Abstract
Elections are regularly held in countries facing ongoing civil conflicts, including in India, Iraq, Nigeria, the Philippines, and Ukraine. Citizens frequently go to the polls having endured years of violence between armed groups and governments. A growing literature questions how violence conditions voters’ support for incumbents versus challengers, and for hawks versus doves. We analyze this relationship in the context of the 2014 presidential election in Colombia, an election defined by candidates’ positions on negotiations with the country’s largest insurgent group, the FARC. Our results show an inverted-U relationship between past insurgent violence and vote share for President Juan Manuel Santos, the pro-peace candidate: he performed better in communities with moderate levels of insurgent violence and poorly in communities with both very high and very low violence. We also find that areas where the FARC originally mounted attacks 50 years ago more strongly supported Santos. The article concludes by comparing these results with past studies of violence and vote choice in Israel, Turkey, and Spain.
1. Introduction
Elections are regularly held in countries facing ongoing civil conflicts, including in India, Iraq, Nigeria, the Philippines, and Ukraine. Citizens in conflictual societies frequently head to polling stations having endured years of communal or individual targeting by both illegal armed groups and governments. Such elections present us with the paradox of citizens democratically choosing their political leaders even as armed groups violently contest the very same question. Following scholarship on the broader relationship between elections and violence (e.g. Snyder 2000; Varshney 2002; Brass 2003; Wilkinson 2004; Bekoe 2012; Hafner-Burton and Jablonski 2014; Staniland 2015), a new literature has focused on how violence affects voters’ choices in the polling booth (Berrebi and Klor 2006; Bali 2007; Berrebi and Klor 2008; Gassebner et al. 2008; Gould and Klor 2010; Montalvo 2010; Dunning 2011; Kibris 2011; Birnir and Gohdes 2012; Getmansky and Zeitzoff 2014). This newer literature proposes several different mechanisms for how violence within civil war affects vote shares for incumbents versus challengers, and hawks versus doves.1
We explore a number of these mechanisms in the context of the 2014 presidential elections in Colombia. Very few elections hinge so completely on candidates’ positions on peace talks as Colombia’s did in 2014. Candidates in this election offered two very different visions for their country’s future, with campaigns nearly completely defined by their position on the civil war, the longest ongoing conflict in the Western Hemisphere. More specifically, the campaign turned on whether and how the government should negotiate with the Revolutionary Armed Forces of Colombia (FARC), the country’s oldest guerrilla group. The campaign’s singular focus on the peace process allows us to assess how historical patterns of violence by guerrillas conditioned communities’ support for peace, and to compare Colombians’ choice with similar choices made by civilians in Israel, Peru, Spain, and Turkey.2
Our quantitative tests attempt to adjudicate statistically among explanations prominent in the literature. We find that Colombian President Juan Manuel Santos, who opened formal negotiations with the largest insurgent group, the FARC, in February 2012, performed poorly at the polls where insurgents have historically committed high levels and low levels of violence, but did well in places where there were moderate levels of insurgent violence. That is, his vote share followed an inverted U-shape across levels of FARC violence. Vote shares for the anti-negotiation candidate, Óscar Iván Zuluaga, in contrast, display the opposite pattern: he performed far better in communities with very high and very low levels of violence, yet his vote share was small in areas with moderate prior levels of violence.
We proceed as follows. We begin with a synopsis of the 2014 presidential campaign in Colombia. Next, we tease out observable implications of various plausible mechanisms connecting legacies of violence to vote shares. We then present our quantitative data and provide the empirical results. Finally, we consider issues of generalizability and discuss paths for future research.
2. The 2014 Colombian presidential elections
The 2014 Colombian presidential elections were dominated by the question of whether and how to negotiate with the FARC. The election featured two former cabinet ministers under the popular and hardline former president, Álvaro Uribe. The incumbent president, Juan Manuel Santos, had served as defense minister and helped craft Uribe’s “firm hand” (mano dura) policies in the struggle against insurgents. Santos therefore surprised many when he announced in October 2012 that his government had commenced negotiations with the FARC. As the Cuba-hosted talks made slow but steady progress, Santos bet his political future on a “peace agenda” including negotiations with the FARC, reparations for victims of violence, and land restitution for displaced populations. A vote for Santos, then, was effectively a vote for negotiations, although he continued military operations during the talks. Santos’s opponent was Óscar Iván Zuluaga, himself a former finance minister during Uribe’s presidency. Zuluaga proposed a hawkish platform that included suspending negotiations with the FARC until the group met key demands, including a unilateral ceasefire.3 Zuluaga was endorsed by Uribe, the de facto head of Zuluaga’s Democratic Center party.
