Abstract
A large body of literature claims that oil production increases the risk of civil war. However, a growing number of skeptics argue that the oil–conflict link is not causal, but merely an artifact of flawed research designs. This article re-evaluates whether – and where – oil causes conflict by employing a novel identification strategy based on the geological determinants of hydrocarbon reserves. We employ geospatial data on the location of sedimentary basins as a new spatially disaggregated instrument for petroleum production. Combined with newly collected data on oil field locations, this approach allows investigating the causal effect of oil on conflict at the national and subnational levels. Contrary to the recent criticism, we find that previous work has underestimated the magnitude of the conflict-inducing effect of oil production. Our results indicate that oil has a large and robust effect on the likelihood of secessionist conflict, especially if it is produced in populated areas. In contrast, oil production does not appear to be linked to center-seeking civil wars. Moreover, we find considerable evidence in favor of an ethno-regional explanation of this link. Oil production significantly increases the risk of armed secessionism if it occurs in the settlement areas of ethnic minorities.
Introduction
A vast and influential literature argues that oil-producing countries face a significantly larger risk of civil conflict than other states (Koubi et al., 2014; Ross, 2015). 1 Most recently in connection with the Islamic State’s reliance on oil revenue to finance combat in Iraq and Syria, the relationship between oil and violence has also attracted significant attention from policymakers (World Bank, 2011). In addition to its apparent effect on the outbreak of violence, oil has also been associated with autocratic regimes and weak economic growth, giving rise to the concept of an ‘oil curse’ (Ross, 2012). Notwithstanding its status as ‘stylized fact’, the oil–conflict link has recently come under attack.
Critics rightly argue that the statistical evidence underlying much of the oil curse literature may be biased. These skeptics argue that oil production is not a random treatment, but an industrial activity subject to economic and political incentives. Consequently, correlational evidence of the oil–conflict link may be misleading. Indeed, once this inferential threat is addressed with more sophisticated research designs, the empirical support for a causal effect of oil on conflict becomes inconclusive (e.g. Brunnschweiler & Bulte, 2009; Mitchell & Thies, 2012; Cotet & Tsui, 2013).
Despite these suggestive findings, the question of whether there is an oil curse remains open for several reasons. First, most causal identification strategies proposed in the literature are vulnerable to reverse causality. Second, research designs that focus on within-country variance or only on oil discoveries may help eliminate omitted variable bias, but discard so much information that they potentially conceal a positive effect. Third, focusing on empirics at the country level, the criticism of the oil curse literature so far been largely detached from the debate about the theoretical underpinnings of the oil–conflict link.
This article addresses these challenges directly. First, we employ geospatial data on the location of sedimentary basins as a new spatially disaggregated instrument for petroleum production that offers a much more solid basis for causal inference than previous attempts to deal with endogeneity. Combined with newly introduced geocoded information on the exact location of productive oil and gas fields, this identification strategy enables us to re-evaluate whether the oil curse is causal.
Second, beyond merely testing the existence of the oil–conflict link, we also argue that it is best explained as an ethno-regional oil curse. Petroleum extraction in regions inhabited by locally concentrated ethnic groups risks provoking secessionist violence, especially if locals see few of the benefits of oil production, while bearing most of its costs. This account differs from alternative explanations that locate the origins of the oil–conflict link at the individual or governmental levels.
Our empirical analysis confirms that petroleum extraction exerts a large and positive causal effect on the probability of violent conflict. Furthermore, in line with our ethno-regional argument, we find that the oil–conflict link originates in regions inhabited by territorially concentrated ethnic groups that seek secession. Moreover, this effect is particularly large for ethnic minorities that are politically excluded from the national executive. Finally, the results also show that studies that fail to correct for endogeneity tend to underestimate the effect of oil on conflict, especially at the subnational level.
The remainder of this article is structured as follows. The next three sections review the literature, introduce our theoretical framework, and derive testable hypotheses. A subsequent section introduces the main datasets. We then test whether there is an oil curse, followed by an evaluation of our ethno-regional argument. Finally, the concluding section discusses the consequences of our analysis for theory and policy.
The oil curse and its critics
The claim that oil causes the outbreak of civil war is well established (for recent reviews, see Koubi et al., 2014; Ross, 2015). Under the general heading of an ‘oil curse’ (Ross, 2012), this result is often cited in combination with the related findings that oil hinders democratization (e.g. Ross, 2001) and economic growth (e.g. Sachs & Warner, 1995).
Challenging this consensus, however, a growing number of scholars assert that most statistical analyses in support of the oil curse make causal claims based on the incorrect premise that petroleum production is exogenous to societal outcomes (see e.g. Brunnschweiler & Bulte, 2009; Mitchell & Thies, 2012; Cotet & Tsui, 2013; Lei & Michaels, 2014). They argue that whether, and to what degree a given country or region produces oil is not exogenously given, but determined by sociopolitical and economic factors.
