Administrative Shifts

First, partition created new administrative units, as groups were divided between different entities. In many instances, these moves empowered or created new sets of elites, who replaced or competed with pre-existing leaders who found themselves in different countries. Miles (Reference Miles1994) describes this process in divided Hausaland. On the northern side, the French defeated Damagaram, executed the sultan, and replaced him with his more compliant brother (Taithe Reference Taithe2009). Subsequently, the territory was divided into three provinces, each with its own chef de province. One area – Magaria – was itself split into seven jurisdictions (Miles Reference Miles1994, 98).

These changes tended to undermine leaders’ ability to mobilize group-based collective action. First, in the case of those preexisting leaders who held some degree of power after partition, new boundaries made communicating with and exerting authority over their groups more challenging. Administrative headquarters were sometimes moved away from chiefs’ earlier centres of power; in Hausaland, this centre shifted from Damagaram to faraway Niamey. Further, important cultural roles for chiefs, such as participating in celebrations, could become more challenging across borders. In the nearby Adamawa Emirate, which was also partitioned, the emir, upon learning that two-thirds of his subjects now lived across the border, felt as if the ‘Europeans had cut off the body and left him with only the head’ (Kirk-Greene Reference Kirk-Greene1958, 67–68).

Second, in cases where new authorities were appointed to replace or compete with existing leaders, many of these ‘warrant chiefs’ lacked legitimacy and made a mockery of pre-existing norms. One case involved the elevation of ‘a former palace slave and a eunuch’ (Smith Reference Smith1978, 74, cited in Miles Reference Miles1994, 97). Thus, many installed authorities lacked their predecessors’ ability to organize group-based collective action, subsequently weakening identity. These changes – new boundaries and the creation of competing authorities – were not unique to partitioned groups, but they were more pronounced there, and thus more degrading to group identity.

Revenue Collection

Next, authorities in partitioned areas often saw revenue-collection abilities curtailed, which undermined their authority and, consequently, group identity. New borders separated these chiefs from lands and populations, limiting their ability to extract revenue (Barkindo Reference Barkindo and Asiwaju1985). For example, Lewanika, the Lozi monarch whose territory was partitioned in 1890 between German South-West Africa and what would become British-controlled Northern Rhodesia, was deprived of significant opportunities for tribute (Zeller Reference Zeller2010). Conversely, some traditional authorities, particularly those empowered by indirect rule, found their revenue-generating capacity expanded (Acemoğlu et al. Reference Acemoğlu, Reed and Robinson2014; Ali et al. Reference Ali, Fjeldstad, Jiang and Shifa2019). Partitioned groups, though, typically experienced greater barriers in this area.

Revenue was often key for the maintenance of traditional authority and collective identity in pre-colonial Africa. Given general scepticism toward centralized power (Vansina Reference Vansina1990), taxation and tribute were generally only conceded as part of a social contract, with authority providing collective benefits such as order, enhanced economic productivity, and protection (Robinson Reference Robinson2022). Southall (1985) describes how the Alur came together as a group – and, eventually, an identity – through the collection of tribute for an authority that, in turn, provided conflict mediation, rainmaking, and collective defence. As authorities’ abilities to collect revenues were adversely affected by partition, their ability to extend authority, including through the provision of services, declined, and the social contract frayed. Subsequently, collective identity, which was often constructed around these shared features, declined as well. ‘Warrant chiefs’’ ability to reconstruct these dynamics would be limited by legitimacy concerns. After all, even today, individuals are more likely to contribute to collective efforts organized by traditional authorities over ‘formal’ (for example, elected) ones (Baldwin and Mvukiyehe Reference Baldwin and Mvukiyehe2015; Fanthorpe Reference Fanthorpe2006).

Intra-Group Schisms

Finally, in a similar vein, members of partitioned groups often experienced vastly different economic and political regimes from one another, which made ethnic-based collective action and identification less likely. Groups sometimes saw some elites strengthened by colonial and post-colonial policies – or saw no net changes to their powers – while others in the same group were weakened, or outright removed. Hausaland is again a paradigmatic case, with Hausa authority eviscerated on one side of the border, in Damagaram, and empowered on the other, in the revived Daura Emirate (Miles Reference Miles1994).