With the conflict front and center in the campaign, Colombians went to the polls on May 25. Zuluaga scored a major victory in the first round, receiving 29.3% to Santos’s 25.7%, with other parties’ candidates garnering anywhere from 8% to 16% of the vote share each. A second round was held on 15 June, since no candidate won 50% of the vote share. The run-off saw a victory for Santos, by a 5% margin. He seemed to benefit more from the support of smaller parties, especially from the leftist Democratic Pole in Bogotá and the Liberal Party in Valle del Cauca and the Caribbean coast.
3. Theory
Very few elections hinge so completely on candidates’ positions on peace talks as Colombia’s in 2014. Yet Colombia’s election is hardly unique, as voters often vote in the midst of violence. A new literature has recently considered whether violence by non-state groups tips the electoral balance, whether between hawks and doves or incumbents and challengers (Dunning 2011). They have examined several potential mechanisms. First, some have argued that violence raises voters’ doubts regarding the incumbent government’s competence (Ferejohn 1986). Gassebner et al. (2008) find, for example, that terrorist attacks raise the probability of government replacement in democracies. Montalvo’s (2010) analysis of voting in the 2004 Spanish elections, which took place after a terrorist bombing in Madrid, echoes this finding: Spanish nationals living abroad who voted before the bombing were far more likely to vote for the government. Kibris (2011) examines the response of Turkish voters to military and police casualties resulting from terrorism by the Kurdistan Workers’ Party (PKK), finding that Turkish voters living in districts with higher casualties more likely voted against the incumbent government.
Second, scholars have questioned the impact of violence on voters’ preferences for concessions to perpetrators and support for hawkish politicians and parties. Berrebi and Klor (2006, 2008) find that Israelis living in localities suffering from terrorist attacks within three months of an election were more likely to vote for right-wing political parties. Getmansky and Zeitzoff (2014) agree, finding that exogenous expansions in Palestinian rocket barrages increased the vote shares for more hawkish political parties in Israeli elections. Kibris (2011) finds that Turkish voters exposed to violence supported right-wing parties opposed to concessions to the PKK. Psychologists have also noted that prolonged exposure to violence can create a propensity to support more militant policies and parties (Canetti et al. 2013; Lavi et al. 2014).
Not all scholarly analysis, however, agrees that violence necessarily nudges citizens towards more belligerent parties. Pape (2003) argued that suicide terrorism can succeed by placing popular pressure on democratic governments for concessions. Bali (2007) and Montalvo (2010) find that the 2004 Madrid train bombings decreased support for the incumbent Popular Party due to its hawkish policy on Iraq. Gould and Klor (2010) actually suggest a non-linear relationship between terrorism and support for hawkish political parties in Israel; low and moderate levels of attacks induced Israelis to support concessions to Palestinians, but high levels of attacks reduced support for concessions. Birnir and Gohdes (2012), meanwhile, suggest a more contingent impact of civil war violence on vote choice, based on their research in Peru: local victims of violence punish parties associated with perceived perpetrators, while voters isolated from such events blame incumbent governments.
Colombia’s 2014 presidential elections, a virtual referendum on the peace process with the FARC, provide an excellent laboratory in which to study these dynamics. As in Turkey and Israel, Colombia’s continuing conflict represents a highly salient political issue, one which clearly divided the political establishment between more hawkish and more dovish candidates. Previous scholarship suggests three mechanisms connecting FARC violence and Colombians’ vote choices. First, insurgent violence might have increased support for the candidate opposed to peace talks (Zuluaga) and decreased support for the pro-negotiation candidate (Santos). Scholars suggest that voters punish incumbents for security failures: we should expect that they especially do so in areas with more frequent insurgent attacks. Evidence from Israel (Berrebi and Klor 2006, 2008; Getmansky and Zeitzoff 2014), Turkey (Kibris 2011), and Peru (Birnir and Gohdes 2012) also implies that FARC violence should encourage voters to support more hawkish candidates. Santos’s vote share should have been lower in areas with high levels of insurgent violence, while Zuluaga’s should have been higher, as Colombians punish Santos both for his incumbency and his support for peace.
