2.1 The P index of conflict potential
Consider a population partitioned into \(n\ge 2\) non-overlapping groups. Let \(\pi _{i}\) be the relative population size of group i, where \(i=1,2,\ldots ,n\), and \(\Pi =(\pi _{1}, \pi _{2},\ldots ,\pi _{n})\) denote the vector of groups’ population shares. As in ER we conceptualize conflict potential as the sum of all effective antagonisms between individuals or groups in the society. The antagonism or alienation felt by one individual towards another is a function of the distance between them. Since, by assumption, individuals within each group are all alike, the strength of alienation at the group’s level is obtained as the sum of all individual alienations. The alienation becomes effective once it is translated into some form of organized action, such as political mobilization, protest or rebellion. The power of a group to translate the overall alienation into effective voicing depends on the degree of cohesiveness within the group, which in turn depends on the group’s relative size.
Here, we extend the ER approach and assume that the power of a group depends also on the relative sizes of all the groups listed in
\(\Pi \). As in ER, we specify a function
\(\Phi \) that combines the group’s power, that depends on
\( \pi _{i}\) and
\(\Pi \), with the alienation felt towards other groups. Following Montalvo and Reynal-Querol (
2005), we define the distance
\(D_{ij}\) between individuals belonging to two groups
i and
j, using a discrete metric, i.e.,
$$\begin{aligned} D_{ij} := \left\{ \begin{array}{ll} 0 & {\text {if }} \ i=j, \\ 1 & {\text {if }} \ i\ne j. \end{array} \right. \end{aligned}$$
As pointed out in Montalvo and Reynal-Querol (
2005), when the population is partitioned according to some categorical attribute like ethnicity, language or religion, identifying groups according to the so-called “
belong—does not belong to” criterion is less controversial than defining the distances between them simply because it reduces significantly the measurement error. Moreover, as argued by social institutionalists, the salience of collective identities (i.e., distances) may be fluid and context dependent.
1 The potential for conflict in a society then derives from the interaction between power and alienation, aggregated over all pairwise comparisons:
$$\begin{aligned} P(\Pi )=\sum _{i}\sum _{j\ne i} \pi _{i}\pi _{j} ~ \Phi (\pi _{i}, \Pi , D_{ij}). \end{aligned}$$
(1)
We assume that
\(\Phi (\pi _{i}, \Pi , 0)=0\) and let
\(\Phi (\pi _{i}, \Pi , 1):=K\phi ^{n}(\pi _{i}, \Pi )\) for some constant
\(K>0\). We add a superscript
n to
\(\phi \) to distinguish between distributions characterised by a different number of groups. Since
\(\sum \pi _{i}=1\), the index defined in (1) can be written as:
$$\begin{aligned} P(\Pi )=K \sum _{i} \phi ^{n}(\pi _{i}, \Pi ) ~ \pi _{i}(1-\pi _{i}), \end{aligned}$$
(2)
and will be called the
P Index of Conflict Potential. The function
\(\phi ^{n}(\pi _{i},\Pi )\) for
\(i=1,2,\ldots ,n\), will be referred to as the
effective power associated with group
i. The
\(P(\Pi )\) index, hence, is obtained as a combined effect of two different elements, namely the groups’ effective power and the between group interactions,
\(\pi _{i}(1-\pi _{i})\), measuring the probability of randomly selecting an individual from group
i that interacts with an individual from another group. The sum of these components gives the probability that two individuals randomly selected from a population belong to different groups. Special cases of (2) are the Reynal-Querol (
2002) and Montalvo and Reynal-Querol (
2005) discrete polarization index,
\(RQ=4\sum \pi _{i}^{2}(1-\pi _{i})\), where each group’s power is equal to its relative population size, and the fractionalization index,
\(FRAC=\sum \pi _{i}(1-\pi _{i})\), which derives from (2) when each group’s power is constant and equal to 1.