Intra-group divisions, in terms of orientations towards the colony (and, later, independent state) and preferred strategies, often widened. Some found it beneficial to cooperate and identify with colonial or state authorities, while others saw resistance and distance as more beneficial (Boahen Reference Boahen1987). In the divided territories of Ovamboland (Dobler Reference Dobler2010) and the Ngoni and Chewa lands (Phiri Reference Phiri and Asiwaju1985), differing colonial legal regimes spurred members of the same ethnic group to migrate in various directions based on economic needs, scattering populations and further weakening traditional leaders’ authority. These differential pulls might also have undermined traditional leaders’ roles as ‘cognitive authorities’ (Martin Reference Martin2002), whose status contributed to the coherence of belief systems and, ultimately, group identity. Collective action declined as interests fractured. In the long-term, these changes – new units and elites, disrupted revenue generation and widened intra-group cleavages – tended to reduce partitioned groups’ opportunities for ethnic-based collective action and, thus, ethnic identity.

Illustrative Cases: The Mandara and Ewondo in Present-Day Cameroon

To illustrate the potential divergent paths of divided and undivided groups, we turn to two groups: the Mandara and the Ewondo. The Mandara Kingdom emerged in the early 1400s and was governed by a Tlikse, who was a unifying figure for the population living south of Lake Chad (Barkindo Reference Barkindo and Asiwaju1985, 30). The experience of the Mandara, who saw their territory divided in 1912–13 between Nigeria and Kamerun (after 1916, French Cameroun), illustrates the three mechanisms discussed above.

First, this partition generated administrative shifts that significantly weakened the power of the Tlikse to govern the Mandara. The Germans in Kamerun empowered the Mandara’s historic ethnic rivals, such as the Fulbe and Kanuri (Mbapndah and Samah Reference Mbapndah and Samah2009, 11), through actions like transferring some Mandara territories, including Disa, Gwozo, and Lufua, to the Fulbe district of Madagali; acquiescing to and even supporting the Fulbe occupation of other Mandara territories; deposing one leader in 1911 and burning his successor’s palace in 1915 (Barkindo Reference Barkindo and Asiwaju1985, 35–37); and even dividing the commercial capital, Kerawa, and other major towns, including Ashigashiya and Banki. The British in Nigeria similarly empowered Kanuri and Fulbe leaders at the expense of the Mandara (Barkindo Reference Barkindo and Asiwaju1985, 40). Colonizers cycled through Tlikwe, instating new ones when they felt the need for local intermediaries with some legitimacy and deposing them when deemed ‘un-cooperative’ (Barkindo Reference Barkindo and Asiwaju1985, 40–41).

Second, tax policies differed on both sides of the border, with rates generally higher on the eastern side of the partition. However, there was also greater inter-community variation in rates on the eastern side, adding another layer of complexity to post-partition governance and, thus, economic regimes. Since rates and collection periods were not harmonized, many Mandara crossed the border to reduce their personal burdens (Barkindo Reference Barkindo and Asiwaju1985, 41–42). Ultimately, this meant that Mandara leaders’ ability to extract tribute through their own regimes was undermined.

Third, the Mandara case highlights how people from the same group were affected differently by colonialism depending on which side of the border they fell, thereby undermining group identity and collective action. Europeans gave sub-groups new appellations. Colonizers attempted to wipe out Mandara as the lingua franca, but prioritized different languages: Kanuri or Fulfulde on the western side, and French on the eastern (Barkindo Reference Barkindo and Asiwaju1985, 38–42). By the 1940s, groups that had formerly been united were engaged in violent clashes over farmland, while smuggling and other border-related criminal activities created new problems (Barkindo Reference Barkindo and Asiwaju1985, 43).

Finally, we note that the story of Mandara is not simply one of decapitation and subjugation. Rather, Mandara frequently resisted colonial rule or ignored the border. Village heads whose titles had not been bestowed by the Tlikwe were seen as illegitimate, and their orders were often only obeyed after force was employed (Barkindo Reference Barkindo and Asiwaju1985, 40). As Barkindo writes, ‘the border was mostly indeterminate, being marked either by single trees or by foot-paths’ (Reference Barkindo and Asiwaju1985, 39). And Mandara clearly retained ties with kin on the other side of the new boundary, regularly attending market days and ceremonies marking births and deaths (Barkindo Reference Barkindo and Asiwaju1985, 43). Still, the central point is not that the border dividing the Mandara had to be impenetrable or the colonizers’ power total. Rather, the experience of partition created obstacles to centralized leadership and collective identity that were, on average, more significant than what non-partitioned groups faced, thus weakening group identity over time.