Second and conversely, insurgent violence might have driven support for Santos while reducing Zuluaga’s. Pape (2003) argues that campaigns of suicide terrorism encourage concessions in democratic countries. Evidence from Spain suggests that voters punish hawkish parties when their policies are blamed for continuing violence (Bali 2007; Montalvo and Reynal-Querol 2010). Given this logic, FARC attacks might have convinced those who witnessed them that the likelihood of an outright military victory by the state was rather slim. They might have reacted poorly to Zuluaga’s insistence on essentially ending talks. War-weary Colombians may also respond not with a desire for retribution and continued conflict, but in favor of truth-telling and material compensation for victims (Rettberg 2008). Areas with more frequent FARC attacks, then, might have proved more supportive of Santos’s political platform.
Finally, Gould and Klor (2010) suggest a non-linear relationship between violence and vote shares for hawks and doves: moderate levels of violence will push voters towards concessions but very low and very high levels of violence will leave voters more supportive of a more militant response. The same logic might apply to Colombia’s 2014 election. Santos’s message of peace may have fallen on deaf ears in areas with very low violence. Birnir and Gohdes (2012) suggests that such voters will punish incumbents for violence elsewhere in the country. Areas with moderate levels of violence may have more likely supported concessions to guerrillas, yet Santos’s pro-peace campaign may have struggled to convince voters subjected to high levels of brutality of the benefits of negotiations. This logic suggests that Santos’s vote share should follow an inverted-U pattern: a lower vote share in areas with very low or very high insurgent violence and a higher vote share in areas of medium levels of violence. Zuluaga’s vote share would follow the opposite pattern: greater support in areas with very low or very high violence and less in areas with moderate levels of violence.
A final potential explanation turns to path-dependent patterns of violence and civilian political support for insurgents. The FARC originated in armed peasant groups seeking to defend land rights; Flores (2014), for example, cites evidence that the early “independent republics” created new patterns of collective action among peasants. This civilian support, dubbed “social endowments” by Weinstein (2007), proved essential to the FARC’s survival during its first 20 years and predicts well areas of future activity (Daly 2012). Early insurgent violence likely occurred in areas where the FARC could access this deep-rooted civilian support. These areas might remain political strongholds of the FARC and, if so, we might expect support for the peace process and President Santos to be higher in such areas.4
4. Data
Electoral data were obtained at the municipality level from the Registraduría Nacional de Estado Civil, the governmental body that implements elections. We focus on the run-off returns given that (a) third parties had an important effect upon vote shares in the first round (the Conservative Party drew votes from Zuluaga, the Democratic Pole and Green Party from Santos) and (b) abstention declined 8% from the first to the second round.
Dependent variables for the statistical analysis are Santos Vote Share and Zuluaga Vote Share, the percentage of total votes in a municipality cast for each candidate in the run-off. Although there are only two candidates, we show results for both due to variance in the prevalence of the “blank vote” (voto en blanco), a protest vote against both candidates.
We code independent variables that distinguish among explanations discussed above. Violence data are taken from the Human Rights Observatory Database, compiled by the Presidency of Colombia and based on daily bulletins internally compiled by the Colombian government’s security agency. Note that this violence dataset should be considered a “convenience dataset” and, as such, it cannot be assumed or demonstrated that the data are representative of the true population of episodes of lethal violence in Colombia (e.g. Krüger et al. 2013). For additional details regarding data collection procedures and the reliability of the data, as well as results using a different violence dataset, see the Supplementary Appendix.
Our first independent variable is the total number of attacks carried out by the FARC between 1988 and 2010 in each municipality, Cumulative guerrilla attacks. We then use a quadratic term, Cumulative guerrilla attacks squared, to test non-linear effects.5 In the Supplementary Appendix we show robustness to measuring FARC violence using different temporal windows. We focus on FARC violence, rather than paramilitary violence, because the elections were a de facto referendum on the peace process with the FARC, and because the official demobilization of paramilitaries in the mid-2000s left behind fragmented criminal gangs lacking political objectives and more single-mindedly devoted to the drug trade. As such, coding paramilitary activity post-2005 is prone to measurement error.6
Our final explanation predicts that even earlier periods of armed mobilization will affect current vote choice. Using data from Daly (2012), Insurgent violence before 1970 is a dummy that takes the value of one if there were any violent events perpetrated by liberal and communist guerrillas in a municipality between 1964 and 1970, the earliest period of FARC’s mobilization.