The most interesting specification of the P Index in (2) is obtained when each group’s effective power is neither constant nor proportional to the group relative size but may also depend on the distribution of the relative sizes of the other groups. Here we assume that groups are allowed to form coalitions that generate bipartitions of the population, and that the effective power of each group depends on all the potential contributions of that group to the worth of all the coalitions that it can theoretically belong to. With only two groups, the power of both groups is 1/2 when their sizes are equal, and is non-decreasing in each group size. The extreme case is the one where the power of the bigger group is 1, while the power of the smaller one equals 0. Applying this latter rule to any bipartition of the population when the number of groups is larger than two, the marginal contribution of any group to the worth of a coalition equals 1 whenever the sum of the relative sizes of all the groups forming the coalition exceeds 1/2 and becomes smaller than 1/2 if that particular group leaves the coalition. It equals 1/2 if either the contribution of a group allows the coalition to reach 50% of the population or if without the group the coalition covers exactly 50% of the population. While the marginal contribution to the worth of a coalition is 0 in all other cases. The relative weight of the sum of marginal contributions of any group to all possible coalitions with respect to the sum of the contributions of all the groups in the population, represents then a measure of that group’s effective power. In this particular case, where extreme relevance is given to inequality between the bipartitions, the effective power coincides with the relative Penrose–Banzhaf index of voting power in a simple majority game. The associated index corresponds to the extreme element of the parametric family \(P_{\infty }^{n}\) that will be discussed in detail in the following sections.
Some real-life examples, based on the data on ethnic distributions and conflict onset that we will present in Sect.
4, may help to illustrate similarities and differences in the behaviour of the
\(P_{\infty }^{n}\) index, FRAC and RQ.
2 For instance, when comparing countries like Kuwait where the ethnic group distribution is (0.49, 0.29, 0.22) and Mauritania with distribution (0.43, 0.41, 0.16), the change in terms of fragmentation (FRAC) and of the
\(P_{\infty }^{n}\) index are proportional. This is because, given that no group reaches the majority of the population, with three groups the effective power of all groups is identical and therefore the
\(P_{\infty }^{n}\) index results proportional to the interaction component measured by FRAC. In terms of polarization, however, the Mauritania’s distribution shows a slight increase in the RQ index because of the existence of two groups covering similarly high proportions of population. A more substantial difference in the behaviour of the three indices appears when the comparisons are made with Iraq with distribution (0.64, 0.19, 0.17). The population distribution in Mauritania is more polarized compared to Iraq, and it also results more fragmented. The
\(P_{\infty }^{n}\) index, on the other hand, ranks the two countries differently from RQ and FRAC, assigning to Iraq a significantly higher value with respect to Mauritania. Yet, in the time range considered in our analysis (1946–2005), Mauritania actually did not experience any ethnic conflict while Iraq has suffered from three distinct violent conflict episodes (in 1961, 1982 and 2004). The different behaviour of the P index compared to RQ and FRAC is due to the fact that a majoritarian group emerges in Iraq thereby shifting the power to this group and then making the index proportional to the interaction of this group and the remaining population. A similar argument holds for the comparisons with Mozambique whose ethnic distribution resembles the one of Iraq and that has experienced one ethnic conflict in the time period considered.
We can derive examples that follow an analogous pattern also for countries with four ethnic groups. For instance, when comparing Gabon with ethnic distribution (0.41, 0.38, 0.13, 0.08) to Benin with (0.40, 0.22, 0.18, 0.18), we can see that in both countries no group is majoritarian. Nevertheless, for the computation of the \(P_{\infty }^{n}\) index, the larger group in Benin exhibits a higher effective power because it is determinant in reaching the majority of the population in all coalitions involving two groups, which is not the case in Gabon where the larger group reaches 49% of the population in coalition with the smaller group. For this reason, the increase in fragmentation recorded in Benin because of a less unequal distribution of population shares is also in line with the increase in the value of the \(P_{\infty }^{n}\) index while polarization is higher in Gabon whose distribution is characterized by two groups with similarly large population shares.
