ratio of non-currency money to total money supply). They find the two latter measures to be negatively associated with civil war. The GDP per capita association with civil war remains negative when controlling for these factors, however, although it is no longer significant when controlling for Contract intensive money. Thies (2010) looks at the association between four revenue-based indicators of state capacity (Government expenditure/GDP, Total revenue/GDP, Tax revenue/GDP, and Relative political capacity) and civil war onset using a two-stage simultaneous equations technique. He finds that none of the revenue-based indicators affect the risk of civil war onset, and GDP per capita is still negatively associated with civil war onset when controlling for them. Both studies thus suggest that the development–civil war association is not due to a small public sector in poorer countries. In another cross-national study, Kocher (2004, ch. 3) finds that controlling for the percentage of the population living in urban areas, the GDP per capita coefficient is no longer statistically significant. Settlement pattern may thus be a central characteristic making poorer countries more prone to civil war. Few studies have looked at the association between infrastructure and the risk of civil war. Buhaug and Rød (2006), taking a geographically disaggregated approach, find that higher road density was associated with lower risk of armed conflict over territorial issues, but had no association with the risk of conflict over government in Africa. Raleigh (2010) finds that greater road density actually increased the risk of an armed conflict event occurring in areas of six African countries, where events included battles, violence against civilians, and establishment of rebel bases. Notably, the association between roads and rebel bases was negative although not significant. 10
Although these studies are useful, none of them assess the relative empirical support of the economic opportunity cost and politico-military opportunity arguments, as formulated above. To fill this empirical gap, I first identify indicators which allow for distinguishing between the two accounts’ empirical implications.
Indicators and Hypotheses
The measurement of economic opportunity costs is a weakness of the literature. In addition to GDP per capita, Collier and Hoeffler (2004) use Male secondary school enrollment and Economic growth as proxies for the concept. Both are found to be negatively associated with the risk of civil war. However, they are far from ideal indicators of potential rebels’ foregone income, since (i) they do not measure income, and (ii) they are very broad indicators. 11 According to economic opportunity theory, the poorest should be the first to enlist as rebel soldiers, as they have the least to lose. A national-level indicator of economic opportunity costs for rebelling should therefore be targeted on the income level of the poorest segment of society. Accordingly, I use the mean income or expenditure of the lowest income decile, Bottom decile income (BDI) per capita, as a cross-national proxy for economic opportunity costs (see details below). 12
I emphasize three conditions giving political and military opportunity for insurgency in poor countries: low state penetration and capacity; a rural, scattered settlement structure; and poor communications. The first is the most difficult to measure. Oft-used state capacity indicators like revenue extraction, bureaucratic quality, and military capability measures do not capture the concept well. 13 Available tax measures say little about the state’s reach, since they do not separate taxes on income, property, and capital gains from taxes on international trade, and the latter do not require high administrative capacity throughout a country’s territory (Fauvelle-Aymar, 1999: 392; Hendrix, 2010: 279). The Political Risk Services Group’s Bureaucratic quality measure is also not fitting, since it does not tell the degree to which state administration penetrates the entire country. 14 The most used military capability measure, Military personnel per capita, is also problematic. The counterinsurgency literature suggests that police and intelligence capabilities are more important for defeating nascent insurgencies than traditional military capabilities (Leites and Wolf, 1970: 154; Sepp, 2005). 15 I instead use roads per capita (Road density) as an indicator of state penetration. It may capture the reach of state bureaucracy as well as the state’s ability to project military force. 16 Herbst (2000) emphasizes the importance of roads for broadcasting state power, arguing that low density of roads, effectively cutting large areas off from the capital, is at the heart of most African governments’ weakness. Fearon and Laitin (2003: 80) also argue that poor countries are prone to civil war because their terrain is less “disciplined” by roads.