The experience of the partitioned Mandara contrasts with that of the non-partitioned Ewondo, of what is now the Centre Region of Cameroon. Ewondo chiefs like Atangana collected taxes and compelled labour, while often keeping excess beyond the Germans’ quotas (Rudin Reference Rudin1968, 183). They also rendered justice in the Court of First Instance and enjoyed higher status than other chiefs and village heads (Mbapndah and Samah Reference Mbapndah and Samah2009, 13–15). In these cases, groups’ centralized authority experienced gains – or, at the very least, reduced losses, when compared to some of their peers – in the new colonial dispensation, which could translate to a relatively stronger in-group identity after independence.

We recognize that some non-partitioned groups, such as Dahomey, also saw their institutions significantly weakened by colonialism (Manning Reference Manning1982). However, we argue that, on average, partitioned groups were more likely to experience these degradations than non-partitioned ones, and therefore should exhibit lower levels of ethnic identification today relative to them.

Data and Methodological Approach

We first test the general expectation that partitioned groups have weaker ethnic attachments than non-partitioned ones. We employ data from rounds 3–6 of the Afrobarometer. The unit of analysis is the respondent. The analysis covers about 99,000 respondents and up to 26 countries and 250 ethnic groups over a decade, between 2005 and 2015. The Afrobarometer contains information on respondent ethnicity as well as the relative identification with ethnicity and nation.

Table A2 of the SI reports descriptive statistics. Our dependent variable, Ethnic Identification, is categorical and has five values:

We recode the question such that higher values indicate a stronger attachment to ethnicity.

Figure A1 of the SI shows the distribution of Ethnic Identification. Thirteen per cent of respondents identify ‘mostly’ or ‘only’ with their ethnic group, 47 per cent ‘mostly’ or only’ with their nation, and 40 per cent with both equally. Identification with one’s ethnicity ranges from 3 per cent in Tanzania (Round 5) to 35 per cent in Nigeria (Round 3).

One potential issue with this measure is that it assumes that higher ethnic identification is always associated with lower national identification. Unfortunately, this is the only variable in the Afrobarometer that captures ethnic attachment. However, it is used widely to measure ethnic identification (for example, Eifert et al. Reference Eifert, Miguel and Posner2010; Depetris-Chauvin et al. Reference Depetris-Chauvin, Durante and Campante2020; Gibler et al. Reference Gibler, Hutchison and Miller2012; Higashijima and Houle Reference Higashijima and Houle2018; Robinson Reference Robinson2014). Moreover, there is evidence that, in Africa, ethnic and national identities are negatively related. For example, Miguel (Reference Miguel1994) has shown that, on the one hand, Tanzania was able to weaken ethnic identities by fostering a strong national attachment. On the other hand, in Kenya, ethnic identities remained relevant (or were even strengthened) because of the failure to develop a national identity.

Since the dependent variable is categorical, we employ ordered probit models. The ordered probit model is as follows:

$$y_{i,k,g,w}^* = {X_{i,k,g,w\;}\beta}\;{ +\alpha _k} + \;{\delta _w} + {\mu _{i,k,g,w}}\;$$

where $y_{i,k,g,w}^*$ is an unobserved indicator of ethnic identification for respondent i from country k, ethnic group g and survey-wave w. ${X_{i,k,g,w\;}}$ is the set of explanatory variables, $\alpha_k$ are country fixed-effects, ${\delta _w}$ are survey-wave fixed-effects, and ${\mu _{i,k,g,w}}$ is the error term. These fixed-effects control for temporal and country-level heterogeneity.

Although the true ethnic identification level ( $y_{i,k,g,w}^*$ ) cannot be observed, we can observe ${y_{i,k,g,w}}$ which is the answer to the question asked by the Afrobarometer:

$${y_{i,k,g,w}} = \left\{{\matrix{ {1\;if - \infty {\rm{ }} \lt y_{i,k,g,w}^*{\rm{ }} \lt {\tau _1}} \cr {2\;if\;{\tau _1} \le y_{i,k,g,w}^*{\rm{ }} \lt {\tau _2}} \cr {3\;if\;{\tau _2} \le y_{i,k,g,w}^*{\rm{ }} \lt {\tau _3}} \cr {4\;if\;{\tau _3} \le y_{i,k,g,w}^*{\rm{ }} \lt {\tau _4}} \cr{5\;if\;{\tau _4} \le y_{i,k,g,w}^*{\rm{ }} \lt \infty } \cr } } \right.$$

$$\eqalign{ & P\left( {{y_{i,k,g,w}} = t} \right) = P\left( {{\tau _{j - 1}} \le y_{i,k,g,w}^* < {\tau _j}} \right) \cr & \quad \quad \quad \quad \quad \quad = F\left( {{\tau _j} - {X_{i,k,g,w\;}}\beta \; - {\alpha _k} - \;{\delta _w}} \right) - F\left( {{\tau _{j - 1}} - {X_{i,k,g,w\;}}\beta \; - {\alpha _k} - \;{\delta _w}} \right) \cr} $$

where $F\left( . \right)$ is the standard normal cumulative density function and $t = 1,\;2, \ldots, \;5$ .

