Why ethnic parties form: evidence from Bolivia

Research on the effects of ethnic politics abounds, but much less attention has been paid to where and why ethnic parties form. This article tests the explanatory power of rational-choice and social-movement informed approaches to ethnic party formation that, it argues, differ in their assumptions about the location of agency (elite vs. grassroots) and motives for party formation (officevs. policyseeking). The assumptions are tested through an analysis of original data on party registration and socio-economic factors in 327 Bolivian municipalities during the 2004 local elections. The elections took place under new electoral rules during a period of political restructuring, allowing an analysis of party entry decisions per se. Through a series of logistic regression models and various robustness checks, this article finds that social-movement approaches are better able to explain ethnic party formation, and in particular that grievances over political maladministration and socio-economic inequalities drive ethnic party formation.


Introduction
Why do ethnic parties form? Particularly in new, unconsolidated democracies, ethnic cleavages are feared to be politicised in the form of ethnic parties, polarising ethnic relations and endangering political stability (e.g . Brancati 2006;Horowitz 1985;Rabushka and Shepsle 1972). But while research on the effects of ethnic politics abounds, much less attention has been paid to where and why ethnic cleavages translate into party politics: not everywhere where there is ethnic diversity are there ethnic parties. The party system literature sees ethnic diversity as driver of party system fragmentation more generally, but studies tend to focus on consolidated democracies and rarely go beyond establishing a correlation between the number of ethnic groups and the number of parties in general (e.g. Clark and Golder 2006;Lago Penas 2004). Research on new party formation, too, concentrates on established party systems of Western democracies and cannot be directly translated to developing democracies, where the probability of ethnic parties is thought to be the highest (e.g. Lago and Martínez 2011;Tavits 2006). Research on elections quantitative, spatially disaggregated analysis of the relation between local contexts and ethnic party formation. To fill this gap, I use new municipal-level data on party registration, socio-economic factors and demographic differences in 327 Bolivian municipalities.
Through a series of logistic regression models and various robustness checks, I document that a variety of factors contribute to ethnic party formation, but that political calculations of individual ethnic entrepreneurs are not likely to be one of them: neither electoral strategy nor office-seeking incentives affect the probability of ethnic party formation. Instead of greed, grievances with the current administration are driving ethnic party formation, over and above ethnic group size and other contextual factors like the population size of the polity. Overall, the findings suggest that individual leaders have a stronger agency in party formation than often assumed in social-movement approaches, but that they are not as strongly driven by their own individual interests as often assumed in instrumental approaches. With this finding, the paper contributes to the broader literature on ethnic mobilisation.
The following section first specifies the instrumental and social-movement approaches to ethnic party formation, with emphasis on differences regarding the ethnic target group and motives for party formation. The next section provides background on the Bolivian context and outlines its suitability as a plausibility probe for the two approaches. The paper then operationalises the factors outlined above and presents the results. The conclusion discusses the findings in the light of the theory and lists limitations, as well as directions for future research.

Ethnic party formation
Research on ethnic party formation is sparse. To date, the literature on ethnic parties has focused mainly on their strategies once formed (Bochsler and Szöcsik 2013;Zuber 2013); their success (Chandra 2004;Madrid 2012); and their influence on ethnic relations and conflict (Birnir 2007;Ishiyama 2009). Yet most such studies are informed by assumptions on why ethnic parties form, making an empirical analysis of the dynamics of formation imperative. Notably, the literature on ethnic party success often assigns the same explanatory factors to party emergence as it does to success, even though the two processes are distinct in practice: explaining success requires considering not only party or elite behaviour but also that of voters (Hug 2001), and voting behaviour in turn is affected by the parties on offer. Focusing on ethnic party success alone also introduces the question of selection bias, since only those parties may form for which there is indeed a demand (Bernauer and Bochsler 2011). The present paper hence examines the supply side of ethnic parties, that is, party formation, in more detail. Party formation is here defined as first-time entry to an electoral competition, and in the following party entry and formation are used synonymously.

