Abstract
Social scientists in general and conflict researchers in particular increasingly combine multiple datasets to study ethnic politics and conflict in Africa. We facilitate these efforts by systematically linking over 8,100 ethnic categories from 11 databases, including surveys, geographic data, and expert-coded lists. Exploiting the linguistic tree from the Ethnologue database, we propose a systematic solution to the grouping problem of ethnicity. An analysis of political exclusion, mistrust of state leaders, and ethnic grievances highlights different ways of linking ethnic categories from multiple datasets. The LEDA open-source software package allows researchers to link ethnic groups from any database with explicit rules and to add their own data on ethnic groups.
Introduction
Ethnic identity constitutes one of the most salient political cleavages in developing countries, in particular in sub-Saharan Africa. Not surprisingly, social scientists investigate the effect of ethnic differences on outcomes such as national identification (Robinson, 2014), trust (Nunn & Wantchekon, 2011), voting (Huber, 2012), and distributive politics (De Luca et al., 2018). Ethnic groups and their attributes have been especially relevant to the study of civil war (Cederman, Gleditsch & Buhaug, 2013; Horowitz, 1985; Østby, 2008; Stewart, 2008) and communal violence (Fjelde & von Uexkull, 2012; Fjelde & Østby, 2014; Hillesund et al., 2018), but also one-sided violence (Fjelde & Hultman, 2014) and international dynamics of ethnic civil wars (Cederman et al., 2013). Combining meso- and micro-level datasets, scholars explore the effects of ethnic group-level characteristics on individual outcomes (Franck & Rainer, 2012), measure group-level attributes through micro-data (Cederman, Weidmann & Bormann, 2015), or enrich one meso-level dataset with information from another (Wig, 2016; Wig & Kromrey, 2018).
When studying questions related to ethnicity, it is inherently difficult to link ethnic categories from two datasets to each other.
1
Due to the socially constructed nature of ethnic identities and different conceptual approaches, we lack a common definition of the universe of ethnic groups in Africa. Thus, any social scientist faces the ‘grouping problem’ of ethnic identities (Posner, 2004a: 850–851). Put differently, each dataset comes with its own list and resolution of ethnic categories. Some, for example the Ethnic Power Relations data (EPR; Vogt et al., 2015), focus on a theoretically Meta-structure of the dictionary approach
In this article, we introduce the Linking Ethnic Data from Africa (LEDA) project. We match more than 8,100 ethnic categories from the 11 most prominent datasets on ethnic groups in Africa to the list of known language families, languages, and dialects from the 16th volume of the Ethnologue database (Lewis, 2009). Using the Ethnologue linguistic tree as a relational master dictionary allows us to link groups at different resolutions, gauge the degree of linguistic overlap between any two groups, and create continuous measures of linguistic distance between them, within and across country borders. Figure 1 depicts our approach with the datasets linked to each other. Online appendix Table A3 provides additional information on the inclusion criteria and substantive contents of these datasets.
LEDA aims to improve empirical research on ethnicity by increasing the conceptual clarity, transparency, and efficiency of linking ethnic data. First, scholars who merge two datasets hard-code several decisions into their data, such as the resolution at which groups are linked or the required degree of overlap between two groups. LEDA allows scholars to explore the robustness of their results to these decisions. Second, matching tables are often not accessible to other researchers, which limits replication attempts. Third, the current fragmentation of links between group lists makes it difficult to leverage the information they contain for linking new group lists to existing ones. With LEDA, researchers who want to establish new ethnic links can draw on the information contained in all prior links.
The open-source LEDA R package 2 allows researchers to query different links between any two existing datasets and to add new data to the language tree, thus creating links to all 11 datasets of ethnic identity that are already covered. This flexibility permits scholars to draw on the large pool of ethnic group-level data when working with geographic or survey data. Thus, LEDA increases the number and scope of research questions that can be studied with currently available and newly collected data on ethnic groups in Africa.