Several mechanisms may cause such endogeneity. The most obvious one is the immediate negative effect that large-scale violence exerts on a country’s ability to extract oil and gas. Unless accounted for, these effects may attenuate estimates of the oil–conflict link (Ross, 2004). A second potential source of endogeneity is that oil-producing states may be structurally different in ways that affect their conflict propensity, but are difficult to observe or measure. In which direction this type of endogeneity biases estimates of the oil–conflict link depends on the underlying argument. Some authors argue that weak states are more likely to rely on an oversized oil industry because they are unable to establish competitive manufacturing and service sectors (Ross, 2004: 338; Haber & Menaldo, 2011: 2). Since weak and inefficient states are also more likely to experience violence, this mechanism would imply that previous studies have overstated the oil–conflict link. In contrast, Cotet & Tsui (2013: 50) and Torvik (2009: 245) suggest that fragile institutions may be associated with less oil production, because weak property rights disincentivize investments in extractive infrastructure. In this case, previous studies would have underestimated the gravity of the oil curse. Finally, another plausible source of endogeneity is long-term reverse causality. Specifically, the mere anticipation of conflict may deter prospective investors from financing oil exploration and extraction projects (Brunnschweiler & Bulte, 2009: 654). This mechanism leads to an underestimation of the oil–conflict link, as particularly conflict-prone countries will systematically host less oil production.
A number of recent studies attempt to address endogeneity by re-evaluating the oil–conflict link with more sophisticated causal identification strategies. First, some authors employ fixed effects panels, hoping to eliminate bias due to unobservable cross-country differences. Based on this strategy, Cotet & Tsui (2013) find no effect. Similarly, Haber & Menaldo (2011) conclude that once unobserved cross-country differences are accounted for, there is no evidence for the claim that oil hinders democratization. Yet, fixed effects do not solve the issue of reverse causality. Moreover, removing all cross-country variance could produce false negatives, as differences between countries are an important source of explanatory power for analyzing political outcomes.
Another way to address endogeneity is to use an instrumental variable design. The merit of this approach depends on the quality of the instruments, which should be good predictors of oil production, exogenous to the outcome under investigation, without affecting the outcome through any other channel than oil. The latter two conditions are commonly known as the exclusion restriction. Yet, it is doubtful whether existing instruments meet these requirements. Instrumenting oil production through macroeconomic variables, Mitchell & Thies (2012) find no support for the oil–conflict link. However, it is debatable whether their instruments are good predictors of oil production, and these types of measures clearly do not meet the exclusion restriction.
As an alternative, other scholars use proven oil reserves as an instrument for resource dependence and oil production. Brunnschweiler & Bulte (2009) report that once instrumented, natural resource production no longer affects civil conflict. Similarly, Haber & Menaldo (2011) find no effect of oil production on regime type when using reserves as instrument. While oil reserves certainly predict oil production, this variable is hardly exogenous to conflict. First, reverse causality may be an issue, as ongoing or anticipated conflict may deter not only oil production, but also oil exploration. Second, reserves are typically endogenous to structural economic or political factors, since the concept of ‘proven reserves’ explicitly accounts for local socio-economic conditions. 2
Finally, Cotet & Tsui (2013) and Lei & Michaels (2014) address endogenity by relying on information about oil discoveries, rather than production. This method is appealing because assuming that exploration is taking place, discoveries are largely random, and thus approximate the ideal of an experimental ‘random treatment’. Based on this approach, Cotet & Tsui (2013) find no evidence of an oil–conflict link, whereas Lei & Michaels (2014) report a positive effect for the case of giant fields. Yet, being limited to the analysis of newly discovered fields rather than the effects of total oil production, these studies arguably underestimate the true impact of petroleum on conflict.
If there is an oil curse, then which one?
Since most skeptics question the very existence of the oil–conflict link, they say little about its underlying causal mechanisms. This theoretical deficit contrasts with the ‘embarrassment of mechanisms’ (Humphreys, 2005) that characterizes most of the conventional literature. The question, then, is not only whether there is such a curse, but also which one.
Structuring our theoretical discussion according to levels of analysis, we start by considering the most influential explanations at the level of individual motivations and governmental structures, before discussing mechanisms located at the regional level. First, we summarize a set of mechanisms that highlight the motives of individuals under the heading of the individualist oil curse. Following the pioneering contributions by Collier & Hoeffler (1998), many scholars argue that natural resources trigger conflict by affecting the cost–benefit rationale of prospective rebels and warlords. Specifically, the expected revenue from trading valuable minerals and gemstones is assumed to motivate individuals to take up arms and challenge state authority. Moreover, this interpretation proposes that the financing of ongoing conflict by looting natural resource deposits and extorting resource extractors lowers the marginal costs of fighting (Collier & Hoeffler, 2004: 565; see also Ross, 2012: 151).
The main problem with these individualist interpretations of the resource–conflict link is that they are generally more convincing as accounts of conflict duration through rebel financing than as causes of conflict onset. While oil-fueled start-up funds are in principle possible, in most cases such a scenario is hardly feasible. In contrast to alluvial diamonds and drugs, the main problem pertains to lootability (Humphreys, 2005). Indeed, oil extraction typically requires territorial control, or at least physical access. Most states and oil-extracting companies are capable of defending their installations. If rebels come to control petroleum production sites, they are already running successful campaigns, as illustrated by the Islamic State. As a rule, then, effective exploitation and marketing of oil require considerable resources that only a state or a state-like actor can field, thus casting doubt on the relevance of the individualist oil curse (Fearon, 2005: 500).
At the country level, a class of mechanisms can be summarized as the governmental oil curse. The most prominent explanation is based on the weak-state mechanism, which assumes that resource extraction does not affect the occurrence of rebellion directly, but prevents the state from prohibiting the rise of violent challengers in its periphery. Initially proposed by Fearon & Laitin (2003), this mechanism holds that the conflict-inducing effect of resource extraction runs via the latter’s impact on government revenue. Freed from the need to generate tax-based revenue, oil-endowed states abstain from creating the type of intrusive institutions that are necessary for tax collection. Yet, the absence of a low-level administrative apparatus also prohibits the state from effectively policing its population, thus facilitating the organization of violent resistance against the state. An alternative account at the same level, sometimes referred to as the ‘honeypot effect’, views the state as a lucrative target of enrichment, thus accounting for why greedy rebels would be motivated to topple the government (Fearon, 2005; Le Billon, 2005).