Why ethnic parties form 3
Corresponding to the focus on formation, I use a minimal definition of ethnic parties, according to which a party is ethnic if it explicitly appeals to a specific population group united by an ethnic attributerather than, for example, if it attains the majority of votes from specific ethnic groups (see e.g. Chandra 2011;Horowitz 1985). While ethnic appeals may also be implicit or covert, for instance, due to social norms or political regulations (Chandra 2011;Ishiyama and Breuning 2011), explicit appeals are those referred to most often in the existing theories on ethnic parties, as discussed next. Moreover, Bolivia, where the theories are tested, does not show spatial variation regarding norms or regulations that would systematically affect the existence of explicit appeals, making the latter appropriate to use as defining factor of ethnic parties.
Research on ethnic parties approaches the issue of party formation from either of two directions. On the one hand, we have instrumental accounts influenced by formal models of party entry in rational choice approaches, most fully developed by Cox (1997;see also Hug 2001;Tavits 2006). According to instrumental accounts, party entry is the result of a strategic decision taken by rational actors to obtain the benefits associated with holding office. Rational actors only decide to enter the fray, and hence to bear the costs associated with registering and campaigning, if they think that they are likely to win enough votes to enter office. Accordingly, political actors also only form ethnic partiesrather than a non-ethnic party or no party at allif they perceive a chance of winning based on this cleavage. This argument assumes that it is clear at the time of entry which party or issue is viable on election day. While Cox (1997) argues that the assumption does not hold, by definition, in situations of uncertainty as in early elections, scholars of ethnic parties maintain that it is exactly in these situations that ethnicity is used as basis for party formation. Precisely because there is no information on the viability of parties according to their policy positions, candidates tend to campaign on the basis of their personal attributes to attract voters of similar backgrounds (Moser 1999;Mozaffar, Scarritt, and Galaich 2003). Drawing on theories of political learning and information shortcuts, these studies theorise that, in low-information settings, ethnic cues are readily available sources of information and may be strategically emphasised by politicians to guide the political choice of voters (Birnir 2007;Chandra 2004). Ethnic cleavages do not translate 'automatically' into politics: from a repertoire of cleavage dimensions like region, religion, language or skin colour, politicians tend to emphasise that dimension that promises the necessary number of votes to win office (Chandra 2004;Posner 2005). 1 On the other hand, there are approaches to ethnic party formation that align with social-movement theory. The mobilisation of new political demands, whether through parties or movements, has often been explained by the opportunities and organisational capacity available for mobilisation to address an underlying dissatisfaction with the status quo (Eisinger 1973;Kitschelt 1986). These studies, too, assign a role to political actors taking the conscious decision to form a party; however, this decision is seen to be more Anaïd Flesken 4 strongly influenced by the political structures in which the actors find themselves (Morris and Staggenborg 2006). Moreover, in these approaches, the actors form parties to seek policies addressing collective, rather than their own, individual interests. That is, (ethnic) party formation is much more a bottom-up rather than top-down process (e.g. Rice and van Cott 2006;van Cott 2005). This is not to say that the distinction between the two approaches is always clear cut (see Opp 2009). It is hence not surprising that their different concepts are often measured with similar indicators or that both have been extended with similar arguments, such as political learning. However, both approaches have different underlying assumptions, with different implications for what we would have to observe if ethnic parties were formed either by instrumentally driven, office-seeking or by collective-interest driven, policy-seeking political actors. It is hence worth spelling out the differences in more detail. In particular, the two approaches have different implications with regard to their motives, ethnic group size and institutional permissiveness.

Motives
One observable difference between instrumental and social-movement approaches lies in the incentive structure for party formation. According to the former, politicians are instrumentally rational and driven by the financial rewards and prestige that come with political office (Tavits 2006). In new democracies in particular, patronage expectations are anticipated to structure incentives (Allen 2012;Chandra 2004). Accordingly, the probability of entering an electoral race, either with an ethnic or a non-ethnic party, should be higher the more personal gain politicians expect to derive through such practices.
Social-movement approaches, in contrast, see collective grievances of the ethnic group vis-à-vis the traditional elites as driving force for party formation, as it is these grievances that are to be addressed once in power. That is, parties are primarily policy-seeking rather than office-seeking. Such grievances may include socio-economic inequalities but also political or cultural marginalisation (Bilinski 2015;van Cott 2005). That is, here, we would expect to see a higher incidence of ethnic parties where there is more dissatisfaction with the current system.

Ethnic group size
An important factor in both approaches is the population share of the ethnic group that is targeted by the party. Social-movement approaches see the ethnic group's population share as an indicator for the pool of potential electoral candidates and their organisational backing. That is, the larger the ethnic group, the more likely the presence of dedicated individuals to mobilise politically, and the more likely their collective support. This is not just a numbers game: Why ethnic parties form 5 ethnic group size itself is of little use where the group's organisational capacity is limited, and vice versa. For example, indigenous parties emerged in Colombia despite the country's miniscule indigenous population of less than 3 per cent, while they did not emerge in Peru, which has an indigenous population share of almost 40 per cent, but in which organisational structures are only weakly developed (van Cott 2005). However, quantitative analyses do include indigenous population share as indicator of organisational capacity, with the hypothesis that the higher the share, the higher the chances for party emergence (Birnir 2004;Rice and van Cott 2006). In instrumental approaches, in contrast, the ethnic group share is important as pool of voters rather than of candidates. This is expressed in the information-shortcut hypothesis that states that political entrepreneurs use ethnic traits in the voting population to conduct a head count to see whether they would attain sufficient votes for office. That is, this literature sees the political entrepreneur as strategic in the sense that she may also mobilise in non-ethnic terms if the demographic variation on the ground is not fruitful for ethnic mobilisation (Chandra 2004;Posner 2005). Where ethnic group share is the pool of potential voters, the relationship between voters and ethnic party emergence is not linear: ethnic parties would only form in constituencies when the ethnic group share passes a certain threshold after which entering office is virtually guaranteed. But ethnic parties would also not form in constituencies with very high ethnic group shares, since then the utility of ethnic traits to differentiate between voters vanishes (Chandra 2004).