In analogy with the case of three groups, a better alignment of the \(P_{\infty }^{n}\) index with conflict onset with respect to FRAC can be observed when comparing Benin and Bolivia. While Bolivia has experienced a violent social arrest in 1952 prior to the revolution, Benin did not go under any inter-ethnic conflict. The presence of a dominant group in Bolivia, covering 61% of the population leads to an increase of the \(P_{\infty }^{n}\) index compared to Gabon and Benin, while FRAC decreases. Here the \(P_{\infty }^{n}\) index appears more in line with the RQ index that also exhibits a sharp increase because the distribution becomes more polarized given the presence of two large groups while the other two groups cover very low shares of population. Comparison of Bolivia with the ethnic distribution of Sri Lanka shows a reduction in all indices because of the increased share of population (71%) covered by the dominant group.
2.2 Axiomatic derivation of the effective power function
Let
\(N:=\{1,2,\ldots ,n\}\) denote the set of all groups. The set of all vectors
\( \Pi \) is in the
n dimensional unit simplex
\(\Delta ^{n}\). The effective power
\(\phi ^{n}(\pi _{i},\Pi )\) of group
\(i\in N\) with relative population size
\(\pi _{i},\) given
\(\Pi ,\) is defined as:
3$$\begin{aligned} \phi ^{n}(\pi _{i},\Pi ):[0,1]\times \Delta ^{n}\rightarrow \mathfrak {R}_{+}, \end{aligned}$$
and satisfies the next properties.
Normalization requires that the powers of all groups sum to 1, and implies that the effective power of each group is bounded in the interval [0, 1].
The Monotonicity axiom implies that the effective power of a larger group cannot be lower than the effective power of a smaller group. This property implies the Symmetry of \(\phi ^{n}\) which requires that, if two groups are of equal size, then their effective power has to be the same. The reverse, however, is not necessarily true: the effective power could still be equal for groups of different relative size. Monotonicity in combination with Normalization implies that if all groups have identical relative size, each one of them has an effective power equal to 1/n. This result will provide a reference point for all the indices that we will obtain from the axiomatization. In fact, a common feature of these indices is that they all exhibit the same value for distributions where all the groups are of equal size. Moreover, this value will be proportional to the fractionalization index divided by n.
We now introduce two crucial assumptions: (i) groups can either act individually or through a coalition, and (ii) if any two or more groups form a coalition, the remaining groups belong to the “opponent” block. So we consider only bipartitions of the population.
What is the
rationale behind these two assumptions? Suppose that there are 3 groups involved in a contest with only one strategic endowment, namely human resources. A relatively smaller group that is interested in winning the contest may find profitable to join the forces with some other group in order to contrast the adversary, even at the cost of the future division of power within the winning block. Consequently, a group that is large enough to ensure the victory alone will act as an independent actor. Hence, one block or coalition may be formed in order to contrast or challenge the other block. Skaperdas (
1998), Tan and Wang (
2010) and Esteban and Sakovics (
2003) show that in a three groups contest, parties will have an incentive to form a coalition against the third if the formation of the alliance generates
synergies that enhance the winning probability of the coalition. Skaperdas (
1998) and Christia (
2012) argue that this tendency is not only theoretical but also frequent in many real life situations and provide an example of the “... on and off alliance of the Bosnian and Croat forces against the more (strategically) well endowed Serb forces in Bosnia during the recent past ...” (Skaperdas
1998, pg. 27). Along similar lines, a close cooperation between the Eritrea People’s Liberation Front (EPLF) and the Tigray People’s Liberation Front (TPLF) during the Ethiopia’s civil war led to a victory against the authoritarian rule of the Mengistu government (DERG). These two rebel groups had different ideological and territorial objectives but at the same time they recognized the benefit, in terms of material and tactical capability, from joining the forces through an alliance. Another example relates to the two rival Kurdish groups, Patriotic Union of Kurdistan (PUK), and Kurdistan Democratic Party (KDPI) who started to collaborate in mid-1980s against the Ba’ath regime in Iraq, conversely the brief alliance between two Chadian groups, Forces of the North (FAN) and Popular Armed Forces (FAP) broke down in 1979 turning the two groups into bitter enemies.