Our main independent variable, Partition Dummy, takes the value one if a group has been partitioned during colonization and zero otherwise. The variable comes from Michalopoulos and Papaioannou (Reference Michalopoulos and Papaioannou2016), who use Murdock’s (Reference Murdock1967) distribution of ethnic groups across Africa at the time of European colonization. To be considered partitioned, at least 10 per cent of the area of the group’s homeland needs to be situated in more than one country.Footnote ² They code 229 out of 825 ethnic groups as partitioned, and 45 per cent of respondents are members of partitioned groups. We use alternate indicators of partition in Table A5 of the Supplementary Information (SI).Footnote ³

We include a set of individual-, ethnic group-, and country-level control variables that others have found predict identity salience. First, we take several individual-level variables directly from the Afrobarometer, including education, gender, age, and urban residence. We also include the Respondent’s Wealth, which is constructed by summing the number of items – television, radio, and motor vehicle – the respondent reports owning (we then normalize the index such that it ranges from 0 to 1).Footnote ⁴ This variable has been used in several other studies (Dionne et al. Reference Dionne, Inman and Montinola2014; Houle Reference Houle2015; Østby Reference Østby2008).

We also use the geo-coded information in the Afrobarometer to construct two other respondent-level variables: distance between the respondent and the capital (Respondent’s Distance from Capital),Footnote ⁵ and distance between the respondent and the closest border (Respondent’s Distance from Border). We include these to rule out the possibility that estimated associations between partition and identity stem from the fact that members of partitioned groups are likely to be concentrated in regions further from the capital and closer to borders, which could affect access to state-provided resources and feelings of proximity to power centres.Footnote ⁶ Miles and Rochefort (Reference Miles and Rochefort1991), for example, hypothesize that proximity to the border could decrease the salience of ethnic identities as such boundaries highlight the dichotomy between ‘fellow citizen’ and ‘foreigner’, while Herbst (Reference Herbst2000) suggests that, all else being equal, national identity should grow as distance from borders increases and distance from capitals decreases.

We also include several ethnic group-level control variables. The size of the ethnic group may affect its salience (Posner Reference Posner2004). Moreover, it is important to make sure that results are not driven by the possibility that split groups are smaller and thus perhaps less likely to identify with ethnicity. Therefore, we measure Group Size using the Joshua Project, which gives the proportion of the population of the country that are members of the group. We also control for the area covered by the group’s homeland (Area [km sq.]), as used by Michalopoulos and Papaioannou (Reference Michalopoulos and Papaioannou2016).

While we previously discussed individual-level controls regarding distance from borders and capitals, these same factors might also be important at the group level. Ethnic groups whose homelands are close to borders are more likely to have been split, and groups whose homelands are near borders might have different levels of internal cohesion from groups further from them, regardless of whether they are partitioned. The same logic could apply to groups with homelands close to capitals versus those whose centres are far from them. Thus, we include distance between homeland and the nearest border (Group’s Distance from Border) and a dummy, Capital in Ethnic Homeland. Both are drawn from Michalopoulos and Papaioannou (Reference Michalopoulos and Papaioannou2016).

Finally, we include two country-level controls: GDP per capita logged (Bolt et al. Reference Bolt, Inklaar, de Jong and van Zanden2018), since economic performance could affect the likelihood that an individual sees value in associating with a state, and the Polity score, which ranges from -10 to 10 and indicates political regime type, with higher values indicating more-democratic regimes. We do not include more country-level variables because models already include country and year fixed-effects. We also run other analyses with additional control variables capturing, among other things, a group’s history of political inclusion and exclusion as well as its past involvement in ethnic civil wars.

Since the analysis includes a large number of controls, and it is important to make sure that the results are not driven by post-treatment bias, we show in the SI that results are unchanged when we do not include controls and when we only include individual-level controls (Table A3). The SI also demonstrates that the results are unchanged when we control for the number of years between the survey and the closest election (Electoral Distance, Table A4). Eifert et al. (Reference Eifert, Miguel and Posner2010) find that elections increase the salience of ethnicity.