Institutional permissiveness
The permissiveness of the institutional system also plays a role in both approaches. Particularly the electoral system is identified as limiting factor. In instrumental approaches, actors' strategic calculations depend not only on the pool of potential voters but also on the number of voters needed to enter office. That is, electoral institutions do not only distribute votes into seats following the elections (mechanical effect), but work even before the elections take place, by shaping party and voter strategies in anticipation of the institutions' mechanical effect (psychological effect) (Clark and Golder 2006;Duverger 1954). In districts with a small number of seats, and hence a high number of necessary votes and low probability of winning a seat, the incentives are higher to form broader alliances rather than to enter an individual party. In districts with a larger number of seats available, in contrast, more parties are potentially viable and even minority groups may feel encouraged to register their own party to attain representation (Cox 1997). This 'anticipated strategic voting' hypothesis has received considerable empirical support from crosscountry comparative studies in consolidated democracies, which shows that party system fragmentation is the product of an interaction between ethnic fractionalisation as measure of diversity and average district magnitude as Anaïd Flesken 6 measure of electoral permissiveness (Amorim Neto and Cox 1997;Clark and Golder 2006).
In social-movement approaches, too, institutional permissiveness in the form of registration costs or electoral thresholds affects party formation, but here, not so much in altering actors' strategic choices but in hindering the registration of parties as well as their electoral success after the decision to run has been made. van Cott (2005) makes institutional permissiveness an important part of her explanation of ethnic party emergence in Latin America (also Birnir 2004, but see Madrid 2012.
In summary, then, we can formulate two expectations. First, the instrumental approach leads us to expect ethnic party formation where both the material or symbolic incentives for holding office and the probability of winning office are perceived to be high. The latter depends on the extent to which the ethnic group size aligns with the vote count necessary to win office. Second, the socialmovement approach leads us to expect that the probability of ethnic party formation increases with ethnic group size and with the degree of collective grievances and institutional permissiveness. In the following, the two hypotheses are tested with ethnic party entry in the 2004 Bolivian municipal elections.

Ethnic parties in Bolivia
Bolivia is one of the most ethnically diverse countries in Latin America. Shortly before the 2004 elections, around 62 per cent of the population identified themselves in the national census as belonging to one of the country's now thirty-five recognised indigenous peoples (Bolivia 2001). Bolivia is also one of the most unequal countries in Latin America, and indigenous origin and socioeconomic status are closely linked: the indigenous population fares consistently worse in land ownership, poverty, malnutrition or infant mortality (CERD 2006;ECLAC 2006). But despite this inequality, ethnicity has only recently begun to play a role in electoral mobilisation, with first attempts in the 1970s and more assertive mobilisation from the 1990s onwards (van Cott 2005;Weber, Hiers, and Flesken 2016;Yashar 2005). During the 2002 presidential elections, the young party MAS, arising from the coca farmers' movement in the Bolivian countryside, made indigeneity a salient issue, and the party's leader Evo Morales gained a surprising 20.9 per cent of the vote. Less successful but nevertheless present and active was the MIP led by Felipe Quispe Huanca, which, as the party name suggests, puts an even stronger emphasis than the MAS on indigenous issues (Albó 2002;Singer and Morrison 2004).
Yet before the 1990s, party formation around indigenous identity was rare and not electorally viable; ethnic cleavages in Latin America had not translated into ethnic parties. Instead, indigenous Bolivians were linked either to traditional parties through patron-client relations; to leftist parties that directly appealed to (but rarely represented) indigenous interests; or to populist Why ethnic parties form 7 parties that appealed to indigenous Bolivians' class background against their common opponent in the form of the ruling elite (Rice and van Cott 2006). The change from this situation to one of indigenous political mobilisation is often attributed to the possibility of strong local movements to use the opportunity of a political opening accorded by institutional reforms in the 1990s. Yashar (2005) suggests that the combination of political and economic liberalisation and pre-existing social networks triggered the politicisation of indigeneity in the Andes. van Cott (2005) focuses more strongly on the role of the political institutional environment, like the degree of decentralisation and ballot access or the presence of reserved seats, but also stresses the importance of community-level organisational structures and strategic political actors to use political opportunities. However, while these and similar accounts recognise that organisational capacity differs in space, the empirical exploration is limited to the differences between the two major geographic areas in the region: the Andean highlands and the Amazonian lowlands. Further exploration has so far been hampered by lack of suitable sub-national data indicating organisational capacity (Rice and van Cott 2006). Previous studies were therefore not able to explain why ethnic parties form only in some municipalities but not others. An analysis of the 2004 municipal elections accords us that opportunity. Prior to the elections, electoral reforms further opened the political arena to new political actors. The electoral law now, for the first time, allowed citizens' associations and indigenous peoples to compete in elections alongside national, traditional parties. 2 Low registration hurdlesrequiring signatures equivalent to only 2 per cent of the valid votes in the previous elections, or at least five signatures (Bolivia 2004a)ensured that this opening did not just exist on paper alone. Moreover, the preceding general elections of 2002 amid an anti-neoliberal protest wave had dealt a blow to the traditional parties and increased the visibility of indigenous political organisations (Singer and Morrison 2004), which both increased the incentives for new (indigenous) actors to compete in the local elections.
The 2004 municipal elections present a useful opportunity for three reasons. First, with such drastic changes in the political environment, it is possible to consider them as relatively independent from previous elections, and hence to study spatial differences in party formation itself rather than, say, their continued success. This context, a low-information setting in which the utility of ethnic mobilisation was becoming increasingly clear, allows focusing on whether actors decide to enter the electoral race and whether they do so by emphasising ethnicity.
Second, the Bolivian socio-economic and political environment makes it an ideal context to probe the plausibility of the instrumental and social-movement approaches: ethnically structured inequalities together with a history of political patronage and of strong social movements make Bolivia a 'most-likely' context for both explanations, and any absence of evidence for either link particularly telling. 3

