Civil wars cause humanitarian crises and change the course of state development, with generational effects (Acemoğlu and Robinson Reference Acemoğlu and Robinson2012; Black Reference Black2013; Englehart Reference Englehart2009; Herbst Reference Herbst1996). As these events can be so catastrophic, many have sought to better understand the drivers of civil war to anticipate their onset, using a range of methodologies (Fearon, Kasara, and Laitin Reference Fearon, Kasara and Laitin2007; Goldsmith and Butcher Reference Goldsmith and Butcher2018; Goldstone et al. Reference Goldstone2010; Hegre et al. Reference Hegre2017). Quantitative predictive methods have shown that models are able to generate better-than-random predictions about future conflict onset, an important finding for researchers and policymakers (Goldstone et al. Reference Goldstone2010; Hegre, Nygård, and Landsverk Reference Hegre, Nygård and Landsverk2021), though temporal variability in predictive accuracy should contextualize model application (Bowlsby et al. Reference Bowlsby2019).

Civil war onset has various established “drivers,” including patterns of human development (Fearon and Laitin Reference Fearon and Laitin2003; Fox and Hoelscher Reference Fox and Hoelscher2012; Goldstone Reference Goldstone2008), economic development (Alesina and Rodrik Reference Alesina and Rodrik1994; Muller and Seligson Reference Muller and Seligson1987; Sambanis Reference Sambanis2004), horizontal inequality (Cederman, Weidmann, and Gleditsch Reference Cederman, Weidmann and Gleditsch2011; Forsberg Reference Forsberg2014; Rüegger Reference Rüegger2019), and political institutions (Goldstone Reference Goldstone2002; Gurr Reference Gurr1993; Saideman et al. Reference Saideman2002). These variables are typically used to predict conflict civil war onset by using a “single model” approach that relies upon a set of independent variables with a temporal lag.

However, an examination of onset cases used in quantitative models shows that civil war cases are caused by a wide range of factors (Marshall, Gurr, and Harff Reference Marshall, Gurr and Harff2021). For example, the UK experienced civil conflict in 1971 driven by state-led discrimination against Irish Catholics by Ulster paramilitaries and British armed forces (Loyle, Sullivan, and Davenport Reference Loyle, Sullivan and Davenport2014). Contrast this with the roles of agitated youth and regional instability in the 2011 Egyptian Revolution (Salih Reference Salih2013) or the 1996 insurgency in Nepal, partly attributable to high rural inequalities (Murshed and Gates Reference Murshed and Gates2005).

While global quantitative models have demonstrated utility in predicting civil war onset, the wide variation in drivers suggests that there may be unique patterns or risk profiles that are common to some, but not all, onset cases. It may be that there is not one set of universal structural factors driving instability, but alternative sets of factors that drive vulnerability, depending on a particular constellation of factors. Rather than identifying the “gas on the rag” (O'Brien Reference O'Brien2002), it may be the case that there are different types of flammable substances that lead to civil war onset that are particular to levels of development, or a confluence of factors unique in some cases but not others, rather than universal factors that explain all failures.

In this article, we use cluster analysis to identify patterns in factors that may correlate with civil war onset. We ask: is there evidence that persistent and unique clusters of drivers of civil war exist, calling into question the use of “single models” to predict onset? If we find no persistent driver clusters, that provides further evidence that “single model” approaches to predicting onset should continue to be the focus of research in this field, possibly strengthening their prominence in both research and policymaking communities.

We begin by creating a “base” model, using agglomerative clustering on six theoretically motivated variables that have been shown historically to predict civil war onset, finding three persistent clusters. We test how persistent these clusters are by using alternative model specifications. We find a persistent cluster of drivers of civil war onset associated with young, poorly developed, and anocratic countries, a grouping of onset cases that is persistent across all model specifications. We also find a cluster of countries with above-average levels of development in consolidated regimes that suffer moderately higher levels of horizontal inequalities. This suggests that wealthier, politically consolidated states more commonly have civil conflict onset associated with minority repression. Our third cluster is characterized by higher neighborhood conflicts, younger populations, and higher levels of horizontal inequalities.