The second variable, settlement structure, can be proxied by the percentage of the population living in urban areas (% Urban). The third, communication infrastructure, can be proxied by the number of telephone lines per 1,000 inhabitants (Telephone density). Telecommunication infrastructure eases the flow of information over distance and may facilitate anonymous denunciation. It should therefore increase state agents’ ability to locate insurgents. 17 Rebel actions also bear witness to the importance of infrastructure for counterinsurgency: rebels have often turned to destruction of key communications such as bridges and telecommunication towers, despite the unpopularity of such actions among local populations (c.f. Hall, 1990; Gersony, 2003: 61).
None of these three indicators—Road density, % Urban, and Telephone density—sufficiently capture how large a part of the population is easily accessible to the state. State reach is co-determined by these factors. A formative, rather than a reflexive, measurement model is reasonably evoked here. It is difficult to imagine an underlying latent concept giving rise to these indicators; rather, the indicators together determine the reach of the state. No particular pattern of correlation between the indicators can therefore be assumed, making factor analysis inappropriate (Diamantopoulos and Winklhofer, 2001). 18 I construct a simple additive index—State reach—where the indicators are given equal weight. The indicators were standardized by minimum and maximum values before calculating the index. 19 Road density and Telephone density were also log-transformed before standardization because their marginal impact on the reach of the state is likely to be decreasing.
A simple quantitative implication can be derived from both theoretical accounts. According to the economic opportunity account, GDP per capita is related to the risk of civil war because it acts as a proxy for economic opportunity costs. If this holds, we should observe cross-nationally that:
HEc.Opp.: The statistical relationship between GDP per capita and the risk of civil war disappears when controlling for BDI per capita.
The politico-military opportunity account, on the other hand, suggests that GDP per capita is a proxy for how accessible populations are to state agents. This suggests that:
HPol-Mil.Opp.: The statistical relationship between GDP per capita and the risk of civil war disappears when controlling for State reach.
This study does not aim at tracing the exact pathways through which economic development increases the risk of civil war. It takes on the more tractable task of assessing the plausibility and strength of two prominent explanations of the development–civil war association. For this task, the nature of the relationship between GDP per capita and the two indicators of interest, BDI per capita and State reach, is not decisive. It is plausible that economic development, measured by GDP per capita, to a large extent causes variation in the other two indicators. If this is the case, BDI per capita and State reach could be seen as intervening variables through which GDP per capita has an indirect effect on the risk of civil war. It could also be that the relationship between GDP per capita and the other two indicators is not primarily causal, but that they are indicators of the same underlying variables, state reach or economic opportunity cost. Either way, we should expect the hypotheses to hold true.
Operationalizations and Data
Data on BDI per capita, the average income of the poorest decile, are taken from household surveys. 20 For low- and middle-income countries, I primarily use survey data reported in the World Bank’s Povcalnet (World Bank, 2008a). 21 For most high-income countries, I use data from the Luxembourg Income Study (LIS) (2009). 22 The temporal extent of the sample is limited by the availability of income survey data. Few countries have surveys in the 1980s, but coverage increases rapidly in the early 1990s. Taking this into account, I delimit the sample to the 1989–2006 period. I choose 1989 as the initial year because there were many new civil wars in the 1989–1993 period and because BDI values can be estimated back to that point without risking severe measurement error. 23 The sample period ends in 2006 because data for the subsequent years are lacking for BDI and some other explanatory variables. The spatial extent of the sample is also somewhat curtailed by data availability. I include the 133 countries having at least one survey observation and a population over 500,000.
The BDI per capita time-series have many missing values, which should be estimated to avoid bias (Honaker and King, 2010). The average number of surveys in a country is 3.7, and 16 countries have only one survey observation. As a first step, I linearly interpolate between the surveys. Next, I use the existing BDI observations as reference points and extend the time-series using the growth rate of household consumption expenditure per capita (PPP) from national accounts (World Bank, 2008b). 24 If household expenditure data is missing, I use the growth rate of GDP per capita (PPP) to extend the time-series. After this, only six country-years, and no onset cases, are left missing. 25 BDI per capita is log-transformed since it presumably has a decreasing marginal impact on the risk of civil war.