A further motivation underlying the assumption on bipartitions is the following. According to the theory of alliance formation in a multi-group context (Riker
1962; Axelrod
1970; Christia
2012), when forming a coalition, alliance partners expect more benefit (in terms of odds of winning) from joining the forces against third parties than from a conflict against each other. One of the basic mechanisms underlying this theory assumes that coalitions’ relative power depends on the sum of their relative population size, military power and/or territorial control capacity. The relative power in turn determines the probability of winning, the smaller the difference between the coalitions’ relative power, the more similar the odds of winning the contest. The complexity of the model increases with the number of groups and potential coalitions. Since smaller relative power differentials translate into lower relative odds of winning, even with more than two potential alliances, coalitions will have an incentive to join the forces and form supra-coalitions, i.e., coalitions of coalitions, collapsing the coalition structure to a bipartition scheme (Christia
2012).
Here we do not model any endogenous mechanism of coalition formation nor we are interested in which coalition is more likely to form. The probability distribution over coalitions, hence, is assumed to be symmetric. Symmetry is a plausible assumption in our case since we do not have or use information about the differences among groups and we use a discrete metric to define these distances.
4 Considering only the bipartitions of the population, we rule out the possibility that more groups run on their own against the rest. However, we do not rule out any coalition between two or more groups. As previous examples clearly show, this assumption is not unrealistic, since in many situations that involve coalition formations in conflicts, even unmatchable parties often coordinate their interests in order to contrast the opponent, even when they are aware that the coalition is temporary (Esteban and Sakovics
2003). As we will show later, even under these simplifying assumptions the distribution of the effective power between groups will depend on the characteristics of the population distribution across them. This important feature of the effective power function will make the
P index substantially different (both theoretically and empirically) from the existing distributional diversity indices based on the assumption of groups as independent actors.
5
In order to characterise the effective power for any arbitrary number of groups we first consider the simpler case of a distribution with only two groups. The results that we obtain will then be used to generalize the analysis for any arbitrary number of groups.
Consider a population divided into two different groups (
\(n=2\)) with population shares
\(\pi \) and
\(1-\pi \). Denoting with
\(\phi ^{2}(\pi )\) and
\( \phi ^{2}(1-\pi )\) the effective power of the groups, we define the
relative effective power between them as:
6$$\begin{aligned} \frac{\phi ^{2}(\pi )}{\phi ^{2}(1-\pi )} = r(\rho ) \quad {\text {where }} \, \rho := \frac{\pi }{1-\pi }. \end{aligned}$$
(3)
Thus, the relative effective power between groups is a function
\(r(\cdot )\) of the groups relative population size
\(\rho \) that coincides with the population shares
odds ratio. From Monotonicity it follows that whenever
\(\pi =1/2\), hence
\(\rho =1\), the groups will equally share the power, that is,
\(r(1)=1\).
7 The relative effective power is supposed to satisfy the following property:
In order to interpret the 2GRPH axiom, suppose that we start from a population distribution \(\Pi \) in which, for instance, the size of the smaller group is 40% of that of the larger group (i.e., \(\rho =0.4\) ). Now imagine that a portion of the population from the second group migrates in a neighboring country such that the size of the smaller group is now 80% of that of the larger group (i.e., \(\rho =0.8\)). Thus \( \rho \) has doubled (i.e., \(\lambda =2\)). Such a variation in the relative population size may affect the relative effective power between the two groups. Now imagine a similar situation where \(\rho \) doubles but the relative size of one group moves from 30 to 60%. The 2GRPH axiom requires that the variation in the relative effective power is the same in both cases. In other words, no matter from where we start with respect to the relative size \(\rho \), the variation in the relative effective power is always the same as long as the change in \(\rho \) is of the same proportion across the two distributions.
We can now state the first result proved in the Appendix.
This functional form for the effective power is similar to the
ratio form contest success function commonly used in the rent-seeking literature (Tullock
1980 with
\(\alpha =1\), Skaperdas
1996,
1998; Nitzan
1991). However, the axiomatization of the effective power function differs from those in the literature.