Anaïd Flesken 8
Third, examining sub-national, rather than cross-country variation, addresses several methodological problems usually faced in the examination of party entry. It allows testing the effect of electoral permissiveness directly at the district level where it occurs (Cox 1997) and avoiding the endogeneity problem between the electoral system and the nature of the existing party system (Benoit 2002): municipal-level electoral rules are dictated by the state, and hence independent of the local party system. This has the added benefit that electoral rules do not vary aside from the district magnitude, which makes it easier to trace its effect. Last, sub-national units of analysis are also better comparable with regard to data on ethnicity since they rely on only one source, which further reduces non-random measurement error (Jones 1997).

The data
This paper uses an original compilation of municipal-level data on party registration, socio-economic factors and demographic differences in 327 municipalities for the 2004 elections. The following presents the measurement of the dependent variable (ethnic party entry) and of party motives, ethnic group size and institutional permissiveness, as well as of control variables (compare Table 1).

Ethnic party entry
The dependent variable indicates the entry of an ethnic party into the electoral race. Entry refers to the registration of the party for competition, as recorded by the Bolivian electoral court (Bolivia 2004b). I use this minimal definition of entry rather than a minimum vote count since the focus of analysis is on organisations' decision to enter the race. This is a priori unrelated to its performance, particularly so in an open political context during which many organisations form and enter for the first time. In the 327 municipalities examined here, 419 different political organisations entered the elections, made up of 21 traditional parties, 58 indigenous peoples and 340 citizens' associations. The number of organisations per municipality ranges from one to twenty-one.
A party was coded as ethnic party if it explicitly mobilises around ethnicity, by i) registering as representing an indigenous people, ii) explicitly referring to indigeneity in its name or iii) using an indigenous-language name. Following the identification of ethnic parties, it was further necessary to assign them to the specific ethnic group they claim to represent, to be able to see whether ethnic group characteristics like size would make a difference. Although the Bolivian constitution of 2009 recognises thirty-five indigenous peoples, the highland-lowland distinction is the most salient division in both politics and society (Flesken 2013;van Cott 2005), and I hence categorised parties into highland and lowland indigenous parties. For this, it is possible to use candidates' surnames as categorising device since they signal ethnic group Why ethnic parties form 9 Anaïd Flesken 10 membership even where other information is not available (Chandra 2004;Dunning and Harrison 2010). In Bolivia, the most common indigenous languages in the highlands, Aymara and Quechua, are phonologically similar and can be distinguished from the languages spoken among Amazonian lowland indigenous peoples, like Guaraní or Mojeño (Cerrón-Palomino 2015). With the assistance of a Bolivian linguist, I distinguished ethnic parties accordingly based on the linguistic background of the highest-ranked candidate's surnames on the party list, as it was this name that appeared on the ballots, visible to voters on election day. We coded surnames with discernibly Aymara or Quechua roots, like Condori or Mamani, as highland and surnames with roots from Amazonian indigenous languages, like Bejarano or Mayja, as lowland indigenous. Where candidates of indigenous organisations had surnames of Castilian origin, like Cruz or Lopez, I judged the ethnic background of the organisation based on i) the surnames of the lower ranked candidates on the list, ii) descriptors or language of the party name itself or iii) in-depth case knowledge.
The lowland and highland indigenous populations do not, despite their designations, live in separate parts of the country; in a regression design, ethnic party formation could therefore be attributed to either the highland or the lowland indigenous population, or to both. Since the number of lowland people and parties is small, I restrict the analysis to those of the highlands, but note that the models provide substantively similar results even without the differentiation between highland and lowland. In total, eighty different highland indigenous parties entered the elections in 156 municipalities, ranging from one to five organisations per municipality. As the distribution of highland indigenous parties is skewed across municipalitieswith more than two ethnic parties in only three municipalitiesit does not allow for meaningful analysis of party counts with Poisson regression. Instead, I recoded the data into a binary indicator for the presence or absence of ethnic parties.