For Road density, I use data on the length of roads per capita from Canning (1998). It combines information from the International Road Federation, the UN, and national sources. I fill in missing observations where possible using World Development Indicators (WDI) (World Bank, 2008b). Since the time-series end between 2000 and 2002 for most countries, I lag the last value in each country’s time-series for the last 2–5 years. This should not cause discernible error, as the variable changes slowly over time. 26 For % Urban of the population I use WDI data (World Bank, 2008b), which is based on national reports. 27 The variable has complete sample coverage. 28 Telephone density, the number of telephone lines per 1,000 inhabitants, is also taken from WDI (World Bank, 2008b). 29 GDP per capita is measured in PPP-adjusted constant 2005 USD and taken from WDI (World Bank, 2008b). 30 The variable is log-transformed.
Civil war can broadly be defined as a conflict between a government and one or more domestic opposition organizations within a sovereign polity involving reciprocal violence of a considerable scale (Kalyvas, 2006: 5; Sambanis, 2004b; Small and Singer, 1982: 210). But what is a “considerable” scale? I concur with Sambanis (2004b) and Fearon and Laitin (2003) that the traditional threshold of 1,000 fatalities per year may be too high. It excludes some conflicts displaying organized armed opposition and extensive reciprocal violence, like that in Croatia from 1992. I use the somewhat lower and more flexible fatality threshold suggested by Sambanis (2004b). Onset of civil war is coded in the first year a conflict causes more than 500 battle-related deaths. If it causes below 500 (but over 25) deaths in its first year, onset is still coded in that year if cumulative deaths over the three subsequent years reach 1,000. A conflict must see at least two years of non-activity before a new onset can be coded. 31 Since Sambanis’s dataset is not updated beyond 1999, I use the UCDP/PRIO conflict data (Harbom and Wallensteen, 2010) and the PRIO battle deaths dataset to construct the variable (Lacina and Gleditsch, 2005). 32
Due to the limited sample size and degrees of freedom, I include only the two most significant standard control variables in the literature: total Population (logged) and Proximity of war in the main models (Hegre and Sambanis, 2006). 33 Population figures are taken from WDI (World Bank, 2008b). 34 Proximity of war is a decay function of the time since the last year of civil war in the country, or since independence if the country has not experienced civil war (Raknerud and Hegre, 1997: 393). 35 Table 1 shows descriptive statistics for the variables in the main models.
Descriptive Statistics
Exploratory Analysis
The utility of statistical analysis depends on the degree of independent variation between the two independent variables of theoretical interest, State reach and BDI per capita, and between these and GDP per capita. If they cannot be empirically distinguished, statistical analysis can only tell us that GDP per capita may be a proxy for both state reach and economic opportunity cost. The correlation between State reach and BDI per capita turns out not to be very high (r = 0.81). GDP per capita is quite highly correlated with State reach (r = 0.91), and somewhat less correlated with BDI per capita (r = 0.86). This independent variation is, as shown below, enough to see that the risk of civil war is more closely associated with State reach than with BDI per capita.
Since the outbreak of civil war is a rare event, the most valuable information lies in the onset cases. Figure 1 shows deviation from mean BDI per capita and State reach for the country-years with civil war onsets in the sample, with the number of onsets in parenthesis for countries with more than one onset. 36 The line shows the predicted values from a regression of State reach on BDI per capita for the entire sample, including the non-onsets. The onset cases above the line have higher State reach than expected from their BDI per capita and vice versa. The graph clearly shows that onset cases tend to have relatively lower State reach scores than BDI per capita scores. 37 In total, 37 out of 48 onset country-years have lower State reach than predicted from their BDI per capita. Most countries with civil wars in this period, such as Sri Lanka, Yemen, Nepal, Ethiopia, and Rwanda, are relatively poor, but what really sets them apart from the average country is a remarkably low State reach. This provides a first indication that the reach of the state is more strongly associated with the risk of civil war than is poverty.