However, there are strong reasons to doubt the relevance of these country-level theories. While the perverting effects of petroleum on state-building appears plausible, such accounts overlook the obvious possibility that strategic governments should be able to deploy their often formidable oil revenues to defend against rebellious challenges (Colgan, 2014: 7; Paine, 2016). The other country-level mechanism also suffers from serious shortcomings. As explained by Ross (2012: 161), the honeypot effect relies on unrealistic assumptions that fail to account for how collective action problems can be overcome.
Because of the theoretical weaknesses undermining both the individualist and the governmental oil curses, we shift our theoretical attention to the regional level. Our approach builds on previous ethno-regional explanations to develop a theoretical framework that accounts for when and why oil production leads to violent conflict. On this basis, we then derive a set of hypotheses that allow testing not only whether the oil–conflict link exists, but also whether the ethno-regional approach provides more explanatory leverage than the alternative individualist and governmental oil curses.
The political economy of industrial oil and gas extraction has a number of properties that set it apart from other economic activities, making it particularly likely to evoke secessionist demands. First, whether oil extraction benefits the resident population in productive regions is almost entirely determined by policy. Unlike labor-intensive industries, such as plantation agriculture, oil extraction typically offers few employment opportunities for locals. However, because producers are relatively easily taxed, oil extraction typically generates large windfalls for the central government. Consequently, whether locals benefit from oil production is almost entirely a question of redistributive policy. As argued by proponents of the regional explanation of the oil–conflict link, this situation creates considerable incentives for locals to support secessionist movements. If oil-producing regions succeed in establishing an independent state, then the appropriation and redistribution of oil rents is relocated to the regional level, thus yielding a significant increase in per-capita payoffs to local residents (Collier & Hoeffler, 2006; Sorens, 2011; Ross, 2012).
Absent strict regulation, oil production in populated areas is often associated with drastic negative externalities for the resident population. Numerous case studies document how oil production in developing countries may threaten local livelihoods through environmental pollution and large-scale land expropriation. For instance, Human Rights Watch (1991: 54) reports oil spills and widespread pollution in the Niger Delta, Kell (2010: 37) discusses environmental and social externalities in Indonesia’s Aceh province, and Ramos (2012) documents the effects of oil production on fishing grounds in the Angolan Cabinda exclave. However, few central governments face incentives to address these externalities. While strict regulation benefits only a small part of the government’s constituency, lax regulation provides the government with more funds to ensure its access to power. This dilemma creates additional incentives for locals to support groups advocating secessionism, simply as a means to ensure that local interests are represented in the regulation of oil extraction. 3
The problems of rent redistribution and local externalities are particularly likely to spark violent secessionism if oil is extracted in the settlement area of territorially concentrated ethnic groups (Sorens, 2011). This setting offers particularly fertile breeding grounds for local elites to mobilize support for secessionist goals by referencing petroleum-related grievances.
Territorially concentrated ethnic identities facilitate mobilization because they make promises over future oil-rent payoffs more credible. Clearly visible ethnic markers allow an unambiguous assessment of the future beneficiaries of independence, and thus attenuate the collective action problem associated with fighting for secession (e.g. Caselli & Coleman, 2013). Further, ethnic differences between the ‘foreign’ beneficiaries and the local victims of oil extraction will allow secessionist leaders to frame oil production as an issue of ethno-nationalist self-determination. Specifically, local elites may gather support for secession by framing grievances over unfair rent-distribution and the state’s apparent indifference to local externalities as a problem of ‘internal colonialism’ (Hechter, 1975), whereby the ‘foreign’ state is accused of illegitimately plundering locally owned resources (see also Aspinall, 2007). Finally, in regions inhabited by territorially concentrated ethnic groups, pre-existing intra-ethnic political and societal networks may be used to overcome the collective action problem inherent in political mobilization (Bates, 1983).
In summary, these arguments imply the existence of an ‘ethno-regional oil curse’: oil production in areas inhabited by locally concentrated ethnic groups is likely to cause widespread and easily mobilizable support for secessionism. Moreover, this mechanism is particularly relevant in regions inhabited by ethnic groups with limited control over government policy, which face the issues outlined above to their greatest extent.
Apart from its theoretical appeal, the ethno-regional account has also received tentative empirical support in recent years. Re-analyzing the relationship between oil production and center-seeking civil wars, Paine (2016) finds no evidence of an oil–conflict link once the positive impact of petroleum on governments’ financial capabilities is taken into account. This finding runs counter to the weak-state and honeypot explanations of the oil curse. Furthermore, suggesting that the oil–conflict link may indeed operate at a regional level, Asal et al. (2016) and Morelli & Rohner (2015) find that oil-rich ethnic groups are particularly likely to engage in rebellion.