We use an ordinary least squares estimator and control for factors that might influence both insurgent attacks and electoral support. These controls include the log of municipal population, area, altitude, and average annual rainfall, all of which condition recruitment and the effectiveness of state policing (Fearon and Laitin 2003); variables that account for the municipality’s relative economic isolation, such as distance to the departmental capital and an index accounting for the proportion of a municipality that is rural, as both vote shares and rebellion are expected to be conditioned by state reach and the dynamics of economic activity (e.g. Kalyvas 2007). We include the number of judicial institutions, social and economic variables such as gross enrollment in primary and secondary education, and the poverty rate, which also help proxy state capacity and development (Hendrix 2010). We control for the availability of legal and illegal rents like per capita tax revenue, per capita royalties received by the municipality for natural resource exploitation, and the share of the municipal area under cultivation by coca crops, given that such resources make rebellion both more attractive and more viable (e.g. Weinstein 2007). We also include inflow of internally displaced individuals, as this may both change the composition of communities and increase disputes over land, incentivizing civilians to use armed groups to settle personal scores (Kalyvas 2006).7 Department fixed effects are used to account for unobserved heterogeneity, provided that heterogeneity is common to all municipalities within a department.
5. Results
Model 1 in Table 1 demonstrates that without controls Cumulative guerrilla attacks is positively and significantly correlated with President Santos’s vote share. The effect size is nearly the same in Model 2, which includes controls. The opposite result can be seen for Zuluaga in Model 3, which includes no controls, and Model 4, which does: Cumulative guerrilla attacks is statistically significant and negatively correlated with Zuluaga’s vote share. These results clearly support the second explanation for the link between violence and voter choice proposed above, that insurgent violence encourages support for conciliatory candidates, not the first explanation, that the public punishes incumbents for security failures.
Guerrilla activity and vote shares.
Regarding substantive effects, a one standard deviation increase in Cumulative guerrilla attacks (equivalent to 5.6 attacks) is associated with a 1.6% increase in vote share for Santos in a given municipality,8 and a 1.5% decrease in Zuluaga’s vote share.9 To provide some context, a number of Colombian elections were sufficiently close such that even small margins of difference could have swung the election. In the first round of the 1994 election, for example, the difference between Ernesto Samper and Andrés Pastrana was 0.3%, or 18,000 votes, while in the first round of the 1998 presidential election, Horacio Serpa and Andrés Pastrana’s vote totals were separated by 0.4%, a difference of only 43,000 votes. Given that we are examining a long temporal period of 22 years, one additional attack every other year in a given municipality would correspond, ceteris paribus, to a 3% change in local vote share. In short, the substantive size of these local effects is large, and could scale up quickly at the national level to change electoral outcomes.
Table 2 explores potential non-linear effects. Models 1 and 2, demonstrate an inverted-U-shaped relationship between FARC violence and vote share for Santos: Cumulative guerrilla attacks squared is highly significant and indicates that Santos received fewer votes in low violence municipalities, more votes in municipalities with moderate histories of violence, and fewer votes in municipalities with high levels of prior FARC violence. Models 3 and 4 show the opposite effect for the challenger, Zuluaga: his vote share was strongest with low and high levels of prior FARC violence, and lowest where moderate FARC violence occurred. This echoes findings from Israel which demonstrate a non-linear relationship between violence voting (Gould and Klor 2010).
Guerrilla activity and vote shares: exploring non-linearities.
Figures 1 and 2 display predicted vote shares for both candidates given increases in Cumulative guerrilla attacks, as well as a histogram of intensity of attacks. Results are taken from Models 2 and 4 in Table 2, with other variables held at their means. Figure 1 demonstrates that Santos lost in municipalities with low levels of violence, and won where there were medium and high levels of violence. Because most municipalities experienced low levels of violence, confidence intervals are large for high-violence municipalities. The opposite holds for Zuluaga, displayed in Figure 2.
Non-linear effect of legacies of violence on Santos vote share.
Non-linear effect of legacies of violence on Zuluaga vote share.
Finally, Table 3 provides evidence for the long-run effect of violence on vote patterns. Pre-1970 insurgent violence is positively associated with vote share for President Santos. A municipality that experienced insurgent activity prior to 1970 had an almost 3% higher vote share for Santos. In contrast, such a municipality saw 2.8% fewer votes for Zuluaga.
Historical guerrilla activity and vote shares.
Our findings should be interpreted as suggestive correlations rather than causal relationships. Finding a suitable instrumental variable in this context is challenging, let alone designing an experiment. Yet we have taken several steps to alleviate endogeneity concerns. First, our coding of insurgent violence is temporally prior to the dependent variable, ending 4 years before the vote in the closest proximity case. Second, our results survive the inclusion of control variables and department fixed effects, which capture unobserved heterogeneity. However, we cannot rule out omitted factors that may account both for prior violence and current electoral outcomes. For instance, if the unobserved underlying political affiliation of a municipality determines both past violence and current vote shares, our estimates would be biased. Alternatively, FARC violence could be driven by expectations of future electoral outcomes. We think confounding stories of this sort are unlikely, but we cannot rule them out.