Deviation from Mean State reach and BDI per capita for Countries with Civil War Onset, 1989–2006
Regression Analysis
I test the hypotheses using pooled logit analysis of cross-national time-series data for 133 countries from 1989 to 2006. The country-year is chosen as the unit of observation because more disaggregated data with sufficient spatial and temporal coverage do not exist for the variables of interest. Country-level analysis may involve measurement problems since some civil wars are primarily fought in regions where conditions differ from the country average (Buhaug and Lujala, 2005). The aggregation problem is not likely to be overwhelming here, however. In a global study, Buhaug et al. (2011) find that the negative relationship between development and the risk of civil war holds using geographically disaggregated data. They find that within countries, conflict begins somewhat more often in areas that are poorer than the country average. For the least developed countries, however, conflict begins more often in wealthier areas. Since there seems to be no strong unidirectional tendency, country-level aggregates should not yield substantial bias. The observations are pooled because the explanatory variables change relatively little over time, making within-country comparison futile. Standard errors are clustered on countries and a robust “sandwich” estimator is used.
First, I estimate separate logit models including only one variable of interest and the two control variables, Population and Proximity of war. 38 Figure 2 shows the risk ratio (RR) for the central independent variable of each model. The risk ratio is here defined as the ratio of the probability of onset when xi is at the upper quartile to the probability of onset when xi is at the lower quartile, holding control variables at their means. 39 A higher GDP per capita is strongly related to a lower risk of civil war. Its RR estimate of 0.41 means that countries in the upper quartile of GDP per capita have an estimated 59% (+/– ca. 20%) lower probability of civil war onset than countries in the lower quartile of GDP per capita, holding Population and Proximity of war at their means. The 95% confidence interval of the estimate, shown by the line, is well below a risk ratio of 1 (no risk difference). The association between BDI per capita and the risk of civil war onset is weaker and much less certain. The risk ratio estimate is not significantly different from 1 at the 5% level. State reach, on the other hand, has a risk ratio estimate very similar to GDP per capita, which is clearly significant. 40 The same is the case for Telephone density and % Urban. Roads density has a somewhat higher risk ratio of 0.61, but it is clearly significantly different from 1.
Risk Ratio Estimates for Central Variables from Separate Logit Models
Next, I estimate models where two of the three central variables are included in addition to the two standard control variables. Risk ratio estimates for each variable in these models, again holding all other variables at their means, are shown graphically in Figure 3 and numerically in Table 2. 41 Model 1, shown with black dots, includes GDP per capita and BDI per capita. The results go against the idea that the development–civil war relationship can be attributed to a low economic opportunity cost for the poor. The significantly negative GDP–civil war association actually becomes stronger (mean RR = 0.22) when controlling for BDI per capita, contrary to the economic opportunity hypothesis. The relative risk estimate of BDI per capita is highly uncertain, with most, but not all, of the confidence interval stretching across positive RR values. The mean RR estimate is 3.06, meaning that the risk of civil war is tripled when BDI per capita increases from the lower to the upper quartile, holding GDP per capita at the mean. It is not statistically significant, however.
Risk Ratio Estimates from Multivariate Logit Models
Risk Ratio Estimates from Multivariate Logit Models
Robust standard errors in parentheses. * p < 0.10; ** p < 0.05; *** p < 0.01.
The estimates show the ratio of the probability of onset when xi is at the upper quartile to the probability of onset when xi is at the lower quartile, with other x variables held at their means.