The grouping problem and its solution
The grouping problem of ethnic identities highlights multiple characteristics of an optimal link between two sets of ethnic groups
The first step towards solving the grouping problem is to limit ourselves to linguistic identity categories. Most social science definitions stress subjective beliefs in common descent or (descent-based) membership criteria as defining features of ethnic as opposed to other social groups (Barth, 1969; Chandra, 2012; Weber, 1978). Although individuals in Africa subscribe to multiple putatively descent-based identities including tribe, religion, and race (McCauley, 2014; Posner, 2004b), language is arguably the most widespread ethnic identity marker globally (Gellner, 1983), and is particularly pronounced in sub-Saharan Africa due to, not least, missionary activity (Vail, 1989). More importantly, other ethnic markers often closely align with language. In many African states, language mirrors tribal affiliations at the local level, yielding the smallest identity category with reliable data. The more fine-grained our measurement of the constituent parts of ethnic groups, the easier it is to bridge differences in group definitions between datasets. Our purely language-based approach is leads to false positive matches in contexts where non-linguistic categories are more salient than or further divide linguistic ones. 3 Future research may extend LEDA by adding ethnic categories such as religion or race to the matching dictionary.
The second step of linking ethnic categories leverages the structure of the linguistic tree. This tree is constructed by linguists based on the lexicographic similarity of any two languages/dialects and reflects the ‘genealogy’ of world languages (e.g. Gray & Atkinson, 2003). The language tree helps us to assess the distances between any two languages, which proxy cultural (Fearon, 2003) and genetic distances (Cavalli-Sforza, 1997).
We illustrate the utility of linking different ethnic group lists via the language tree with an example from Ghana in Figure 2. Subfigure 2a depicts the simplified subtree of the Akan language cluster in Ghana (black), comprising the Abron and Akan languages as well as the Ahafo, Asante, and Fante dialects. To the right of the tree, we list four ethnic labels from four lists: the Akan from the Afrobarometer, the Asante/Akan from DHS, the Brong from Murdock’s Map, and the Asante from the EPR data. We link each of these labels to the relevant level on the language tree according to the similarity of the labels and other important clues such as demographic size and information from datasets’ codebooks. In cases in which the appropriate tree-level link was ambiguous, we gave preference to more encompassing links, that is, linking the Akan to the Akan language cluster rather than to the Akan language. Any link to a higher-level language category implies a link to its subsidiary nodes. Thus, linking the Akan from the Afrobarometer to the ‘Akan’ node on level 9 simultaneously links them to the language and dialect nodes below.
Once we have linked all datasets to the linguistic tree, we can merge any two datasets via three systematic rules. Researchers can adopt these rules according to their needs and fine-tune the trade-off between precision and completeness. When the goal is to achieve high levels of precision, researchers will encounter some groups for which no precise links exist. Conversely, keeping as many groups as possible from one dataset comes at the cost of matching groups that are only weakly related.
Importantly, these links can be asymmetric, connecting multiple subgroups in B to a broader superordinate category in A without creating a reverse link. For example, researchers studying economic inequality between ethnic groups might measure groups’ income from survey data and link it to an expert-coded list such as EPR. While the income estimates for large groups in EPR depend on correctly identifying all constituent survey groups, researchers might want to avoid income estimates for a small group in EPR from a large survey category, which comprises many respondents from other ethnic categories than the narrow one that EPR identified as politically relevant.
More concretely, we distinguish between three linking rules. These are implemented in the LEDA R package, which is documented in the Online appendix:
Partial linguistic tree from Ghana and link rules Matched ethnic group lists Because of spelling inconsistencies, groups in the Afrobarometer, DHS, IPUMS, and SIDE lists include ‘duplicate’ entries. Groups that span multiple countries are counted multiple times.
These three general rules allow for specifying the precision and coverage of links between any two group lists within or across countries in a theoretically informed manner that reflects the needs of a research project. Researchers may also explore the impact of alternative linking rules by replicating their analyses across various ethnic links. Lastly, researchers can incorporate measures of uncertainty of any match into their analyses by weighting one-to-many matches by the linguistic distance between group a and linked categories b.
Coding procedure and reliability
The quality of links between any two datasets depends on the quality of their links to the Ethnologue dateset. The main challenge is to correctly match different names or spellings that describe the same category. We link 8,119 distinct ethnic categories from the 11 datasets in Figure 1 and Table I to the Ethnologue tree of African languages that features 15,200 nodes, 2,154 primary languages and 4,822 dialects.
To establish the link between a dataset and the Ethnologue tree, we follow a four-step procedure.
5
First, we use fuzzy string matching to create link suggestions between ethnic categories and Ethnologue entries and their alternative names. Second, we assign all ethnic group lists to research assistants who code and justify links between ethnic categories and language tree nodes. The coders draw on the fuzzy string matches, information on group size, qualitative descriptions in codebooks, Proportion of groups matched to Ethnologue and other lists
Third, an algorithm checks that coded links actually exist in Ethnologue and adds new links as suggestions for ethnic categories with similar names in other datasets. This procedure increases the consistency of our coding across different datasets, while allowing coders to deviate from these automatic suggestions, for example when secondary sources suggest more plausible links. Fourth, we check groups without a match, potentially inconsistent links of groups that share the same name, and inconsistent links of groups that cross borders.