Despite these results, however, the empirical record of the ethno-regional explanation remains inconclusive. Most importantly, because the above-named studies do not account for endogeneity, it is questionable whether they permit strong theoretical conclusions. For instance, Paine’s (2016) non-finding may be due to reverse causality since particularly vulnerable governments may be unable to maintain a large-scale petroleum industry. Another open question is whether the ethno-regional oil curse is limited to secessionist civil wars. While we expect this to be the case, Asal et al. (2016) and Morelli & Rohner (2015) do not distinguish between different types of conflict in their empirical analyses. Finally, it remains unclear whether the oil–conflict link is a function of political representation. Though these studies find that the oil–conflict effect is only statistically significant for groups without governmental representation, they do not test whether the oil–conflict effect is significantly different between politically included and excluded groups.
Observable implications of the ethno-regional oil curse
In this section, we derive a number of observable implications from the ethno-regional account, and analyze how these differ from those yielded by the individualist and governmental explanations.
First, our theoretical discussion clearly implies that there is a causal oil–conflict link: H1: Oil extraction increases the risk of civil conflict.
Second, our theoretical framework implies that oil should be associated primarily with secessionist conflict. If locals support resistance against state rule due to the prospect of gaining exclusive access to local oil rents and being able to curb the externalities of oil extraction, then the more limited aim of secession is sufficient (Sorens, 2011). In fact, even if assuming power over the central government were feasible, it implies that some fraction of local oil revenue would again have to be redistributed. In addition, the ethno-regional oil curse predominantly affects groups that are unlikely to succeed in center-seeking conflicts. Hence, we expect that H2: Oil extraction increases the risk of secessionist conflict.
Evaluating H2 is important for two reasons. First, it serves as a critical test of our theory: if oil is not associated with secessionist, rather than center-seeking, civil wars, then the ethno-regional approach cannot explain the oil–conflict link. Second, this hypothesis evaluates the explanatory leverage of the ethno-regional argument in relation to other approaches. Specifically, evidence in favor of H2 implies that the honeypot explanation for the oil–conflict link is insufficient.
Next, the ethno-regional argument implies that whether oil causes conflict depends on its evoking secessionist demands among the resident population in extractive regions. Whereas petroleum extraction in uninhabited areas neither threatens the livelihood of locals, nor evokes strong local ownership claims, in populated areas it should be associated with an increase in conflict risk. Thus, we postulate that H3: Only oil production in populated areas increases the risk of secessionist conflict.
Evaluating this hypothesis is important because it is inconsistent with the state-weakness and ‘honeypot’ accounts. Since these alternative explanations rely on government revenue, rather than oil extraction, they imply that the local conditions of oil production should be inconsequential. Thus, H3 suggests that there is an effect of oil on conflict that cannot be explained by these state-level mechanisms (cf. Lujala, 2010).
Next, the ethno-regional mechanism implies that even if occurring in populated areas, oil extraction is more dangerous in some places than in others. Our theoretical argument suggests that territorially concentrated ethnic groups that lack governmental representation should be at a particularly high risk of reacting to oil extraction with secessionist demands. Hence, we expect that H4: Oil extraction has a particularly large effect on the risk of secessionist conflict in areas inhabited by locally concentrated ethnic groups that are excluded from central government.
Finally, demographic size also influences whether petroleum-producing ethnic groups engage in violent secessionism. Thus, we expect the ethno-regional oil curse to be particularly relevant for locally concentrated ethnic minorities. First, oil extraction in the territory of small, locally concentrated ethnic groups ensures that a significant proportion of group members are affected by the accompanying externalities, thus facilitating mobilization along ethnic lines. Second, demographic weight also has implications for the potential payoffs associated with independence. The smaller a petroleum-producing group in relation to the rest of the country’s population, the larger the potential increase in redistributive transfers if secession succeeds. Finally, limited demographic weight may also be relevant as very small minorities may feel that they are unable to defend their interests vis-à-vis larger ethnic groups even if they are presently represented in government. Thus, minorities may fear that even if the central government agrees to acceptable petroleum-related policies now, larger ethnic groups may easily renege on these promises in the future, thus creating an acute commitment problem (see e.g. Sorens, 2011).
Given these considerations, we expect that H5: Oil extraction has a particularly large effect on the risk of secessionist conflict if it occurs in areas inhabited by an ethnic minority.
In order to distinguish our argument from alternative explanations, it is important to note that Hypotheses 4 and 5 are again incompatible with the state-weakness and honeypot mechanisms, as they postulate that the conditions under which oil is extracted determine whether it causes conflict. Moreover, evidence in support of H4 and H5 is also difficult to explain with individualist explanations, because there is no reason to expect that opportunities for oil theft and extortion are affected by the ethnicity, relative demographic weight, or political representation of local residents. In contrast, these moderating factors are inherently regional, and evidence that they matter would support our argument that the oil–conflict link operates by causing widespread support for secessionist policies in oil-producing regions, rather than by affecting the incentive structure of prospective rebels.
Instrumenting and measuring oil: New data
Addressing the endogeneity concerns discussed earlier, our identification strategy is to use information on the location and thickness of sedimentary basins to construct an instrumental variable (IV) for onshore oil and gas production. Sedimentary basins clearly meet the requirements for an IV design.
First, sedimentary basins are highly predictive of the location and extent of petroleum production, as they hold almost all of the world’s recoverable oil and gas reserves (Hyne, 2012: 17). The source of all oil and gas is organic matter that has been deposited in sedimentary material and then exposed to sufficiently high temperatures. These conditions are only present in particularly thick layers of sedimentary rock – called sedimentary basins – where geothermal energy is sufficiently high to trigger the formation of hydrocarbons.