In Model 2 (white circles), State reach is included and BDI per capita is excluded. The results are in line with the politico-military opportunity hypothesis. Controlling for State reach, the relationship between GDP per capita and civil war disappears. The high correlation between GDP per capita and State reach makes their estimates very uncertain. For both variables, the confidence interval crosses RR = 1; they have no statistically significant relationship with the risk of civil war, controlling for the other variable. The mean risk ratio estimate is 0.99 for GDP per capita, whereas it is 0.57 for State reach. Still, little can be inferred from these point estimates because of their uncertainty.
In Model 3 (squares), GDP per capita is removed and BDI per capita is included together with State reach. The relative risk for BDI per capita is not significantly different from 1, but its mean estimate is strongly positive. State reach, on the other hand, has a strong and clearly significant negative association with civil war risk. The mean relative risk is 0.25, implying that countries in the upper quartile of State reach have 75% (+/– ca. 13%) less chance of civil war than countries in the lower quartile, holding BDI per capita at the mean. This suggests that State reach can plausibly account for the relationship between BDI per capita and civil war onset, but not vice versa.
Robustness Checks
Next, I test whether the results are sensitive to the choice of control variables, imputation method, and operationalization of the dependent variable. Figure 4 shows risk ratio estimates for BDI per capita and State reach from models where they are both included. As above, risk ratios are estimated holding all other variables at their means. In Model 1, the battery of variables from Fearon and Laitin’s (2003) first model are included as controls: New state dummy, Noncontiguous state dummy, % Mountainous terrain (logged), Democracy, Instability dummy, Ethno-linguistic fractionalization, Religious fractionalization, Prior war, and Oil dependency dummy. 42 Including these controls, the results (shown with black dots) even more strongly support the politico-military opportunity account. A higher BDI per capita becomes significantly related to a higher risk of civil war, with a mean risk ratio estimate of 3.33. State reach has a very strong negative risk ratio estimate: an increase in State reach from the lower to the upper quartile gives an 87% (+/– ca. 12%) reduction in the risk of civil war, holding BDI per capita and the control variables constant.
Risk Ratio Estimates for BDI per capita and State reach from Multivariate Logit Models
In the second model (white circles), missing BDI per capita values are estimated in a multiple imputation model using Amelia II (Honaker and King, 2010). Only the two standard control variables are included. The results are very similar to the first model, with the State reach risk ratio being strongly negative and clearly significant and BDI per capita having a positive risk ratio that is barely significant. The third model employs a slightly different operationalization of the dependent variable. The Sambanis definition is kept, but the period of conflict non-activity required for a new onset is increased from two to four years. Again, results are stable. In the third model, a different adjustment is made to the dependent variable: onset is coded if cumulative battle-related deaths over the preceding three years reach 1,000 (in the original, onset is coded in the first year with at least 25 deaths if the next three years reach 1,000 deaths). The positive risk ratio of BDI per capita is no longer significant, but the main result holds: only State reach has a significant negative effect. Model 4 uses the civil war onset operationalization of Strand (2006) based on the UCDP/PRIO data, which has a higher intensity threshold of at least 1,000 battle-related deaths in at least one year of the conflict. This gives only 25 onsets in the sample (with the preferred onset operationalization, there are 48). The estimates are therefore less certain, as seen by the wider confidence intervals. State reach still has a risk ratio below 1 which is barely significant and the mean risk ratio estimate of BDI per capita remains positive. 43
Conclusion
This article has assessed two prominent accounts of why poorer countries are more prone to civil war than wealthier ones: the economic opportunity account, suggesting that poverty makes rebellion financially viable because recruits with lower foregone income of rebelling will demand lower salaries, and the politico-military opportunity account, suggesting that weak state penetration of rural society in low-income countries makes it possible for rebels to survive by hiding from government forces and to gain local control or influence, which greatly enhances resource mobilization. I argued, based on the case study literature, that the latter account is more powerful than the former. By focusing on a single pre-war characteristic of the population—poverty—the economic opportunity account largely ignores the politics of resource mobilization. Armed organizations use several tools—persuasion, coercion, organization, and economic rewards—to mold the incentives and allegiances of people. Although low economic opportunity cost might ease recruitment, it may not be necessary if insurgents can use such tools effectively. Their ability to do so depends foremost on state reach and capacity, rather than poverty. Adding to this, beyond a certain level of state reach, insurgents can hardly survive through the initial phase of mobilization.