To ensure reliability of our coding decisions, we repeated these four steps, and rotated coders between countries. Between the two rounds we recover 70% of all links. Signaling difficulties in determining the ‘resolution’ of ethnic groups, 20% of all cases differ by language tree level but identify the same broader linguistic category for a group. In about 4% of all cases, we link a language in one of the coding rounds but not in the other. In the remaining 5% of cases, we match ethnic categories to divergent sets of languages. This problem occurs most often in the AMAR dataset, which includes many highly disaggregated (historical) ethnic categories that are hard to identify in Ethnologue. Finally, the authors double-checked the 30% of mismatches in a third round and decided on the optimal match based on the comments and sources provided by our coders and, where necessary, additional investigation.
Moreover, we compare the links between ethnic group lists derived from our coding to three links between the EPR dataset and the Afrobarometer, DHS, and Fearon’s list (Cederman, Weidmann & Bormann, 2015), one link between EPR and DHS (Müller-Crepon & Hunziker, 2018), and another between Murdock’s map and the Afrobarometer (Nunn & Wantchekon, 2011). Using the set overlap rule at the dialect level, we recover at least 90% of these earlier links between ethnic categories. Our recovery rate further increases as we link ethnic categories at lower levels of the tree.
Descriptive results of ethnic group links
After linking all ethnic datasets to Ethnologue, we can match ethnic categories from any two lists to each other. Figure 3 shows that our language-based approach successfully links most ethnic categories from any specific dataset to at least one category in any other dataset. The share of successfully linked groups decreases wherever we match fine-grained ethnic lists from census or survey data to more broadly defined groups.
For each ethnic list pair A and B, we calculate the share of ethnic categories
The remaining columns (3–13) in Figure 3 encode the population share of groups a (rows) successfully matched to groups b (columns). This disaggregation reveals how the choice of baseline ethnic categories matters for the ability to make connections between two datasets. Consider the Afrobarometer to EPR link (row 1, column 6) and the EPR to Afrobarometer link (row 4, column 3). We only match around 83% of the fine-grained ethnic categories enlisted in the Afrobarometer survey data to EPR groups. In contrast, we match essentially all EPR categories to at least one group from the Afrobarometer. Without population weighting, match rates decrease because of fewer matches between many small groups in fine-grained datasets (AMAR, IPUMS, DHS) and groups in datasets with large ethnic categories (EPR and GREG) (see Figure A2 in the Online appendix).
Different types of errors arise due to missing links between ethnic group lists and the language tree. Two broad classes of false negatives exist. First, some definitions of ethnic categories do not have linguistic equivalents in Ethnologue. For example, we could not find a suitable match for the religiously defined ‘Muslims’ in EPR’s group list of Mauritius. 8 Second, some non-matches occur because the list of languages is too detailed. It is often difficult to identify all the constituent languages of big ethnic clusters. For example, we probably miss some of the links between the EPR cluster ‘Hausa-Fulani and the Muslim Middle Belt’ in Nigeria and the hundreds of corresponding Ethnologue languages, many of which have a few thousand speakers only. Conversely, false positives also exist. They affect links between ethnic groups wherever two groups speak the same language but differ along other historical, phenotypical, or religious markers. Important examples include the Hutu and Tutsi in Burundi and Rwanda, as well as Arab and Somali-speaking groups. Researchers should take note of such cases and correct language-based links accordingly.
Empirical illustration
To illustrate the utility of LEDA, we investigate whether exclusion from political power leads African citizens to distrust their political leaders and develop ethnic grievances. While the empirical link between ethnic exclusion and intrastate conflict is well established at the ethnic group level (see e.g. Cederman, Wimmer & Min, 2010), only few, inconclusive findings on the micro-foundations of the underlying processes exist. Most importantly, it remains contested whether individuals reflect objective ethno-political inequalities in perceived injustice and grievances (Hillesund et al., 2018). We use our ethnic links to test whether group-level political exclusion affects subjectively felt distrust of those in power and perceptions of ethnic discrimination as is often assumed in the conflict literature.