Second, the sediment instrument meets the exclusion restriction. Unlike other potential instruments for oil production, sedimentary basins are absolutely exogenous to human activity. Moreover, there is little reason to believe that sedimentary rock affects conflict through any other channel than the presence of hydrocarbon deposits. One potential exception is the presence of mountainous terrain, which has been linked to the outbreak of civil conflict in various studies (see e.g. Fearon & Laitin, 2003). However, it is straightforward to account for this possibility econometrically by adding a respective control.
A further advantage of the sediment-based IV design is that it allows causal inference at the subnational level. Because geographically disaggregated information on the presence and thickness of sedimentary rock is available, we may construct corresponding instrumental variables for arbitrary subnational units. Despite its advantages, the sediment instrument is also limited in that it is time-invariant, and thus does not allow distinguishing between the effects of long-term and recent petroleum extraction.
To identify regions featuring thick layers of sedimentary rock, we employ the CRUST 1.0 dataset by Laske et al. (2013). CRUST 1.0 is a raster map providing information on sediment thickness at a resolution of 1 decimal degree grid cells for the entire globe, as shown in Figure 1. More detailed information for selected countries is shown in Figure 2.
Sedimentary thickness (in km) from the CRUST 1.0 dataset
Sedimentary thickness (in km) and ACOR fields (white/black dots) in Nigeria (left) and Burma (right)
Our theoretical argument implies that the location of oil extraction within countries affects whether and where conflict erupts. To test this claim, geocoded information on the location of oil production sites is indispensable. For this purpose, we introduce the new ACOR (Automatically Coded Oil Reserves) dataset, which provides precise geocoded information on the location of oil and gas fields. Specifically, the dataset records the shape, size, and location of all known onshore oil and gas fields in 1982, and is based on an atlas by Mayer (1982). 4 To code these roughly 13,000 polygon features, we developed automatic vectorization software leveraging recent advances in computer vision and machine learning.
ACOR is not the first geocoded petroleum field dataset, but builds on the important work of Lujala, Rød & Thieme (2007), who introduced the PETRODATA dataset. The reason we collected new data is twofold. First, PETRODATA has limitations with respect to spatial precision. Lujala, Rød & Thieme (2007: 246ff) generalize petroleum field locations by buffering and aggregating them into large polygons. This approach is problematic for the empirical analyses conducted later in this article because it overestimates ‘petroleum producing’ subnational units. Second, PETRODATA does not permit approximating the intensity of petroleum production in any particular region. In contrast, being coded on a field-by-field basis, the ACOR data allow creating a rough estimate of total production by counting the number of fields in a given area.
Another dataset providing information on the location of petroleum fields is offered by Horn (2010). While Horn’s dataset encodes fields as points, and even provides estimates of their volume, its key shortcoming is that it is limited to giant fields. While giant oil fields do yield a majority of the world’s oil production (65%, see Robelius, 2007: 82ff), relying solely on the Horn data implies ignoring the vastly greater number of smaller fields. Furthermore, while giant fields are responsible for the majority of oil production in some regions, this is not the case for Africa and Asia, where in 2005, the majority of oil produced originated from other fields (Robelius, 2007). Consequently, similar to PETRODATA, the Horn data do not permit measuring the actual extent of onshore oil production in a given region with high precision. 5
The precision gains associated with using ACOR may be illustrated visually. Figure 3 maps discovered onshore oil and gas fields, as coded by ACOR and PETRODATA, for Burma in the year 1982. 6 Superimposed is the settlement area of Burma’s majority Bamar ethnic group, as identified by the GeoEPR dataset by Wucherpfennig et al. (2011). Based on the ACOR data, it is evident that all Burmese oil production was located in regions inhabited by the Bamar at the time. In contrast, the PETRODATA polygons also overlap with the settlement area of other ethnic groups.
Comparison of ACOR and PETRODATA data of discovered Burmese petroleum fields in 1982
Finally, the ACOR data allow us to demonstrate the close relationship between sedimentary basins and the location of petroleum deposits. Figure 2 displays the CRUST sediment thickness data together with the ACOR oil and gas field information for Nigeria and Burma. In both cases, visual inspection confirms the impressive predictive power of the sediment information.
Does oil cause conflict?
This section evaluates the country-level evidence of the oil–conflict link using our newly proposed causal identification strategy. Specifically, we analyze whether the data support our expectations that there is a causal relationship between petroleum and civil war (H1), that this relationship is limited to secessionist conflicts (H2), and that it is due to oil production in populated areas (H3). To do so, we pursue a cross-sectional research design, and estimate the effect of onshore oil production on whether countries experienced civil conflict in the 1990–2013 period. There are several reasons why we limit the analysis to this time frame.