Research Design

Clustering is an unsupervised quantified classification technique that aims to discover patterns within data based on the similarity of attributes between observations. All clustering techniques follow similar steps: (1) scaling input data; (2) calculating the “distance” between observations; (3) using clustering algorithms, which help determine how clusters are constructed based on the “distance” between observations; and (4) using the output of the preceding analyses to identify an appropriate number of clusters. Each step can be completed using several techniques. For our base model, we elected to normalize the data using a min–max, 0–1 scale. Next, we used Euclidean distance to measure the similarity between each onset case in a multivariate space. Euclidean distance calculates the square root of the sum of the squared differences between input variables for any two observation pairs and is particularly appropriate for noncategorical variables (Abbott Reference Abbott2014).

Thirdly, we used an agglomerative hierarchical clustering algorithm (more commonly, “Agnes”) to group observations. Agnes clustering identifies local patterns in observations, building from the “ground up” by identifying smaller groups of observations that are then merged into bigger groups at each clustering stage (Kaufman and Kousseeuw Reference Kaufman and Kousseeuw2009). Alongside the Agnes clustering algorithm, we used Ward's (Reference Ward1963) linkage method to determine the distance between clusters because it produces more refined and balanced clusters (Wolfson, Madjd-Sadjadi, and James Reference Wolfson, Madjd-Sadjadi and James2004).

Determining the appropriate number of clusters can be achieved using various approaches, including visual heuristics, quantitative indices, and confirmation through domain knowledge (Abbott Reference Abbott2014; Kodinariya and Makwana Reference Kodinariya and Makwana2013). We primarily relied upon an approach that creases thirty indices that evaluate different aspects of model behavior using the NbClust R package (Charrad et al. Reference Charrad2020).Footnote ¹ The output consistently pointed to the use of three clusters in this analysis.

We used the Political Instability Task Force's (PITF) event data for our onset cases (Marshall, Gurr, and Harff Reference Marshall, Gurr and Harff2021). Our case selection criteria included ethnic wars, revolutionary wars, and genocides/politicide, and excluded cases that were exclusively adverse regime change.Footnote ² We then reduced the overall sample to first-year onset (avoiding conflict continuation years) with a minimum of two years of peace prior to the civil war onset year.

We selected variables measuring theoretically supported drivers of civil war onset, intentionally selecting from a range of competing alternative explanations in the literature: economic development (Buhaug, Cederman, and Gleditsch Reference Buhaug, Cederman and Gleditsch2014; Collier and Hoeffler Reference Collier and Hoeffler2004; Van de Walle Reference Van de Walle2004), human development (Badie Reference Badie2000; Rotberg Reference Rotberg2002; Urdal Reference Urdal2005), demographics (Bricker and Foley Reference Bricker and Foley2013; Dyson Reference Dyson2012; Mesquida and Wiener Reference Mesquida and Wiener1999; Weber Reference Weber2019), governance (Bogaards Reference Bogaards2009; Cook and Savun Reference Cook and Savun2016; Dahl Reference Dahl1971; Mansfield and Snyder Reference Mansfield and Snyder1995), neighborhood effects (Buhaug and Gleditsch Reference Buhaug and Gleditsch2008; Cederman et al. Reference Cederman, Gleditsch, Salehyan and Wucherpfennig2013; Wolff Reference Wolff2011), and state-led discrimination/horizontal inequalities (Gurr Reference Gurr1970; Østby, Nordås, and Rød Reference Østby, Nordås and Rød2009; Reynal-Querol Reference Reynal-Querol2002). The sample covers 191 countries from 1960 to 2012 (see Table 1).Footnote ³ The final dataset for the base model includes 76 onset cases.Footnote ⁴

