Measuring the Effect
of “Ethnicity”
Kanchan Chandra
New York University
Steven Wilkinson
University of Chicago
Most tests of hypotheses about the effects of “ethnicity” on outcomes use
data or measures that confuse or conflate what are termed ethnic structure and
ethnic practice. This article presents a conceptualization of ethnicity that
makes the distinction between these concepts clear; it demonstrates how confusion
between structure and practice hampers the ability to test theories;
and it presents two new measures of ethnic practice—ECI (the ethnic concentration
index) and EVOTE (the percentage of the vote obtained by ethnic
parties)—that illustrate the pay-offs of making this distinction and collecting
data accordingly, using examples from the civil war literature.
Keywords: ethnicity; identity; conflict; civil war; constructivist; theory;
operationalize
What is the effect of ethnicity on violence? On voting? On economic
growth? On democratic stability? On the distribution of public
goods? On the nature of party systems? Measuring the effect of ethnicity on
economic and political outcomes and processes such as these, across countries,
has become one of the most important research questions in comparative
politics in the last 20 years.1
To measure the effect of ethnicity on anything, we require a conceptualization
of the term. First, we need a conceptualization to collect meaningful
data—without it, how will we know that the identities that we count as
ethnic are indeed ethnic and that the identities that we do not count as ethnic
are not ethnic? Second, we need a conceptualization to design reliable
measures for the effect of ethnicity—otherwise, how will we know whether
Comparative Political Studies
Volume 41 Number 4/5
April/May 2008 515-563
© 2008 Sage Publications
10.1177/0010414007313240
http://cps.sagepub.com
hosted at
http://online.sagepub.com
515
Authors’ Note: We are grateful to the several scholars whose work and comments have influenced
the ideas here, and, in particular, to Ericka Albaugh and Lenka B. Siroky, the discussants
on the manuscript at the Conference on “Research Frontiers in Comparative Politics” at
Duke University in April 2007.
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
the assumptions that we make in designing our measures are reasonable?
Third, we need a conceptualization to tell us how to design empirical
models to test for the effect of ethnicity using these data and measures.
Finally, we need a conceptualization to interpret the results of these models.
Most cross-national empirical research measuring the effect of ethnicity,
however, has proceeded independently of a conceptual foundation. We use
data on ethnic identities across the world collected without a definition of
what those identities are. The authors of one of the influential crossnational
data sets on ethnic groups, for instance, declare ethnicity to be “a
rather vague and amorphous concept” (Alesina, Devleeschauwer, Easterly,
Kurlat, & Wacziarg, 2003, p. 160) that resists definition, and they proceed
to count ethnic groups around the world without one. When we do propose
definitions, they do not match our classifications, including those in our
own previous work.2
Our principal measure for the effect of ethnicity—the
index of ethnolinguistic fractionalization (ELF)—employs nonobvious
assumptions not justified by a conceptualization of the term. We design
empirical models somehow, without knowing how the data were generated
and what they mean. Our interpretation of models that employ these data
and measures is similarly ad hoc.
As a result, we do not know what we know: Our studies produce a set of
statistical associations but not meaningful conclusions. We also do not
know how to improve them. Studies of the effect of ethnic diversity on the
onset of civil war, for instance, find that different measures of ethnic diversity
are positively related to conflict (Collier & Hoeffler, 2001), positively
related to low-intensity armed conflict but not civil war (Hegre & Sambanis,
2001), or not at all significantly related to the probability of civil war onset
(Fearon & Laitin, 2003). Studies about the effect of ethnic diversity on democracy
or economic growth are similarly contradictory. If we had some yardstick
for gauging the match between concepts, data, methods and interpretations,
we could adjudicate between these contradictory studies, treat as “provisional
findings” the conclusions of the studies which have the closest match, and then
revise and refine these findings through cumulative research. But the opaqueness
of the ELF measure and the data on which it is based makes cumulative
progress much more difficult. Without knowing what the data mean, or how
the measures, models and interpretations relate to our concepts, it is hard to
know how to improve upon them.