The role of the MAS and the MIP
In the Bolivian context, special consideration is merited by the parties MAS and, perhaps to a lesser extent, MIP. 4 Owing to their emphasis on indigenous issues, both parties were, at the time of the municipal election, ethnic parties. However, the focus of this paper is the new formation of ethnic parties, and by then, both MAS and MIP had already been present and built supra-local organisational frameworks. That also means that, despite the parties' grassroots nature, the decision to enter the different municipal elections was not made by local but by national actors (Do Alto 2008). The MAS was present in all but 31 municipalities, and the MIP still in 123 municipalities. Their inclusion into the dependent variable thus does not make sense for theoretical nor methodological reasons, since there is little spatial variation in their registrations.

Why ethnic parties form 11
Yet, the presence of two country-wide parties that would also field candidates in the municipal elections could affect political entrepreneurs' considerations of whether to form an(other) ethnic party. On the one hand, local ethnic parties may be crowded out: where local actors perceive that ethnic rhetoric has already been used successfully, they may decide against forming a party or at least against using ethnicity as politicised cleavage. On the other hand, local actors may have learnt from previous MAS or MIP success and now try to imitate their political rhetoric. Imitation as a factor in new party formation has been observed elsewhere (Erlingsson 2008), and the MAS itself has been argued to have learnt from ethnic mobilisation attempts in the past and in neighbouring countries (van Cott 2005). I hence account for prior success of the MAS and MIP by including their 2002 vote share as control variable. 5 A positive relation with the probability of ethnic party presence in a municipality would support the imitation hypothesis, while a negative relation would suggest a crowding-out effect.

Ethnic group size and institutional permissiveness
To account for an ethnic party's potential support base, I use the share of the highland indigenous population. Ethnic group shares are measured according to the number of people per municipality identifying themselves in the national census of 2001 as belonging to either the Aymara or Quechua highland indigenous population (Bolivia 2001). The census data on indigenous selfidentification has been criticised for overestimating the number of indigenous Bolivians, largely due to the lack of a 'mestizo' (mixed) response category (Toranzo Roca 2008). 6 Alternative measures of ethnicity, such as language skills, however, may have the opposite problem of underestimating the indigenous population size: not everybody identifying as indigenousand hence potentially mobilising as indigenousmay have learned to speak an indigenous language during childhood, particularly where, as in Latin America, doing so has tended to limit social mobility (Garcia 2003). Self-identification catches best the availability of ethnicity as mobilisational tool at the individual level (see also Mähler and Pierskalla 2015), and there is no reason to suspect that overestimation varies systematically by municipality. The highland indigenous population shares ranges from under 1 to 100 per cent per municipality.
Following the theoretical discussion, the effect of ethnic group size depends on the institutional context, with different logics in the social-movement and the instrumental approach. In the former, highland indigenous population share is included as such since the probability of party formation increases with group size. A potential limiting factor in this regard is here only the permissiveness of the institutional environment, that is, whether the group is large enough to vote their candidates into office. To account for this, I also include a binary indicator 'sufficient indigenous' into the model, with 1 indicating that the indigenous population is larger than the population share needed to vote one Anaïd Flesken 12 candidate into office, which in turn is a function of the total population and the district magnitude in each municipality.
In the instrumental approach, group size is also important to the extent to which the group share is equal to or larger than the vote share required, but here, the size of the majority is decisive. I therefore include a continuous indicator 'indigenous majority' that captures this relation by subtracting from the indigenous population size that needed to obtain a seat. 7 According to the instrumental approach, the utility of ethnicity as mobilisational base, and hence the probability of ethnic party presence, follows an inverted U-shaped function: while the probability increases with ethnic group size, it decreases again where the group has such a vast majority that ethnic traits cannot serve as marker of difference anymore. To model this curve, I include both its value as well as its square. If the distribution is as expected, the first should be significantly positive and the latter significantly negative.

Motives
I also account for the motives for running for office. In social-movement approaches, motives for party formation are grievances with the current administration, which are to be addressed through new sets of policies. Grievances with the current administration are here understood in two ways: through levels of social vulnerability and of municipal maladministration. The social vulnerability index (SVI) measures differences in levels of development, including levels of education, health, demographic structure and housing (Bolivia 2009). 8 In the Bolivian context, it is particularly important to include this indicator because the indigenous population tends to live in the less developed regions of the country. Low numbers indicate low levels of social vulnerability, that is, grievances should be relatively low.
Municipal administration is composed of three indices provided by the Ministry for Autonomy (Bolivia 2010). First, the efficacy index refers to the degree to which each municipal government uses its resources efficiently in the provision of public services, and hence the absence of waste and corruption, as indicated by a municipality's spending efficiency, fiscal effort, financial independence and investment per capita. 9 Second, I measure administration through the degree of municipal accountability to the national government, that is, compliance with administrative-legal formalities as indicated by the number of cases of frozen accounts and the delivery of annual operational plans and of budget reports. Finally, an index of control processes shows whether municipalities have in place audit processes and procedures to handle cases of embezzlement or mismanagement of funds. As the efficacy, accountability and control indices increase, the better the municipal administration. To build a grievances index together with social vulnerability, I reverse their coding to indicate maladministration, rescale all variables to the same scale, multiply each with their factor loading from a factor analysis and add them together.