Afrobarometer analysis: mistrust in president
Dependent variable standardized to mean 0 and s.d. 1. Control variables include age, age squared, education level indicators, a female and an urban dummy. Standard errors clustered on ethnic group in parentheses. Significance codes: *p < 0.05; **p < 0.01; ***p < 0.001.
Ethnic grievances: unfairly treated by government
Dependent variable standardized to mean 0 and sd 1. Control variables include age, age squared, education level indicators, a female and an urban dummy. Standard errors clustered on ethnic group in parentheses. Significance codes: *p < 0.05; **p < 0.01; ***p < 0.001.
First, we use the set overlap rule requiring that a respondent’s language shares at least one node on the dialect level of the language tree with an EPR group (see Figure 2b). We then construct dummy variables indicating, for each respondent, whether she is linked to an EPR group coded as at least government senior partner. 10 Second, we calculate respondents’ linguistic distance to the closest EPR senior partner group or higher to measure their cultural proximity to the most high-ranking government elites.
We then estimate linear models with country-survey and, in some specifications, ethnic group-fixed effects along with common individual-level control variables (Tables II and III). In line with existing theories, co-ethnicity with government senior partner increases trust in the president (Model 1 in Table II). The estimates imply .25 points greater mistrust on a standardized scale between 0 and 1 among less represented groups. Results remain stable when only exploiting temporal changes in the ethnic composition of governments between survey rounds (Model 2), reducing the risk that our co-ethnicity variables capture unobserved differences between groups. Models 3 and 4 demonstrate that larger linguistic distances to the most powerful ethnic groups similarly increase mistrust in leaders. Notably, we find separate effects when introducing both variables into the same model (Models 5 and 6), suggesting that cultural distance to political power matters beyond direct co-ethnicity.
Results for the more direct measure of ethnic grievances about unfair treatment by the government are substantively similar but statistically somewhat weaker (Table III). The linguistic distance results appear more robust to the inclusion of group fixed effects than our binary measure of political representation (Models 1–4) and the estimates in Model 6 lose statistical significance. Additionally, we conduct the same analysis with data on leaders’ ethnicity from Francois, Rainer & Trebbi (2015). Due to the temporal restrictions of their data, we retain just 6% of respondents from our original analysis. Nevertheless, we still estimate statistically significant and very similar effects if we include only the binary or continuous ethnic representation measure. Estimates from models including both terms show the same pattern but fail to reach significance (Tables A7 and A8 in the Online appendix).
Overall, these results are consistent with the notions that (1) exclusion from power translates into distrust and grievances among ordinary citizens and (2) ethnic dominance by culturally distant elites may spur even stronger frustration than exclusion from power per se. Our findings thus provide novel evidence for the first step of the causal chain that links ethnic inequality in political representation to conflict via widespread grievances among members of disadvantaged groups.
Conclusion
In this article, we introduce LEDA, a new tool that systematically links 11 datasets on African ethnic groups to each other. The LEDA R package facilitates research on the origins and consequences of ethnic identity in Africa and enables scholars to make the most out of existing datasets. Our approach and technical infrastructure also enable researchers to link their own ethnic group data – for example on the ethnic identities of violent actors and their victims – to the language tree and directly combine it with information from all other linked datasets.
More generally, the LEDA project presents a versatile solution to the grouping problem of ethnic identities that permeates existing datasets. As different lists of ethnic groups are based on differing definitions of ethnic identities, linking them becomes cumbersome and often involves non-replicable, arbitrary decisions. Drawing on the tree of languages as a ‘dictionary’, LEDA helps researchers who combine various datasets to address the grouping problem of ethnic identities in a transparent and replicable manner. While currently based on linguistic markers among ethnic groups in Africa, the approach is generally extendable to other world regions and ethnic markers.
Footnotes
Replication data
The R-package and code for the empirical analysis in this article, along with the Online appendix, can be found at http://www.prio.org/jpr/datasets and 
. All analyses have been conducted using R 3.4.
Acknowledgements
We thank Cullen Hendrix and two anonymous referees for excellent feedback and Paola Galano Toro, Vanessa Kellerhals, Benjamin Füglister, Lukas Dick, Carlos Mairoce, and Julian Seitlinger for invaluable research assistance. We are furthermore grateful for comments from Levke Aduda, Matthew Gichohi, and participants of the 2018 AFK Workshop in Hamburg and the 2019 APSA Annual Meeting.
Funding
This research was supported by the Swiss National Science Foundation through the grants P0EZP1_165233 (Carl Müller-Crepon) and P0EZP1_159076 (Yannick Pengl).