I carried out a test of cross-national implications of the two accounts using data for 133 countries from 1989 to 2006. I proxied the economic opportunity cost of rebelling by the income of the poorest segment of society (Bottom decile income (BDI) per capita) since, according to the theory, the poorest have the lowest foregone income and should therefore be the first to rebel. To measure state reach and capacity, I used an index (State reach) composed of Road density, the % Urban of the population, and Telephone density. I found that State reach was more closely associated with economic development than was BDI per capita. Moreover, BDI per capita could not account for the cross-national statistical association between GDP per capita and the risk of civil war, whereas State reach could. The results are robust to theoretically reasonable changes to model specification, dependent variable operationalizations, and imputation method.
This suggests that although civil war is largely a “problem of the poor”, poverty per se is probably not the crux of the problem. State infrastructural weakness and a rural settlement pattern are likely to be more important for explaining the high prevalence of civil war in low-income countries. Clearly, other interpretations of the statistical results are possible. For instance, state public goods provision may also be important because it provides state legitimacy or removes the popular demand for a rival organization offering public goods (Berman and Laitin, 2008; Fjelde and de Soysa, 2009). The development–civil war association may indeed have several, compatible causes. Still, the case-based and quantitative evidence shown here suggests that the political and military opportunity to effectively organize insurgency is a key explanatory factor whereas the economic opportunity cost of rebelling is, at best, secondary.
Footnotes
1
I would like to thank Håvard Hegre, Halvard Buhaug, Lynn P. Nygaard, and three anonymous reviewers for their helpful comments.
2
In the last two decades (1989–2009), only 2% of the world’s country-years with civil war (as defined by UCDP/PRIO) took place in upper middle- or high-income countries.
3
This point is acknowledged by Collier and colleagues in later work (Collier et al., 2009).
4
Note that the economic opportunity cost argument is part of their broader “economic opportunity” or “greed” theory of rebellion, which I have no ambition of assessing in its entirety here.
5
In recent work, Collier and colleagues maintain that economic opportunity cost is vital for explaining the development–civil war relationship, but acknowledge that state capacity may also be part of the explanation (Collier et al., 2009). No attempt is made at separating and empirically testing the importance of the two accounts, however.
6
It could be partly spurious, however, caused by the rebel organization focusing mobilization efforts in the poorer periphery, not because it is poor, but because this is where government presence is the weakest.
7
A large project involving 21 case studies designed to assess and expand Collier and Hoeffler’s economic opportunity theory also suggested that their economic mechanisms in several cases did not fit the evidence well. Forced recruitment was highlighted as one important mechanism ignored by the economic model (Collier and Sambanis, 2005; Sambanis, 2004a).
8
Notably, some insurgent organizations promise economic benefits to their combatants (Humphreys and Weinstein, 2008; Weinstein, 2007: 111, 125) but several appear to offer no or very few economic benefits (Gutiérrez Sanín, 2004: 268; Stoll, 1993: 138; Weinstein, 2007: 138, 108–110).
9
Note that although most insurgent groups are initially very small (Sambanis, 2004a: 267), in a substantial minority of civil wars since 1945, state power had almost broken down by the start of insurgency, making the rebel–government power balance more symmetric from the outset (Kalyvas and Balcells, 2010: 423).