Why ethnic parties form 13
In instrumental approaches, in contrast, the incentive for party formation is not policy change but the office itself and its associated benefits. One indicator for the prestige of office and of the control over resources that comes with it is the amount of revenues transferred to the municipal governments from the central state, as reported by the Ministry of Natural Resources (Bolivia 2012b). 10 The allocated revenues per capita range from BOB$ 141.4 to 2,427 across municipalities. As alternative indicators for benefits of office, I also include a measure of local councillors' salary as well as total municipal income and assets in robustness checks.

Population size
Finally, I control for the population size of municipalities. For the instrumental approach, Chandra (2004) points out that the size of the constituency may alter its information environment, and hence the utility of ethnicity as mobilisational tool: in municipalities with as little as 221 inhabitants, who likely know each other by sight, ethnicity is not as useful as an information shortcut as in larger municipalities with as many as 1.14 million inhabitants. For the social-movement approach, population size is often taken to indicate organisational capacity: large, urban municipalities have a better infrastructure that may more easily lend itself to political mobilisation, whereas in small, rural municipalities, the costs of forming, competing and possibly working for a political organisation may be too high.
As an alternative measure for political organisation, I also include a new measure of popular participation reported by the Ministry of Autonomy (Bolivia 2010). Besides voter turnout, the number of women in councils and local participation in associations of municipalities (mancomunidades), this measure also includes the number of territorial organisations (organizaciones territoriales de base, OTB) per municipality. This indicates the strength of grassroots organisations in peasant, indigenous and urban neighbourhood committees (Altman and Lalander 2003).

Results
I use logistic regression analysis to estimate the explanatory power of the different factors for the presence or absence of ethnic parties. To do so, I test the two theories against each other by specifying and running two different models and comparing their fit to the data. This approach was chosen over that of specifying a single, omnibus model with variables from both theories since the preceding discussion has shown that although the two approaches consider similar factors to matter, they matter in different waysfor example, compare how ethnic group size and district magnitude are assumed to relate to each other. A single model would be atheoretical, conflating the different explanations and likely impairing the precision of estimates (Achen 2005;Anaïd Flesken 14 Clarke 2001). Table 2 thus presents the results of two binary logistic regressions for the probability of ethnic party presence as well as of a Vuong closeness test to statistically compare model fits (Vuong 1989). 11 The instrumental model includes office-seeking incentives and the size of the indigenous majority as explanatory variables, controlling for first-mover success with the inclusion of MAS and MIP votes. Here, the only significant factor is the share of MIP votes in the previous elections. As the coefficient is positive, this suggests imitation rather than crowding out. 12 Benefits connected to office and the squared indigenous majority, in contrast, are insignificant. When running the same model without the squared term (not reported)that is, with only the size of the indigenous majoritythe latter becomes significantly positive, but municipal resources are still no incentive to form an ethnic party. Moreover, despite the now significant indigenous majority term, this model does not provide a significantly better fit than the former.
Yet, model fit is significantly better for the social-movement model (Vuong's p = 0.018). Besides MAS and MIP votes, this model includes the highland indigenous population share, a binary indicator for whether this share is larger than the vote share needed to attain a seat, population size, grievances and the population's participation. Here, neither MIP nor MAS success in the past play a role. Instead, the indigenous population share is significantly positive, as expected: the larger the indigenous population, the higher the probability that an ethnic party emerges. Further significant factors are population size and grievances. First, the population size is significantly positive. That is, the larger the municipality, the more likely that an ethnic party emerges. This finding contradicts qualitative studies on Latin America that lead to expect that indigenous parties are more likely to emerge in rural, and hence smaller municipalities, where communities are both more traditional and cohesive. Overall, however, the finding conforms to the expectations in social-movement theory that a larger population brings greater mobilisational capacity: large, urban municipalities have a better infrastructure that may more easily lend itself to political mobilisation of subgroups, whereas in small, rural municipalities, the costs of forming, competing and possibly working for a political organisation may be too high. In contrast, the participation index itself is not a significant factor, possibly due to very low variance in the measure across municipalities. 13 Second, and more strikingly, municipal efficacy affects the probability of ethnic party presence negatively. That is, the better a municipal administration is run, the less likely is the emergence of an ethnic party. The model hence confirms the expectations of the social-movement approach in this regard, too.
One may object that the causal direction is indeed the other way around: that municipalities are administered inefficiently precisely when the indigenous population is politically mobilised because efficient decision-making is more difficult in ethnically diverse polities, where group interests tend to diverge. If that were the case, maladministration should increase when indigenous interests are formally represented in the municipal councils following the 2004 elections. However, we have already seen that there is no significant difference between maladministration scores of municipalities in which ethnic parties attained enough votes to enter government, and hence to affect policy accordingly, and those in which they did not. The same holds for changes in maladministration scores until 2009, the end of municipal government tenure; if anything, maladministration decreased more strongly in the former than the latter (see note 9).