8 The coefficient
\(\alpha \) represents the elasticity of the relative effective power with respect to the relative population size. When
\( \alpha =0\) the relative effective power equals 1, for
\(\alpha =1\) each group’s power equals its population share, while for
\(\alpha \rightarrow \infty \) the majoritarian group holds the absolute power. We will call such group with
\(\pi >1/2\) dominant.
Consider now \(n>2\). Groups are allowed to form coalitions that generate the bipartitions of the population. A coalition is defined as any subset of the set N of all groups (including the empty set). In particular the grand coalition contains all the groups; an individual coalition contains only one group; and the empty coalition contains no group. Since we assume that groups can either act individually or form alliances or blocks with other groups, any measure of their effective power should take this possibility into account. This means that a measure of effective power has to consider all the potential contributions of a group to all the coalitions that it can possibly belong to.
Denote with \(C_{i}\) the set of all coalitions c that include group i. This set contains both the grand coalition and the \(i^{\prime }s\) individual coalition. The power of any coalition c is obtained by Lemma 2.1 as \( \phi ^{2}(\sum _{j\in c}\pi _{j})\), where the power of an empty coalition is 0 and the power of the grand coalition is 1.
We next define the
marginal contribution of group
i to the power of any coalition
\(c\in C_{i}\) as (Shapley
1953):
$$\begin{aligned} m_{i}(c):=\phi ^{2}\left( \sum _{j\in c}\pi _{j}\right) -\phi ^{2}\left( \sum _{j\in c}\pi _{j}-\pi _{i}\right) . \end{aligned}$$
(4)
The sum of the marginal contributions of group
i over all coalitions in
\(C_{i}\) is:
$$\begin{aligned} M_{i}=\sum _{c\in C_{i}}m_{i}(c). \end{aligned}$$
(5)
The effective power of any group
i will be a function of
\(M_{i}\) but it will also depend on the marginal contributions of the other groups
\(M_{-i}\). However, as stated in the next axioms, what counts for the
relative effective power between any two groups
i and
j is the ratio between some transformation of the sum of their marginal contributions.
The REP axiom states that the relative effective power between any two groups \(i,j\in N\) depends on their sum of marginal contributions to all the coalitions that they can theoretically belong to. No matter how many groups there are in the population or how the marginal contributions are distributed among them, the relative effective power between any two groups will be determined exclusively by a ratio of a transformation \( g(\cdot )\) of their own Ms. The exclusive role of M in the determination of the relative effective power implies that the strength of one group with respect to another within the same distribution will be determined only by the relative importance they have in terms of the value they add to the coalitions they belong to. Since the same transformation function applies to marginal contributions of all the groups in the population, when two groups have the same M, this is the case also for their effective powers. The property leaves open the possibility that two groups covering different population shares may also be endowed by the same effective power. This could be the case because the \(g(\cdot )\) transformation may attach the same value to different M or because groups with different population shares exhibit the same value of M. The REP axiom considers comparisons across groups made within the same ethnic distribution, for this reason the transformation \(g(\cdot )\) may depend also on the distribution.
The relationship between the ratio of marginal contributions and the relative effective power is clarified by the following axiom where comparisons are extended to groups belonging to different distributions.
According to nGRPI if we compare two population distributions with the same number of groups, and if the ratio between the marginal contributions between any two groups from both distributions is the same, then their relative effective power has to be the same too. That is, the relative effective power is invariant with respect to the distribution of the groups population shares for groups with the same sum of marginal contributions. This axiom emphasizes the fact that the relative effective power between two groups i and j may depend also on the distribution of the other groups. However, this effect arises only if the distribution of the groups affects the ratios of the aggregate marginal contributions, \(M_{i}\) and \(M_{j}\).