10
Somewhat relevant for the this study, Djankov and Reynal-Querol (2010) find that controlling for European colonial settlement, the association between GDP per capita and civil war disappears. This finding is difficult to interpret, however, since European settlement is related not only to future economic development, but also to infrastructural and political development (Lange and Balian, 2008). They also find that when including fixed effects, that is, looking only at temporal variation within countries, the association between GDP per capita and civil war is no longer statistically significant. Still, although using GDP per capita estimates back to 1825, it is unclear whether it is reduced variation that makes the relationship insignificant. One attempt at avoiding the reverse causality problem is Miguel et al.’s (2004) study of civil war in Africa. Using rainfall as an instrument for economic growth, they find a strong relationship between growth and civil war. This finding has later been questioned, however (Ciccone, 2011). Several studies use spatial approaches to assess the association between structural factors, including poverty and unemployment, and violence or other events in civil wars (e.g. Berman et al., 2011; Hegre et al., 2009; Do and Iyer, 2010). They arrive at different conclusions: Do and Iyer (2010), for example, found that district-level poverty was associated with greater intensity of violence in Nepal’s civil war, whereas Berman et al. (2011) found that district- or province-level unemployment rates were not at all or negatively associated with attacks against government forces in Afghanistan, Iraq, and the Philippines. Although useful for understanding the dynamics of violence, one should be careful about drawing inferences about the conditions favoring insurgency from these studies, since violence need not be most intense where rebel mobilization is greatest (Kalyvas, 2006).
11
Male secondary school enrollment has the advantage of being focused on young men, who are overrepresented in most rebel groups. However, school enrollment is probably a poor indicator of the economic opportunity cost of rebelling for these young men.
12
I also refer to this as poverty.
13
14
Its indicators include how regular and meritocratic recruitment is, how insulated the bureaucracy is from political pressure, and the ability to provide services during government changes (Knack, 2001).
15
Also, a swollen army may actually indicate weakness, as rulers in a tenuous position are forced to increase military budgets to maintain the army’s loyalty (Henderson and Singer, 2001).
16
Clearly, Road density is not a perfect measure of state penetration and a very indirect proxy for capacity. I therefore avoid concluding strongly about the role of this particular variable, but rather focus on the composite measure of how accessible the population is to state agents, where road density is one of three components.
17
In the last few years of my sample, the use of mobile telephones could plausibly make telephone lines less important. It is only from the year 2000 that mobile subscribers reachan average of more than 10% for the countries in my sample. For 2000–6, Telephone density is still likely to be a decent measure since it is quite closely related to mobile subscription density (r = 0.80, using 
) data).
18
An important assumption of the formative index is that all indicators are included in the model. This is very difficult to assess, however.
19
Z-score standardization gives similar regression results (not shown for lack of space).
20
Alternatively, one could use GDP per capita along with Gini estimates of income distribution, which would allow for extending the sample. One important problem with this is that GDP underestimates income in poorer countries because they tend to have a large informal economy which is not registered in national accounts.
21
Povcalnet comprises 467 surveys from 114 low- and middle-income countries for 1980–2004 and gives information on mean income or expenditure per capita in PPP 2005 US$ and its distribution in deciles.
22
23
The average lapse from 1989 to the first survey in a country is 2.9 years.
24
I thus assume that the income distribution is stable and that the growth rate of national accounts expenditure is similar to that of the survey income. A similar method is used by on these data. Household consumption expenditure per capita (PPP) is correlated r = 0.90 with BDI per capita. The same correlation is found with GDP per capita (PPP). All variables are logged.
25
This gives 2,323 country-year observations, of which 495 are original survey observations, 1,038 are interpolated, and 790 are estimated using the growth rate of household expenditure or GDP per capita. For a thorough description of the imputation procedure, see the Appendix. One alternative method is multiple imputation—building a statistical model to estimate the missing values (c.f. Honaker and King, 2010). I construct an out-of-sample prediction test of the two methods’ performance, and I find that the method described above performs much better than a multiple imputation model using Amelia II (see Appendix for details). Still, reassuringly, regression results do not change using Amelia imputations.
26
The average annual change is 1.1%.