Robustness checks
To test the analysis for spurious correlations, I ran several robustness checks besides those already reported in the endnotes (see Table A1a, online appendix). First, one may argue that in addition to the perceived benefit of holding office, political actors may need to perceive a certain degree of spending discretion over these revenues. Here, we can use again the municipal maladministration indices as indication to actors to what extent they may be able to divert funds for their personal benefit. Yet adding maladministration to the instrumental model does not change the coefficients, and even decreases model fit (AIC = 316.61).
Second, revenues per capita may not perfectly capture actors' expected benefits of office. I hence examined three further indicators for such benefits, collected by Faguet (2012) for the years 1994-6. Two variables indicate Anaïd Flesken 16 municipal wealth: the total local tax and non-tax revenues and the municipal government's total assets. The third measures direct personal benefits through a councillor's salary. When included in the instrumental model, neither of these variables are significant, further confirming that personal benefits of office are not the driving factor in the formation of ethnic parties in Bolivia. 14 Third, to better reflect the theoretical assumptions of the different approaches, I only examined two separate models. One may speculate that the effect of grievances disappears when including measures for office-seeking incentives such as revenues per capita. However, adding these to the socialmovement model does not result in any changes: the highland population share, the population size and grievances remain significant, but the model fit worsens (AIC = 300.07).
Fourth, I tested the possibility that ethnic party formation is, at least partly, a function of between-municipality collaboration of indigenous networks rather than of within-municipality considerations by including department dummies in the models (Table A1b). Doing so does not substantively change the results presented above: the social-movement model remains better fitting (Vuong's p = 0.027) and grievances remain significant (p < 0.001).
Fifth, while the measures of ethnic group size are derived directly from the literature on ethnic parties, rational-choice inspired research on ethnic politics more generally has focused on the number, fractionalisation or polarisation of ethnic groups (Huber 2012). I therefore also ran models including these measures instead of group share (Table A1c). When comparing these alternative instrumental models to the original social-movement model, the latter's model fit does remain significantly better. And when including the alternative measures in the social-movement modelalthough they do not tend to be theorised accordingly in the literaturegrievances remain significant, too.
Sixth, previous research in other countries suggests that different local dynamics affect whether only one or more than one ethnic party is present (Bochsler 2012). I therefore ran the same models with a multinomial instead of a binomial logistic regression, with none, one and more than one ethnic party as the dependent variable (Table A2). Again, the results do not change substantivelygrievances remain an important factor.
But do grievances explain the formation of ethnic parties or indeed the formation of all new parties, whether ethnic or not? I tested this possibility with a model for non-ethnic party entry as dependent variable, keeping all other variables the same (Table A3). Here, only population size as well as revenues per capita in the instrumental model arise as explanatory factors; the grievances index is not significant.