First, we are only able to measure onshore oil production with reasonably high precision in the post-1982 period. The vast majority of today’s oil reserves were discovered between 1950 and 1980 (Bentley et al., 2000: 170). Consequently, most regions producing oil today have been doing so since at least the early 1980s, and their output is thus approximated reasonably well by the 1982 ACOR data. Using the ACOR data for measuring production in the pre-1982 period, however, is likely associated with considerable measurement error, as we are bound to ascribe oil production to areas where no hydrocarbons were yet discovered. A similar caveat applies to the use of the time-variant PETRODATA or Horn datasets. Due to the aforementioned shortcomings of these datasets, using them to expand the temporal scope of the study is only possible at the expense of measurement accuracy. The reason for preferring to avoid these sources of measurement error – and thus pursue the temporally limited research design – is statistical efficiency. In IV designs, additional noise in the endogenous regressor translates directly into a less efficient estimate of the treatment parameter. 7 A second limiting factor is that geocoded data on settlement patterns – which we require to test the hypothesized importance of oil production in populated areas – is only available for the post-1990 period (CIESIN, 2010). In combination, these restrictions imply that we are only able to measure oil production, and especially oil production in populated areas, with relatively little measurement error for the years after 1990. 8
We rely on a cross-sectional analysis of the data because both our treatment of interest (oil production) and the instrument (sediment volume) are measured in a time-invariant manner. Hence, while pursuing a panel design is feasible (we present respective robustness checks in the Online appendix), doing so has few inferential benefits. Accordingly, as dependent variables, we employ dummies that indicate whether a country experienced at least one civil conflict onset during the 1991–2013 period. 9 These measures are constructed using the Uppsala Conflict Data Program’s (UCDP) armed conflict dataset version 4-2014 (Pettersson & Wallensteen, 2015), which also codes whether a given conflict involves territorial or governmental incompatibilities. We employ the latter distinction to measure secessionist and center-seeking conflicts, respectively. To measure onshore petroleum production, we use ACOR to calculate the logged number of onshore fields per country as a proxy (while adding a unit constant before logging). As an instrument for onshore production, we calculate the logged volume of sediment rock underneath each country’s territory using the CRUST data.
Next, to evaluate H3, which states that only oil production in populated areas should be associated with secessionist conflict, we introduce two new variables measuring the number of petroleum fields in populated and unpopulated areas for each country. In addition, we define two instruments capturing each country’s total sediment volume in populated and unpopulated areas, respectively. We create these variables in two steps. First, we create a global raster map of populated areas in 1990 based on the GRUMP (version 1) settlement data by CIESIN (2010). A raster cell is coded as populated if the population density in its neighborhood exceeds a minimal threshold of 0.1 inhabitants per square kilometer. Second, the (un-)populated oil field and sediment volume variables are generated by intersecting the ACOR and CRUST data with the above-described raster map (see Section A.3.1.5 of the Online appendix).
The IV design requires two types of control variables. First, it is necessary to control for logged country area, as sediment volume is an immediate function of surface area, and vast countries may be more prone to civil conflict simply because larger territories are more difficult to govern. Analogously, when distinguishing between petroleum fields in populated and unpopulated areas, we control for the total populated and unpopulated territory of each country. Further, as a cautionary measure, when analyzing the impact of oil in populated regions, we also control for each country’s total logged population in 1991.
Second, as stated above, it is also necessary to control for mountainous terrain. We do so with a variable measuring the fraction of each country’s territory covered by mountainous terrain, as defined by UNEP-WCMC (2002). These controls are necessary for causal identification because they are causally antecedent to sedimentary volume, and omitting them carries the risk of wrongly attributing their effect on conflict to oil production. However, this does not apply to other commonly employed controls in the conflict literature. Income levels, for instance, are not causally antecedent to the presence of sedimentary rock, nor is there any reason to believe that sedimentary basins affect income levels through any other channel than hydrocarbon reserves.
Econometrically, we rely on standard 2SLS models to implement our IV design, and report Huber-White standard errors to account for heteroskedasticity. Angrist (2001) advocates this approach as the most consistent strategy for estimating treatment effects for binary outcomes. Further, to provide an inferential baseline, the IV results are complemented with estimates from uninstrumented linear probability models.
We now move to the discussion of the country-level results, summarized in Table I, and visualized in Figure 4. For each model specification, we present the results of an ordinary linear probability model next to the 2SLS estimated results. Further, for those IV models featuring only one endogenous regressor, we report first-stage estimates.
(Instrumented) petroleum and civil conflict, country-level, 1991–2013
***p < 0.001, **p < 0.01, *p < 0.05. Standard errors in parentheses. Huber-White robust standard errors reported. 1 F-test of exogeneity of second-stage regressors. 2 For a worst-case size of 10% for a 5% Wald test of the 2SLS estimates.
Estimated increase in probability of civil conflict when moving from 0 to the median value of petroleum fields among producers
Columns 1 and 2 of Table I display the results for the relationship between oil and all types of civil conflict. First, we note that sediment volume is an exceptionally strong instrument for petroleum fields, as evidenced by its large, positive, and highly significant coefficient in the first-stage regression. Further, comparing the F-statistic of the first-stage instrument against the critical value provided in Stock & Yogo (2005) clearly rejects the hypothesis that the instrument is weak. 10 Substantively, once instrumented, the effect of oil on conflict is positive, but only marginally significant (p = 0.099). Thus, though there is some evidence of a causal oil curse (H1), it is associated with considerable uncertainty.
The source of this ambiguity is revealed once we distinguish between conflict types. In line with H2, we find a strong effect of oil on secessionist conflict. The corresponding IV estimate, reported in Column 4, is positive and highly significant. Furthermore, the causal effect of oil on territorial conflict is large: a country featuring a median level of oil fields would have had a 35 percentage-point lower probability of experiencing territorial conflict had it not extracted any petroleum (see Panel B of Figure 4).
Comparing the IV result with its uninstrumented counterpart in Column 3 shows that the positive relationship between petroleum and territorial conflict is only revealed once oil is instrumented. Because countries at risk of secessionist wars tend to see less oil production, we underestimate the strength of the oil–conflict link without proper causal identification. The presence of endogeneity is further confirmed by a Hausman-like test (Wooldridge, 2002: 119), which rejects the null of exogenous second stage regressors at the 0.1% level.