27
There is little information on their definitions and coding criteria. Kocher (: 58) explored this by browsing through the various countries’ census classifications. He found that a settlement is counted as urban if it has more than ca. 2000 inhabitants. Somewhat different thresholds were used. It is unknown how much this affects reliability.
28
For this variable and Telephone density, I recalculate values for states that have experienced border changes. WDI gives values based on present-day states (e.g. 1989 values are based on the borders of Russia, not the Soviet Union).
29
The time-series were interpolated. This should be unproblematic, as the variable changes slowly over time.
30
For states that have experienced border changes, I use the estimates of “Real” (PPP-adjusted) GDP per capita from Gleditsch (2002). I also use PWT 6.3 “Real” GDP per capita estimates (Heston et al., 2009) for a few country-years missing in WDI.
31
Non-activity is operationalized as having less than 25 battle-related deaths. I check whether results are sensitive to this choice using a four-year intermittency period.
32
This gives 48 civil war onsets in the sample. I also check whether the results hold using a higher threshold of 1,000 deaths in at least one single year of the conflict.
33
Having few control variables may increase the risk of omitted variable bias, but I hold that the risk is only significant if there are strong theoretical reasons to expect that a confounding variable is omitted. Including a host of control variables on weak theoretic grounds has pitfalls of its own. It is difficult to find out whether they reduce or rather increase omitted variable bias (Clarke, 2005), and they can make results more difficult to interpret (Ray, 2005).
35
I use the high-threshold UCDP/PRIO definition of civil war as a basis for this control variable due to the difficulty of making a complete civil war list based on the Sambanis definition with the UCDP/PRIO data. The difference would be small and probably have no impact since the variable has very little influence on the coefficients of interest. The half-life of the decay function is set to eight years.
36
For countries with more than one onset, average values of BDI per capita and State reach for the onset years are shown.
37
Plotting GDP per capita against BDI per capita gives a quite similar figure, whereas a figure with State reach and GDP per capita shows the onset cases lying much closer to, and more equally distributed around, the regression line.
38
Regression coefficients for all models are found in the Appendix.
39
Risk ratio estimates are made using simulations with Zelig (Imai et al., 2006).
40
The model with BDI per capita also has significantly worse fit to the data than the models including State reach or GDP per capita. The area under the ROC curve (AUC) is smaller for the BDI per capita model (AUC = 0.746) than for the State reach model (AUC = 0.776) and the GDP per capita model (AUC = 0.771). This difference between the BDI per capita model and each of the other two models is significant at the 5% level. The AUC values of the GDP per capita and State reach models are not significantly different.
41
I do not simulate risk ratios holding the other variables at different levels because it would mean invoking counterfactuals far from the actual data. Using the “WhatIf” software (King and Zeng, 2006) I found that the percentage of observations “nearby” the counterfactual sank considerably when holding the other variables more than +/–10 percentiles away from their means. All the models included here are based on counterfactuals with more than 10% of the data in the convex hull.
42
See Fearon and Laitin (2003: 78–81) for details. Since the time-series end in 1999, I recode the variables that change over time to avoid missing values. Their Oil dependency dummy marks country-years in which oil exports exceed one-third of export revenues, whereas my variable marks country-years in which oil exports exceed 15% of GDP. The Democracy and Instability indicators are recoded using updated data from PolityIV, 2007.
43
Additional robustness tests are not shown for lack of space. I test whether African conflicts are different in terms of the impact of these variables by including an Africa dummy interacted with BDI per capita in one model and with State reach in another. None of the interaction terms are significant, and the BDI per capita estimate only becomes slightly more positive and the State reach estimate slightly more negative (see Appendix). I also check sensitivity to outliers and influential cases. Most vital to the main argument, I exclude the onset cases with the highest State reach compared to BDI per capita score—the four Ethiopian onsets (see ). The results do not change discernibly when excluding these, or other, influential cases.