As mentioned above, to our knowledge, there is only one other cross-national quantitative analysis of the effect of partition on ethnic identities and nationalism. Robinson (Reference Robinson2014), in a broader study of the effects of various factors associated with ‘modernization’ on identities in Africa, finds, among other things, that territorial division is associated with weaker ethnic attachments. Our analysis advances this work in at least five ways. First, we cover four Afrobarometer waves and twenty-six countries, while Robinson (Reference Robinson2014) covers one wave and sixteen countries. This enhances generalizability. Second, we use a measure of partition based on the extent to which a group’s homeland, in the late nineteenth century, was split. Robinson (Reference Robinson2014) operationalizes partition using the percentages of group populations currently living in the same country. Such a measure could be vulnerable to concerns that partition affected these distributions by sparking certain migration patterns and sorting (see Bhandari and Mueller Reference Bhandari and Mueller2019). Our measure is independent of any post-partition developments. Third, and relatedly, we do additional analyses on whether the association is driven by some artefacts of partition. For example, we look at the possibility that divided ethnic groups will tend to be smaller, which could explain why their members are less likely to identify with the group. These analyses enable us to determine that partition itself drives the association, not some of its artefacts. Fourth, we add several key controls to account for the distance between the group/respondent and the border, thereby ruling out borderland effects and the group’s history of political exclusion and civil conflict. Fifth, and perhaps most importantly, our study is the first of which we are aware to develop and test an explanation for why partition decreases ethnic attachment. We also test alternative explanations.

Empirical Analysis

Table 1 reports tests of the association between partition and ethnic attachment. Including both country and year fixed-effects reduces the variation used during the estimation since it only captures within country-year variation (Mummolo and Peterson Reference Mummolo and Peterson2018). Thus, we first estimate the relationship without country and year fixed-effects in model 1. Country and year fixed-effects are introduced in model 2. In both, members of partitioned groups are less likely than those from non-split groups to identify with their ethnicity; the coefficient is significant at the 0.001 level. As explained above, given the nature of the dependent variable, these results should be interpreted as suggesting that partition is associated with higher levels of ethnic attachment relative to national attachment.

Table 1.

Ethnic Partition and Ethnic Identification

Notes: Ordered Probit analyses of the association between partition and ethnic identification. Robust standard errors are in parentheses. ***p<0.001, **p<0.01, *p<0.05.

Based on model 2 of Table 1, the likelihood that members of non-split groups do not identify with their ethnicity at all (that is, they have a score of 1 on the dependent variable) is 34.79 per cent (95 per cent confidence interval: 34.33 –35.24 per cent), while this probability among members of split groups is 37.57 per cent (95 per cent CI: 37.04–38.10 per cent). This association is large relative to that of other variables in the model. For example, the variable that is most strongly associated with identity in model 2 is education. This is consistent with Anderson (Reference Anderson1983), according to whom education fosters nationalism. The difference between members of partitioned and non-partitioned groups (37.57 per cent vs. 34.79 per cent) is about the same as the difference between someone who has completed post-secondary education and someone who has only completed primary education (37.95 per cent vs. 34.87 per cent). Moreover, what is interesting about these findings is not only that partition weakens ethnic identity, but that it does not strengthen it. As discussed above, there are valid reasons to expect partitioned groups should exhibit higher ethnic attachments.

Model 3 of Table 1 redoes model 2 with six additional controls that could be related both to ethnic attachment and Partition Dummy. We do not include these variables in all models because of missing values. All variables are constructed using the Ethnic Power Relations (EPR) dataset.Footnote ⁷ These variables are related to two issues. First, it is possible that groups that have experienced ethnic civil wars are more likely to identify with their ethnicity. At the same time, Michalopoulos and Papaioannou (Reference Michalopoulos and Papaioannou2016) show that partitioned groups are more likely to instigate civil wars. We capture civil war involvement using two variables: Peace Years gives the number of years since the ethnic group has last been involved in an ethnic war, and War History is the number of ethnic wars a group has been involved in since independence. We do not control for whether a group is currently involved in an ethnic civil war because only one group was fighting a war the same year as one of the surveys (Tuareg, Mali).