Next, Column 6 reveals that there is little evidence of a causal relationship between oil and governmental conflict. Though the IV estimate is positive, it is associated with considerable uncertainty, and thus fails to attain statistical significance at conventional levels. Finally, as postulated by H3, the results summarized in Column 8 suggest that only populated oil fields have a significantly positive effect on secessionist conflict. In fact, while the estimate associated with populated fields is 1.5 times larger than the respective coefficient for all types of oil fields reported in Column 4, the estimate for unpopulated fields is negative. 11 A possible interpretation of this result is that as long as oil production does not cause any grievances locally, it may help governments to prevent secessionist conflict elsewhere, either via repression or cooptation.
Finally, Section A.3.2 of the Online appendix shows that these results are highly robust to alternative specifications. We obtain substantively equivalent results when using a different IV method, adding various geographic controls, using petroleum measures based on PETRODATA or Horn, and employing panel data covering the entire period between 1950 and 2013. 12 Moreover, using the inferential procedures proposed by Conley, Hansen & Rossi (2012), we show that the positive causal link between onshore fields and territorial conflict remains intact even under moderately sized violations of the exclusion restriction (see Section A.3.3).
In summary, this section yields three important insights. First, focusing on secessionist conflict, there is strong evidence that the oil curse is causal. In contrast to the literature that rejects the oil–conflict link, we find that uninstrumented models tend to underestimate the true risks associated with oil production. Second, in support of H2, the oil curse appears to be limited to secessionist conflicts. The null of no oil effect can easily be rejected for territorial conflicts, but not for governmental conflicts. And third, in support of H3, whether oil production increases the risk of secessionist violence depends on where it occurs. This result is clearly at odds with the governmental oil curse, which expects oil-rich states to experience conflict more frequently regardless of where oil extraction takes place. Indeed, we find that oil may even inhibit secessionist violence if extraction occurs in uninhabited regions.
Is there an ethno-regional oil curse?
After having established that there is evidence of an oil curse, we now assess whether it is indeed ethno-regional. To do so, we analyze data where the units of analysis are politically relevant and geographically concentrated ethnic groups. We identify these groups using the Ethnic Power Relations (EPR) (version 2014, Vogt et al., 2015) and GeoEPR (version 2014, Wucherpfennig et al., 2011) datasets. EPR enumerates politically relevant ethnic groups for all sovereign countries across the globe from 1946 to 2013, and GeoEPR geocodes corresponding settlement areas.
As in the country-level analysis, we focus on the period between 1991 and 2013. Group-level conflict is identified with the ACD2EPR dataset (version 2014, Wucherpfennig et al., 2012), which links the UCDP ACD (version 4-2014, Pettersson & Wallensteen, 2015) conflict data to the EPR ethnic groups by coding whether the rebel organizations involved in these conflicts made ethnic claims and recruited from a given group. The intensity of petroleum production within each group’s territory is captured by the logged count of the ACOR fields within the groups’ settlement areas (after adding a unit constant).
To instrument oil production, we again rely on the CRUST data to calculate the log of sediment volume per group settlement area, while controlling for group-level surface area and mountainous terrain. Finally, information on whether groups are represented in government (for testing H4), as well as their demographic size (for testing H5), is obtained directly from EPR. 13
We adopt the same econometric approach as in the previous section, with the key difference that we add country-level fixed effects, thus effectively removing all cross-country variation. This step ensures that we do not simply pick up a country-level effect, and addresses country-specific error dependence.
Table II displays the group-level results, presenting both OLS and 2SLS coefficients. Models 1 and 2 test whether the strong relationship between oil and territorial conflict found in the previous section is also detectable at the group level. As anticipated, the 2SLS estimates indicate that the effect of group-level oil production on the probability of territorial conflict is positive and significant. Thus, the oil–conflict link is indeed ethno-regional: groups that host oil production are more likely to engage in secessionist violence. Furthermore, there is strong evidence of endogeneity: comparing the results of Columns 1 and 2 shows that the oil–conflict link is only identifiable with the IV design. Moreover, we can reject the null of exogenous second stage regressors at the 1% level. Thus, petroleum production appears to be less likely in the territory of groups that are particularly conflict-prone, causing unadjusted research designs to underestimate the severity of the oil–conflict link.
(Instrumented) petroleum and territorial conflict, group-level, 1991–2013
***p < 0.001, **p < 0.01, *p < 0.05. Huber-White robust standard errors in parentheses. All models estimated with country FEs. 1 F-test of exogeneity of second-stage regressors. 2 For a worst-case size of 10% for a 5% Wald test of the 2SLS estimates.
Next, we re-evaluate H3 on the group level, testing whether oil production in populated areas is particularly likely to cause secessionist conflict. For this purpose, we recalculate the (un-)populated petroleum field, sediment volume, and surface area variables described in the previous section (see Sections A4.1.4–A4.1.5 of the Online appendix). Note that this analysis is only feasible because the GeoEPR settlement polygons often include unpopulated regions, as areas between population centers are often counted towards a nearby ethnic group. As expected, whereas the coefficient associated with (instrumented) populated fields is positive and significant, the estimate for (instrumented) unpopulated fields is negative (Column 4 of Table II). However, an F-test reveals that we can only reject the null of equal effects at the 10% level (p = 0.059). Thus, in contrast to the country-level results, here we only find limited evidence for H3. This ambiguous finding may be due to the small size of many ethnic settlement regions, and the resulting high correlation between the populated and unpopulated sediment volume estimates for these areas.