Next theorem, proved in the Appendix, provides a role for the sum
\( M_{i}^{\alpha }\) of the marginal contribution to all coalitions of group
i obtained as in (5) making use in (4) of the functional form
\(\phi _{\alpha }^{2}\) derived in Lemma 2.1. Thus,
$$\begin{aligned} M_{i}^{\alpha }:=\sum _{c\in C_{i}} \left(\phi _{\alpha } ^{2} \Bigg(\sum _{j\in c}\pi _{j}\right)-\phi _{\alpha }^{2}\Bigg(\sum _{j\in c}\pi _{j}-\pi _{i}\Bigg)\Bigg). \end{aligned}$$
Group i’s effective power, hence, is defined as the relative sum of the marginal contributions of this group to all possible coalitions, valued according to \(\phi _{\alpha }^{2}\). Given (6), the effective power of a group can be a function of the relative size of all the groups in the population. For \(n>2\) and \(\alpha \notin \{0,1\}\), the effective power of any group i depends on both \(\pi _{i}\) and \(\Pi _{-i}\). As a consequence, the effective power of a group with a fixed population share \(\pi _{i}\) may vary significantly across different population distributions in response to the variation of the relative size of the other groups \(\Pi _{-i}\).
The previous result suggests that the effective power is
not necessarily proportional to the groups’ relative size. This is in line with the literature on
voting power.
9 For instance, when
\(\alpha \rightarrow \infty \), the group
\( i^{\prime }s\) effective power,
\(\phi _{\infty }^{n}(\pi _{i},\Pi )\), coincides with its relative Penrose–Banzhaf index of voting power in a simple majority game (Felsenthal and Machover
1998).
10
As we will highlight in next section a transformation of the fractionalization index and of the RQ ethnic polarization index are special cases of the parametric family of the P index of conflict potential whose effective power functions are characterized in Theorem 2.2. Bossert et al. (
2011) have provided a characterization of a generalization of the fractionalization index where differences between ethnic groups may not necessarily be binary, while Chakravarty and Maharaj (
2011) have provided alternative characterizations of the ethnic polarization index. Both contributions differ from the current work because they assume some form of additivity for the aggregate index, as a result the contribution of the distribution of each ethnic group to the overall index is not affected by the distribution of the other ethnic groups. This is the case for the P index only for some parameter values that make it proportional to the fractionalization index or lead to the ethnic polarization indices but in general
\(\phi _{\alpha }^{n}(\pi _{i},\Pi )\) may depend on
\(\Pi \) and not only on
\(\pi _{i}.\) Other contributions like Desmet et al. (
2009) take into account the interaction between many peripheral/minoritarian groups and a central/majoritarian group. In this contribution also the distance (non necessarily binary) between groups is relevant. In an analogous manner, when distances are binary, the P index for
\(\alpha \rightarrow \infty \) enphasizes the dominance component when the larger group exceed 50% of the population. To conclude, as pointed out the effective power function
\(\phi _{\alpha }^{2}(\pi _{i},\Pi )\) for
\(n=2\) relates to the ratio form contest success function even though, as argued the derivation is different. The generalization to
\(\phi _{\alpha }^{n}(\pi _{i},\Pi )\) however, considers properties defined over the distribution of the aggregate marginal contributions of each group to all bipartitions of the population, instead of being defined over the distribution of the population shares of each group. The consistency between these properties that are set for a generic
\(n\ge 2\) and those valid for
\(n=2\) that lead to
\( \phi _{\alpha }^{2}(\pi _{i},\Pi )\) is verified in the first part of the necessity part of the proof of Theorem 2.2. One may want to consider alternative specifications for the effective power function
\(\phi ^{n}(\pi _{i},\Pi )\) in the spirit of the
n groups contest functions of Skaperdas (
1996) or their generalizations in Chakravarty and Maharaj (
2014). In analogy with the derivation in Lemma 2.1, it could be possible to derive a functional form for
\(\phi ^{n}(\pi _{i},\Pi )=f(\pi _{i})/\sum _{j}f(\pi _{j}) \) for non-decreasing
\(f(\cdot ).\) In this case the relative power of group
i w.r.t. group
j depends only on
\(\pi _{i}\) and
\(\pi _{j}\) but not on the distribution of the other groups. This is not in general the case for the P index if
\(n>3\). In fact for the P index this relative comparison depends on the ratio of the aggregate of the marginal contributions of the two groups to the power of each bipartition of the population. In our case these values could depend also on the distribution of all the other groups.