Conclusion
A variety of studies has examined the link between ethnicity and politics, often with underlying assumptions on the location of and motives for agency. This Why ethnic parties form 17 paper has collected and made explicit these assumptions and tested them with new data from the Bolivian municipal elections in 2004. While the size of the respective ethnic group is related to ethnic party formation, the underlying logic here does not seem to be based on the electoral calculations on the part of ethnic entrepreneursan assumption underlying the instrumental approach. Neither are office-seeking incentives driving ethnic party formation, whether in the form of revenues or personal salary. In contrast, grievances with the current administration, as postulated by the social-movement approach, do contribute to ethnic party formation, over and above ethnic group size and other contextual factors such as the population size of the polity. The findings suggest that individual leaders have a stronger agency in party formation than often assumed in social-movement approaches, but that they are not as strongly driven by their own individual interests as often assumed in instrumental approaches. Although this paper does not directly examine the MAS and MIP, as detailed earlier, the findings do align with qualitative accounts of their formation (e.g. van Cott 2005;Yashar 2005). The MAS arose from the coca farmers' movement, which increasingly drew on frames of indigenous tradition and cultural and religious freedom in their defence against government drug eradication programmes. While most coca farmers had indigenous backgrounds, the organisation only became explicitly 'ethnic' and turned into a party in response to grievances against the government. The MIP similarly grew out of peasant and labour mobilisation, but adopting a more radical indigenous discourse than the MAS.
The methodological design of the study has several advantages over other studies: the focus on sub-national variation reduces the noise often found in cross-country analyses, since the general context is kept constant and the different indicators are measured in identical ways. The choice of Bolivia was not only useful for the condition of political upheaval, which allowed analysing ethnic party formation indeed as first-time entry rather than through continued performance, but also because its most-likely character for the presence of both instrumental and social-movement dynamics lends strength to finding their absence. At the same time, the available indicators cannot always clearly distinguish between the two approaches; arguments may be made to include any in both models. While I conducted numerous robustness checks with different data and specifications, more refined measurement (following clearer theoretical considerations) would, as always, be desirable to strengthen the conclusions.
Moreover, the Bolivian context comes with a set of ethnic structures and practices that may not easily be generalised more widely. I therefore conclude with avenues for future research to examine ethnic party formation in other contexts. First, ethnicity in Bolivia is rather fluid: whether one identifies as indigenous or mestizo depends as much on the individual socio-economic situation as it does on the political context. While the year 2004 already saw the revival of indigenous identity in both rhetoric and individual identification Anaïd Flesken 18 (Flesken 2014), the situation may still be different from that in countries in which ethnicity is not as fluid, as in much of Eastern Europe or in South East Asia. Here, different motivations may drive ethnic party formation, but this remains to be established. The present paper suggests focusing more strongly on different types of grievances rather than on ethnic structure or individual actors' interests alone. Second, ethnicity in Bolivia is strongly linked to socio-economic circumstance, that is, Bolivia is a prototypical ranked society (Gisselquist 2013;Horowitz 1985). One may argue that ethnic party formation follows a different logic in unranked societies, as in much of sub-Saharan Africa. Future research may test this possibility through the analysis of ethnic party formation in unranked societies like Kenya or Tanzania.
Both the points on the fluidity of ethnicity and its relationship to socioeconomic circumstance relate to the nature of ethnic parties more generally. This paper used a minimal definition of ethnic parties, categorising only those parties as ethnic that explicitly appeal to voters on the basis of a common ethnic attribute, leaving open the question whether the appeal or indeed the called-upon ethnic attributes have historical roots or are constructed on the spot for instrumental gain. Constructivist theory aligns with either possibility (see e.g. Chandra 2012). Focusing further research not only on different spatial contexts but also on processes over time can shed light on the questions whether and how ethnic parties contribute to, or are the product of, ethnicity construction, or both. Interview or time-series survey data on motivations of both politicians and ordinary citizens could be particularly useful here.
Finally, the paper focuses on municipal elections. While this has methodological advantages and is of substantive interest, considering that local elections are often 'training grounds' for regional and national electoral scenes, explanations of party formation in the latter likely needs to account for several other factors. For example, inter-constituency collaboration and organisational legacies would need to be considered. Future research may do so through in-depth, over-time analyses of selected national-level ethnic parties.
Endnotes 1 For in-depth discussions of the role of political actors in the construction of ethnic cleavages, see e.g. Chandra (2012) and Wimmer (2013).
2 In the following, all organisations registered to compete in the elections are referred to as political organisations or parties.
3 The Bolivian context was hence chosen for analytical, rather than statistical, generalisation (Yin 2013).
4 The scope of the paper does not allow for a more in-depth discussion of the rise of the MAS and the MIP. However, the interested reader may find information on the topic especially in van Cott (2005), Yashar (2005), Harten (2011), Loayza Bueno (2011) and sources cited therein. 5 The 2002 results are only available at the level of the national electoral district, which include up to twenty-four municipalities (Bolivia 2012a). As a robustness check, I also used a municipallevel binary indicator for the majority of MAS or MIP votes, which yielded similar results. 6 Further categories were Guaraní, Chiquitano, Mojeño, 'other indigenous' and 'no indigenous'. 7 For computational purposes, the values were shifted such that the minimum value of all municipalities is 0.
8 Due to data availability, I use the SVI as measured in 2005, that is, 1 year after the municipal elections. But since the index includes several factors unlikely to change within the span of 1 year, I expect that it differs neither substantively nor systematically with the other variables of interest. A comparison of SVI values from 2005 to 2009 shows that between-year differences are indeed very small. 9 Again, due to data availability, municipal efficacy was measured in 2005. Though unlikely in such a short time span, it may be that the presence of ethnic parties affected municipal efficacy rather than vice versa. This was not the case: a two-sample t-test failed to reject the null hypothesis that the mean efficacy scores of municipalities in which ethnic parties attained enough votes (17.248) was less than those in which they did not (19.433, p = 0.151). The same holds for changes in efficacy scores until 2009, the end of municipal government tenure (p = 0.551); if anything, efficacy increased more strongly in the former (by 2.247) than the latter (by 1.309). 10 Revenues, too, were recorded in 2005. While revenue distribution changed rather quickly between years from 2005 to 2009, this is the result of a new hydrocarbon policy put into practice by the Morales administration from 2006; revenue redistribution before that was based on a hydrocarbon law passed in 1996 (Mähler 2007) and should hence have been relatively stable between 2004 and 2005. I thank Annegret Kuhn for providing me with the data. 11 Note that, while a visual comparison of e.g. the Akaike Information Criteria may provide indications of which model better fits the underlying data, it does not show whether the difference in model fit is statistically significant. 12 I ran the same models with a binary indicator of MAS or MIP majority in the 2002 elections at the municipal level as well as continuous measures of MAS and MIP votes in the 2004 elections. While significance varies, effect sizes are very small and overall model fit does not change. 13 The results do not change with alternative indicators for organisational capacity like the number of indigenous communities, campesino communities, neighbourhood councils and their combined count (Faguet 2012). This suggests that the presence of enterprising actors, more likely in municipalities with higher (indigenous) population counts, is more important than organisational capacity itself. 14 As not all data could be clearly matched to the municipalities, due to spatial reforms due to several municipalities having the same names, the number of observations decreases, such that a direct comparison of model fit is not possible.