This article evaluates previous efforts to measure the effect of “ethnicity”
using the ELF index in cross-national studies and proposes new measures
based on a new conceptual framework to think about ethnic identity. We summarize
that foundation here, but refer the reader to other work for a defense
516 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
and more fully developed discussion (Chandra 2006, 2007 and Chandra and
Boulet 2006). Throughout, we make a plea for narrowness in our concepts and
measures of “ethnicity.” “Ethnicity,” like “politics,” is a big concept – so big
as to be meaningless. When we want to study how “politics” matters, we do
not construct datasets on and measures of “politics” in general. Instead, we
name some specific aspects of political structure relevant to the outcome—for
instance, democratic v/s dictatorial regimes, presidential v/s parliamentary
systems, majoritarian v/s proportional electoral rules, the effective number of
parties in a country—or political practice—for instance, the content of campaign
rhetoric, the size of government expenditure, the number of roll call
votes—and collect data and design measures that operationalize these narrower
concepts. By the same logic, when studying how “ethnicity” matters,
we must also replace that large concept narrower, more meaningful ones.
The key elements of this conceptual framework are as follows: We define
an ethnic identity as a category in which descent-based attributes are necessary
for membership. Nominal ethnic identities are those for which we possess
the attributes of membership while activated ethnic identities are that
subset of our nominal categories in which we profess membership or are
assigned membership by others. Virtually all social science definitions agree
that descent is somehow important in defining an ethnic group, and many stipulate
additional features, such as a common culture, in combination with
descent. The principal innovation in our definition is in its precise specification
of the role of descent—introduced in the distinction between categories
and attributes—and in its elimination of features other than descent. We then
use this definition to introduce a distinction between ethnic “structure” and
different types of ethnic “practice.” The ethnic “structure of a population consists
of the distributions of attributes in that population and is typically multidimensional.
Ethnic “practice” refers to the ethnic categories activated in
different contexts. The ethnic categories activated in any given context need
not belong to any single dimension, or be mutually exclusive (what we define
later as the problem of overlap). They also need not be exhaustive – they can
describe only some individuals in a population, while others are activated on
non-ethnic identities (what we define as the problem of incompleteness).
Based on this foundation, we argue that even the most careful studies using
the ELF index generate indeterminate conclusions about the effect of “ethnicity”
on a range of outcomes, and we introduce and evaluate two alternative measures
of ethnic practice that improve upon it— “Ethnic Imbalance (ECI)”—a measure
of the representation of groups in colonial governments—and “EVOTE”—the
percentage of the vote obtained by ethnic parties. In sum, our critiques of
studies that use the ELF index are as follows: (1) We do not know what the
Chandra, Wilkinson / Measuring the Effect of Ethnicity 517
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
data it is based on are intended to describe – ethnic structure or ethnic practice.
Either way, they are flawed. If intended to describe ethnic structure, they
are flawed in that they ignore the problem of multidimensionality. If intended
to describe ethnic practice they are flawed in that they ignore the problems of
overlap and incompleteness. (2) We also do not know what the ELF index is
intended to measure-ethnic structure or ethnic practice. Either way, it is
flawed. As a measure of ethnic structure must be able to take multidimensionality
into account. But the ELF index does not accommodate multiple
dimensions. A measure of ethnic practice, it must address the problems of
overlap and incompleteness. But the ELF measure employs the strict assumptions
of mutual exclusiveness and exhaustiveness. (3) The interpretations of
the statistical effect of the ELF index are not consistent with our definitions
and classifications, often within the same body of work. By contrast, EVOTE
and ECI are based on data specifically intended to capture very specific
aspects of ethnic practice and make no claim to capturing ethnic structure (we
think that is much more difficult). As measures of ethnic practice, they do not
make any assumptions about exhaustiveness and exclusiveness: both measures
hold up even when the activated ethnic categories are incomplete and overlap.
And, although they have their own flaws, the definitional and operational criteria
guiding their design are transparent enough to produce justifiable interpretations
and to generate ways to improve upon them. The main contribution
of these new measures is not that they change our previous conclusions about
the effect of “ethnicity” but that they generate meaningful conclusions which
can serve the foundations for better ones.
This article builds on the insights of several important critiques of the ELF
index that have been made in recent years (see Posner 2004, Posner and Laitin
2001, Cederman and Girardin 2007). Posner in particular has both identified
several weaknesses of ELF as well as proposed an alternative index—
“politically relevant ethnic groups,” or “PREG”—to understand the role that
ethnicity plays in explaining economic growth rates in Africa (Posner, 2004).