The second group of variables relates to the group’s access to executive power. Ethnic groups that do control (or that have controlled) the executive may be more nationalist, while those that have been excluded may identify more strongly with their ethnicity (Green Reference Green2020). We include four control variables, constructed using the status variable of the EPR, which reports the political status of different ethnic groups in each government. The first (Political Control) takes the value one if the group is the main one controlling the executive. These groups are those with the statuses ‘Monopoly’, ‘Dominant’, or ‘Senior Partner’. The second (Political Exclusion) takes the value one if the group is completely excluded from power (statuses of ‘Discriminated’, ‘Powerless’, or ‘Self-Exclusion’). The omitted category is inclusion in the government as a ‘Junior Partner’. We also include two other controls for the proportion of years since independence that a group has been the main one in control of the executive (History of Political Control) and the proportion of the years that a group has been excluded from the executive (History of Political Exclusion). As shown in model 3, the association between partition and ethnic attachment (relative to national attachment) is unchanged when these additional controls are added.Footnote ⁸

Robustness Tests

The SI presents additional robustness tests. Table A17 shows that the results are not driven by any particular country. Specifically, we rerun model 2 of Table 1 while excluding each country in succession. Regardless of the country omitted, the coefficient on partition remains negative and significant at the 0.001 level.

One concern might be that our measure of partition, as a dummy, is too blunt. This might be of relevance in two ways. First, individuals from partitioned groups might self-identify differently depending on the extent to which they are separated from their coethnics (Robinson Reference Robinson2014, 730). Namely, individuals who live in the same country as an overwhelming majority of their coethnics might be more likely to identify with their ethnicity, while those who are separated from the vast majority of their coethnics might be less likely to do so. The latter type of individual might simply not see ethnicity as salient because they do not communicate as regularly with coethnics and are likely to be more separated from centres of their group’s power and culture (for example, they might be less likely to interact with traditional leaders from their group). It is also theoretically possible that individuals self-select into separation from their ethnic group because they migrate for economic reasons or have a disaffinity toward their own group (Bhandari and Mueller Reference Bhandari and Mueller2019). They might also be exiled from their group, for violating some kind of norm. Any such individuals would be especially unlikely to identify with their group. If any of these mechanisms were operating, they would weaken the argument that it is partition, per se, that is responsible for the observed correlations rather than a separation from coethnics, which could occur for reasons other than partition. The partition dummy, on its own, does not capture this possibility, in that it treats all individuals from partitioned groups equally.

We explore this possibility in models 1–2 of Table A7 of the SI with a continuous variable, per cent Group Live in Country, which measures the proportion of the respondent’s ethnic group that lives in the respondent’s country and is taken from the Joshua Project. It takes the value one if the group is non-split. For example, the Songhai are split between Niger, Benin, Mali, Ghana, and Nigeria. Since 65 per cent of Songhai live in Niger, per cent Group Live in Country takes the value 0.65 for Songhai living in Niger. Model 2 includes the square of this variable, as well, to capture the possibility that the relationship is nonlinear.

The results, which are illustrated in Figure A5 of the SI, suggest that our results are not driven by individuals who are separated from the vast majority of their coethnics. In fact, ethnic attachments (relative to national attachments) are weakest for individuals who share nationality with about 50–60 per cent of their coethnics. The significance of the partition dummy remains in additional tests, reported in models 4–9 of Table A16, in which individuals whose coethnics overwhelmingly live in other countries (that is, per cent Group Live in Country < 5 per cent) are excluded. Combined, these results suggest our findings are not driven by people who are simply separated, for whatever reason, from the majority of their coethnics, but rather that it is partition itself that matters.Footnote ⁹

The other limitation of the partition dummy is that it does not capture the extent to which groups were divided across multiple countries. The long-term effects of partition might differ depending on whether the group was separated between two countries, or three or more. One possibility is that partition between two countries did not weaken ethnic identity as much as partition between three or more countries because groups found it easier to maintain networks and leadership structures across one boundary than they did across two or more.

We test this possibility with a new variable, # Countries Divided. For example, for the Songhai, this variable takes the value 5. Non-split groups are given the value 1. As shown in model 3 of Table A7, members of groups that are divided across many countries are less likely to identify with their ethnicity. However, model 4 of Table A7 shows that this finding is driven by the difference between partitioned and non-partitioned groups. Model 4 uses five dummy variables for groups that have been divided between two countries (One Other Country), three countries (Two Other Countries), etc. As shown in model 4, there is little difference between the dummy variables. In particular, members of groups that are partitioned between only two countries (as captured by One Other Country) are also less likely to identify with their group.

In sum, we find solid evidence that members of partitioned groups are less likely to identify with their ethnicity than those from non-partitioned groups are. Further, we have shown that the results are not driven by individuals whose coethnics overwhelmingly live in other countries or by groups that have been divided among several countries (as opposed to those that have been partitioned between fewer countries).