6. Conclusion
Colombian voters faced a clear choice in the 2014 presidential election: a vote for the incumbent and more dovish President Juan Manuel Santos meant a continuation of the peace process, while a vote for the more hawkish challenger, Óscar Zuluaga, likely meant an abandonment of talks. Our findings suggest that insurgent violence conditioned Colombians’ votes. The simplest interpretation suggests that Zuluaga performed better in municipalities with lower violence and vice versa, because Colombians living in hard-hit areas were convinced of the need for peace. Our findings also suggest that Santos’s vote share suffered in areas with both very low and very high levels of violence, suggesting a more nuanced relationship between violence and vote shares. The results thus support Gould and Klor’s (2010) research on the Israeli–Palestinian conflict, while diverging from arguments that contend that areas experiencing higher violence should more likely vote against doves and incumbents (Berrebi and Klor 2006, 2008; Kibris 2011). Finally, we found evidence that the FARC’s armed mobilization in the mid-1960s had a lasting effect on voting patterns: areas witnessing FARC attacks before 1970 were more likely to vote for Santos than Zuluaga.
The attributes that made Colombia’s 2014 presidential election an ideal test of the relationship between bullets and ballots should caution against an overly eager application elsewhere. Colombia’s election featured two candidates with diametrically opposed positions on the peace process yet similar positions on other issues in a campaign focused on the peace process. This may be a rare occurrence. We suggest several factors likely to condition this relationship.
The salience of civil conflict to voters will almost certainly vary. India’s 2014 elections, which produced an historic defeat for the Congress Party, took place while Naxalites continued to rebel against the Indian state, yet this rebellion remained relatively contained geographically, with slow economic growth instead taking center stage in the campaign. The effect of legacies of violence on voting may also depend upon a conflict’s “technology of rebellion” (e.g. whether it is fought as an insurgency or as a symmetric conflict, whether it features terrorism, the technological sophistication of each side) (Kalyvas and Balcells 2010). The link between vote shares and violence may also depend on whether ethnic, religious, or ideological identities are at issue. Sunni and Shia voters in Iraq, for example, may be more likely to vote for parties representing their sect given continuing violence mobilized along this cleavage and the “stickiness” of ethnic identities.
Our investigation raises several questions for future research. Comparing patterns of violence and vote shares in other contexts would help address questions surrounding generalizability. Researchers might also seek to conduct voter surveys ahead of elections in conflict-affected societies, especially if methods can be used to allow for more honest responses (e.g. Corstange 2009), as a means to tease out the different causal mechanisms we and other researchers have identified.
Article Notes
Declaration of conflicting interest The authors have no conflicts of interest to declare.
Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Supplementary material The online appendix is available at: http://rap.sagepub.com/content/by/supplemental-data
Notes
↵1. Another possibility is the strategic use of violence by armed actors to kill or displace political opponents (Steele 2011).
↵2. We return to the question of generalizability in the conclusion.
↵3. The candidate softened his stance after the election’s first round to gain the Conservative party’s support.
↵4. Territorial control by armed groups might also help explain subsequent candidate vote shares (e.g. Kalyvas 2006), but data limitations regarding territorial control make direct testing nearly impossible.
↵5. Our treatments (insurgent violence) and outcomes (vote shares) are measured at the municipal level. We therefore avoid the ecological fallacy of comparing individual-level voting preferences with municipality-level violence (e.g. King et al. 2004).
↵6. In the Supplementary Appendix we show robustness to the inclusion of paramilitary attacks as a control. On insurgent defection to paramilitaries in Colombia see Oppenheim et. al. (forthcoming).
↵7. Results presented below are also robust to including municipal-level vote shares for Andrés Pastrana, the victorious pro-peace candidate in the 1998 presidential election, as well as outflow of internally displaced persons. These appear in the Supplementary Appendix.
- © The Author(s) 2015
This article is distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License (http://www.creativecommons.org/licenses/by-nc-nd/3.0/) which permits non-commercial use, reproduction and distribution of the work as published without adaptation or alteration, without further permission provided the original work is attributed as specified on the SAGE and Open Access page (http://www.uk.sagepub.com/aboutus/openaccess.htm).