In a final step, we evaluate H4 and H5. Using suitably instrumented interaction terms, we test whether political exclusion and relative demographic size (measured in 1991) mediate the effect of oil on secessionist conflict. Moreover, given the above findings, we focus exclusively on petroleum fields in populated areas. Columns 5 and 6 report the results for political exclusion. In agreement with H4, we find that oil is significantly more likely to cause secessionist violence for excluded groups. Panel A of Figure 5 plots the effect of the populated fields within a group’s territory on the probability of territorial conflict as a function of exclusion. While the effect for groups that are represented in government is only borderline significant (p = 0.052), the effect more than doubles in size and attains statistical significance (p = 0.001) for excluded groups.
Marginal effect of (logged) populated petroleum fields on the probability of group-level territorial conflict, conditional on exclusion (Panel A) and group size (Panel B)
Despite these promising results, there is an important caveat to this analysis. While ethnic groups’ location and demographic size are convincingly exogenous to conflict, this is not necessarily the case for political exclusion. Importantly, if governments systematically exclude conflict-prone groups, we overestimate the conditioning effect of exclusion. However, recent research suggests that the opposite is the case. Wucherpfennig, Hunziker & Cederman (2016) show that governments tend to include threatening groups, which would imply that we underestimate its mediating role.
Finally, Columns 7 and 8 show the results obtained when conditioning on relative group size. The point estimates of the 2SLS estimated model (Column 8) strongly support H5: while the populated petroleum field estimate is positive and significant, the interaction parameter is negative and significant. Panel B of Figure 5 plots the effect of oil production in populated areas on the yearly probability of territorial conflict as a function of relative group size. Visual inspection confirms that the oil–conflict link is associated primarily with ethnic minorities: we can only reject the null of no effect for groups with a relative size of less than 0.5.
How robust are these group-level findings? Section A.4.2 in the Online appendix shows that adding geographic controls, using PETRODATA or Horn-based petroleum measures, and employing panel data covering the longer 1950–2013 time period all lead to equivalent results. Moreover, mirroring the country-level results, we show that the group-level link between onshore fields and territorial conflict also remains significant under moderately sized violations of the exclusion restriction (see Section A.4.3).
In summary, this section yields four important results. First, there is a causal oil–conflict link at the ethno-regional level. Even at the subnational level, we find clear evidence in favor of a causal effect of oil and gas production on the likelihood of secessionist violence. Second, oil extraction appears to provoke secessionist violence only where local residents are immediately affected and, thus, local claims to resource ownership are particularly likely. As in the country-level analysis, there is no evidence of an oil–conflict link for petroleum production in unpopulated areas. Third, small and excluded ethnic groups are significantly more likely to react to petroleum extraction with violent secessionism. And finally, again, we find considerable evidence of endogeneity: conflict-prone ethnic groups host fewer productive petroleum fields than other groups, even in a within-country comparison.
Conclusion
Despite its prominence in theory and policy, the claim that oil abundance causes civil war continues to generate controversy. Major progress has been made, but there is still no consensus as regards the causal mechanisms driving the link, and some scholars even question its very existence. This article aims to overcome this fundamental ambiguity by addressing three outstanding difficulties that afflict the current literature.
First, we reaffirm the existence of a causal ethno-regional oil curse that depends on affected groups’ political and economic grievances and material aspirations. Oil production increases the risk of violent secessionist conflict if it occurs within the settlement area of politically and demographically marginalized ethnic groups. Petroleum extraction in other locations, however, appears to be largely inconsequential, or may sometimes even inhibit the outbreak of violence. In combination, these findings substantiate our theoretical argument, and are largely inconsistent with those explanations of the oil–conflict link that focus exclusively on governmental or individualist mechanisms.
Equally importantly, the current article breaks new empirical ground by introducing an identification strategy that improves on existing attempts to handle endogeneity. Relying on sedimentary basins as an instrumental variable allows us to circumvent reverse causation and other sources of endogeneity. While showing that uninstrumented analyses have underestimated the effect of resource abundance, the sediments-based instrument lends strong support to our ethno-regional explanation.
Third and finally, we introduce the new ACOR dataset that offers precise historical information on the location of oil and gas fields. These empirical advances still leave plenty of room for future progress in terms of data collection and integration. An important area of future research pertains to economic inequality, which we have chosen not to measure directly in this article, primarily because it is very difficult to separate household income from economic activities associated with the oil industry itself.
Pending such advances, the current article offers considerable evidence to suggest that conventional accounts of the oil–conflict link will have to be reassessed. In this light, attempts to prevent and resolve petroleum-related civil wars will need to pay more attention to reducing political inequalities than conventional policy recommendations that attempt merely to strengthen state capacity or to buy off potential protesters. This calls for increased pressure to be put on oil-producing states to induce them to redress, rather than repress, oil-related grievances. This message is all the more important as new oil fields come on line in peripheral areas of developing countries, especially in sub-Saharan Africa.
Footnotes
Acknowledgements
We are grateful to the participants of the Political Economy of Inequality and Conflict workshop at the University of Konstanz and the brown-bag participants at PRIO for their extremely useful comments and suggestions on earlier drafts of this article. We would also like to thank Lukas Dick for his help with compiling the ACOR data. Any remaining errors are the authors’ responsibility.
Replication data
Funding
This research was supported by the Swiss National Science Foundation under COST action IS1107, SERI project C12.0087.
Notes
References
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.