We run additional robustness tests. A number of country-level variables may affect ethnic (and national) attachments. In Table A9, we show that the results are unchanged if we add Gini coefficients (Solt Reference Solt2020), trade openness (World Development Indicatorss), state fragility (State Fragility Index), net migration (Migration Data Portal), or ethnic fragmentation (Fearon Reference Fearon2003).

We also rerun the analysis using Probit models (Table A10),Footnote ¹⁰ ordinary least squares (Table A11), multinomial models (Table A12), multilevel linear models (Table A13), and multilevel Probit models (Table A14). Results on our main model (model 2 of Table 1) are robust.

One potential issue with the analysis is that the question we use to capture ethnic identification may be subject to social desirability bias. Respondents may be reluctant to acknowledge that they identify with their ethnicity. While this bias is likely to affect members of split and non-split groups, we adopt two strategies to alleviate the issue, both of which are borrowed from Robinson (Reference Robinson2014). These are presented in Table A15. First, in model 1, we restrict the sample to respondents interviewed by a member of their own ethnic group. As shown by Adida et al. (Reference Adida, Ferree, Posner and Robinson2016), the social desirability bias on ethnic identification is weaker when the interviewer is of the same ethnicity as the interviewee. Second, in model 2, we restrict the sample to respondents who were alone with the interviewer when answering the questions. One could argue that social desirability bias will be stronger when others are present. In both instances, our results are robust. While these results do not eliminate concerns regarding social desirability bias, they are reassuring.

Partition and Traditional Institutions

To test our argument, we construct the variable Traditional Leaders’ Strength, which captures the degree to which respondents believe their group’s traditional leaders and institutions are strong. It is coded using questions taken from round 4 of the Afrobarometer.Footnote ¹¹ These questions capture (1) whether traditional leaders are in charge of providing a list of services (that, is allocating land, maintaining law and order, solving local disputes); (2) whether traditional leaders have a lot of influence on local government; (3) whether the respondents believe that the influence of traditional leaders should increase, decrease or stay the same, which indicates such leaders’ legitimacy; and (4) whether respondents believe traditional leaders should sit on local government councils (see Table A1 of the SI for questions).Footnote ¹² All questions are recoded to range between 0 and 1, where 1 means that traditional leaders and institutions are very strong. The variable Traditional Leaders’ Strength is calculated by taking the average of all the questions.Footnote ¹³ Figure A2 of the SI shows its distribution.Footnote ¹⁴

The mechanism has two steps: (1) partition weakens traditional institutions and (2) members of groups with weak traditional institutions are less likely to identify with their ethnicity. The first step is tested in model 1 of Table 2. Consistent with our argument, we find split groups have weaker traditional institutions. We also find support for the second step in models 2–3. Members of groups with weak traditional institutions are less likely to identify with their ethnic group.

Table 2.

Ethnic Partition and Traditional Leaders’ Strength

Notes: Model 1 tests whether traditional leaders are weaker among groups that have been partitioned. The model is estimated with OLS. Models 2–3 test whether groups with strong traditional leaders are more likely to identify with their ethnicity. These models are tested using ordered Probit. Robust standard errors are in parentheses. ***p<0.001, **p<0.01, *p<0.05.

However, as shown in model 3 of Table 2, it is important to note that we find that partition is associated with weaker ethnic attachment even when we control for Traditional Leaders’ Strength, meaning that it does not account for its full effect. One possibility is that our measure is imperfect and does not fully capture the argument.

Alternate Explanations

We also test two additional possible explanations for the estimated association between partition and post-colonial identity. First, some scholars have challenged the extent to which colonial-era boundary-making was truly an exogenous process (Nugent Reference Nugent2004; Green Reference Green2012; Paine et al. forthcoming). Colonial powers might have explicitly sought to avoid partitioning ethnic groups with more-centralized authority for two reasons. First, such groups might have been able to resist partition more effectively. Second, in cases in which colonizers embraced co-optation strategies, these groups might have been useful particularly because of their strong identities and adherence to hierarchical structures. If Europeans indeed considered these factors, we might see stronger ethnic identities among non-partitioned groups now, not because of boundary decisions but because these groups had stronger identities before colonialism. Like others (Michalopoulos and Papaioannou Reference Michalopoulos and Papaioannou2016), we therefore test the possibility that certain pre-colonial factors that might be associated with post-colonial ethnic identity predict partition.