Table 1. Variable descriptive statistics (1960–2012, missing values removed, unnormalized)

We pursued various strategies to test the sensitivity of model assumptions to our results.Footnote ⁵ First, we altered the distance metric to use Gower's (Reference Gower1971) measure of distance, which accounts for mixed-type data inputs. We also tested an alternative clustering algorithm called “partitioning around medoids” (PAM) and tested alternative data, including the Varieties of Democracy (V-Dem) liberal democracy index (Coppedge et al. Reference Coppedge2021) and civil conflict onset sample data from the Uppsala Conflict Data Program / Peace Research Institute Oslo (UCDP/PRIO) (Gleditsch et al. Reference Gleditsch2002; Gleditsch et al. Reference Gleditsch2021). Finally, we added two additional variables that have been identified as drivers of civil war onset but are less policy-relevant: population size and peace years. This analysis validates our findings.

Results and Discussion

Table 2 and Figure 1 summarize the base model. Radar plots show the clusters’ mean value by variable, the full civil conflict onset sample mean value, and the full sample mean for both onset and non-onset cases.Footnote ⁶

Fig. 1. Base model results.

Table 2. Descriptive clusters of civil conflict onset

The top-left cluster in Figure 1 highlights civil war onset cases where the state has low levels of economic development, high infant mortality, young populations, low levels of state-led discrimination, few neighboring conflicts, and anocratic governance. The onset case that best represents the conditions described in this cluster is Guatemala in 1966, when communist insurgents launched a protracted guerilla conflict against the ruling military regime (Johnson Reference Johnson1966). This cluster is the second largest of the three, containing 39.47 per cent of civil conflict onset cases (N = 30). We describe this cluster as: Anocratic–Younger–Less Developed.

The top-right cluster in Figure 1 includes cases where the state has consolidated governance (either democracy or autocracy), gross domestic product (GDP) per capita above the global mean, low infant mortality, older populations, and more peaceful neighborhoods. However, it is notable that state-led discrimination in these states is somewhat high relative to the global mean, as well as the conflict-onset mean. This cluster is the smallest of the three in terms of onset cases grouped within it (17.11 per cent [N = 13]). The most central conflict case here is China in 2009, where Uyghur minorities launched violent actions against the government and Han Chinese residents in and around the Xinjiang province, such as the Urumqi riots (Hastings Reference Hastings2011). The characteristics of the states in this cluster at the time of their civil conflict onset can be described as: Older–Wealthier–Moderate Discrimination.

The final cluster in Figure 1 is characterized by states with younger populations, severe levels of neighborhood conflict, and very high levels of state-led discrimination. This cluster is the largest of the three that we found among our PITF sample of civil conflict onset cases, containing 43.42 per cent of total cases (N = 33). The most central onset case in this cluster is Sudan in 1983, wherein the Sudan People's Liberation Army in the country's more Christian southern regions launched an insurgent war against the majority Muslim Arab government in Khartoum, which sanctioned genocidal actions by progovernment militias in the region (Scott Reference Scott1985). This cluster's descriptive characteristics motivate us to describe the group as: Higher Discrimination–Worse Neighborhood–Younger.Footnote ⁷

These three clusters are persistent across various model specifications, with the Older–Wealthier–Moderate Discrimination cluster showing the lowest level of case variation across alternative clustering methods. The Anocratic–Younger–Less Developed cluster was the second most robust across model specifications, followed by the Higher Discrimination–Worse Neighborhood–Younger cluster.Footnote ⁸

We also examine how individual onset cases behave across models by comparing the cluster assignment of cases across all eight model specifications. For forty-four of the seventy-six cases (57.9 per cent), there is uniform agreement on cluster assignment, regardless of model specification. Conversely, in 19.7 per cent of cases (fifteen of seventy-six), onset cases do not neatly cluster for at least two out of eight model specifications, indicating broad disagreement (see Table 8 in the Online Appendix). For all robustness and sensitivity checks, see the Online Appendix.