This possibility is tested in Table 3. We measure pre-colonial institutional centralization using Murdock (Reference Murdock1967). Pre-Colonial Centralization ranges from 1 to 4, where 4 indicates greater political centralization at colonization (for example, pre-colonial kingdoms), while 1 indicates little organization. Figure A4 (SI) shows the distribution of Pre-Colonial Centralization.

Table 3.

Pre-Colonial Centralization, Ethnic Partition and Ethnic Identification

Notes: Model 1 tests whether political units that were centralized were less likely to be partitioned during colonization. The model is estimated using a probit model and the unit of analysis is the ethnic group. Models 2–3 test whether members of groups that were centralized prior to colonization are more likely to identify with their ethnicity today. The models are estimated using ordered probit and the unit of analysis is the respondent. Robust standard errors are in parentheses. ***p<0.001, **p<0.01, *p<0.05.

The argument has two steps: (1) groups with centralized pre-colonial institutions were less likely to be partitioned, and (2) groups with centralized pre-colonial institutions are more likely to identify with their ethnicity today. We estimate each step separately. First, model 1 tests whether centralized pre-colonial institutions reduced the probability of partition. The unit of analysis in this model is the ethnic group (rather than the respondent, as in the other models). Model 1 includes control variables for the main colonial powers because different colonial powers may have had different strategies regarding how they draw borders (taken from Michalopoulos and Papaioannou Reference Michalopoulos and Papaioannou2016).Footnote ¹⁵ We find that Pre-Colonial Centralization has a negative association with partition, but that it fails to achieve statistical significance. These findings are consistent with those of Michalopoulos and Papaioannou (Reference Michalopoulos and Papaioannou2016), who find that pre-colonial centralization indeed does not predict partition. Second, models 2–3 test the association between pre-colonial centralization and ethnic identification. We find that members of groups with centralized pre-colonial institutions are actually less attached to ethnicity today, which is consistent with the previous literature (for example, see Adera Reference Adera2024).

Second, it is possible that authorities under colonial and/or post-colonial regimes anticipated or experienced actual threats from partitioned groups, a number of which were focal points for irredentist collective action (Welch Reference Welch1966; Gondola Reference Gondola2002). In response, these authorities might have targeted integrationist efforts on these ostensibly threatening populations. These efforts might have taken the form of increased spending on local public goods, such as roads and other infrastructure, health clinics, improved sanitary facilities, and water sources. Individual members and communities also might have benefitted from such projects, thereby making them feel more connected with the state and, consequently, the nation. Finally, central authorities might have targeted certain kinds of educational programs at partitioned communities, which might have had dual effects of increasing the perceived benefits of relationships with the state and spreading messages that legitimize the state and nation, at the expense of ethnic identities (Lerner Reference Lerner1958).

We test this possible explanation in Table 4. We use a question included in rounds 3–6 of the Afrobarometer, which asks whether respondents believe their ethnic group is treated fairly by the central government. The variable ranges from 1–4, where 4 means the respondent believes their group is treated very fairly and 1 that it is treated very unfairly. Figure A3 (SI) shows its distribution.

Table 4.

Ethnic Partition and Treatment by the Central Government

Notes: All models are estimated using ordered probit. Models 1–2 test whether members of partitioned groups are more likely to believe that members of their ethnic group are treated fairly. Models 3–4 test whether individuals who believe that their group is treated fairly are less likely to identify with their ethnicity. Robust standard errors are in parentheses. ***p<0.001, **p<0.01, *p<0.05.

Model 1 of Table 4 shows that members of split groups indeed believe that they are treated more fairly, but the association is small and not statistically significant. Model 1 does not include control of the executive and exclusion from the executive (again because of missing values), which are central to the question of whether the group is treated fairly. Groups that have historically controlled the executive should be more likely to believe they have been treated fairly by the central government. When these control variables are included in model 2, members of partitioned groups are found to believe they are treated less fairly, and the association is now highly significant. These results clearly show that members of partitioned groups do not believe they are treated more fairly and are in line with those of Michalopoulos and Papaioannou (Reference Michalopoulos and Papaioannou2016), who find that members of split groups tend to receive fewer public goods.

Models 3–4 test whether individuals who believe that their group is treated fairly are less likely to identify with their ethnicity. We find that they do. Thus, while we do not find support for the first part of this argument – that is, that split groups are treated more favourably by the government – we do find support for the second part – that is, that groups that are treated more favourably identify less strongly with their ethnicity. Therefore, on balance, we find little evidence for these two alternative explanations.