These critiques are important in their own right and we concur independently
with many of them. Where we differ is in grounding our critique in a single,
integrated conceptual framework that often produces different substantive
criticisms with different implications. Take for instance the previous critique
that the ELF index does not capture many important aspects of ethnic diversity
such as the spatial distribution of groups or the depth of ethnic differences
(Posner 2004, 851). This point is not derived from any particular definition of
ethnic identity- and it holds independently of any particular definition. The
principal implication is that the ELF index may overlook important aspects of
ethnic diversity. By contrast, take the instance of our criticism that the ELF
518 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
index cannot reasonably make the assumptions of exhaustiveness and exclusiveness.
This argument is derived from a particular definition of nominal and
activated ethnic identities. Different definitions might yield different critiques.
Indeed, in another effort to ground a critique of the ELF measure in a distinct
conceptualization of ethnic identity, Mozaffar et al. have identified different
criticisms (Mozaffar 2007, Mozaffar, Scarritt and Galaich 2003). And the
import of the point we make is that the ELF measure is logically flawed as a
measure of ethnic diversity on its own terms, even without taking the spatial
distribution of ethnic identities or the depth of ethnic differences into account.
Further, we are able to evaluate new measures using the same yardsticks that
we use to criticize the flaws in previous work—and judge the substance and
nature of improvements we can make in using them. In that sense, starting
from a fixed conceptual position allows us to go beyond critique to the generation
of new research agendas in a cumulative fashion.
The first section in this article proposes a definition of ethnic identity. The
second section proposes one way to break down the large concept of ethnicity
into its component parts using this definition. The third section uses the
conceptual discussion in the first two sections to evaluate the reliability of
the conclusions about the effect of ethnicity reached by using the ELF index
as a measure. The fourth section uses the same concepts to introduce and
evaluate the measures of ECI and EVOTE, situating them in relation to other
alternatives to the ELF index. The fifth section asks, how might these new
measures change our conclusions about the effect of ethnicity?
What Is Ethnicity?
Since the publication of Donald Horowitz’s Ethnic Groups in Conflict
(1985), there has been a convergence among comparative political scientists
on which identities we classify as ethnic. We agree that ethnic identity
is an umbrella concept that includes identity categories associated with one
or more of the following types: religion, sect, language, dialect, tribe, clan,
race, physical differences, nationality, region, and caste.3
We diverge, however, on why we classify these identities as ethnic—that
is, on the definition that justifies placing identities belonging to these types
and only some identities belonging to these types—in a separate analytical
family. We—that is, social scientists interested in some concept related to
ethnic identity as an independent variable—routinely and interchangeably
use the term identity for a category (a descriptive label that can be used to
sort individuals regardless of their own feelings on the matter) and a group
(a subset of categories that describes a collection of individuals who not
Chandra, Wilkinson / Measuring the Effect of Ethnicity 519
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
only share a descriptive label but also think of themselves as a community)
(Brubaker, 2004). Regarding what makes an identity ethnic, a survey of
definitions in the field reveals that we agree that descent is important but
differ over how to specify the rule of descent and which additional features
distinguish an ethnic identity in addition to descent (Fearon, 2003; Fearon
& Laitin, 2000; Horowitz, 1985; Hutchinson & Smith, 1996). Most social
science definitions specify the role of descent in different ways to mean a
common ancestry, a myth of common ancestry, a common place of origin,
a myth of a common place of origin, or a group descent rule for membership,
according to which an ethnic group is one that is defined by a membership
rule of the following form: You are coded as A if your parents and
grandparents were also coded as As (Fearon & Laitin, 2000). We differ over
other characteristics that might define ethnic identities independently or in
combination with descent. These additional features include a common culture,
a common language, a common history, a common territory, and conceptual
autonomy. Of these additional characteristics, a common culture is
perhaps the one most frequently associated with ethnic identities, so much so
that the term ethnic group is taken to be synonymous with the term cultural
community (see, e.g., Gutmann, 2003; Kymlicka, 1995).
We propose here a new definition. We use the term identity to refer to
categories rather than groups.4
Ethnic identities, in turn, are defined as an
arbitrary subset of categories in which descent-based attributes are necessary
for membership. We say an arbitrary subset because although we can write
out the restrictions that wall off the subset of ethnic identities from the
larger subset of descent-based identities (Chandra, 2006), these restrictions
have not so far been shown to have an analytical value. Nominal ethnic
identities are those ethnic identity categories in which an individual is eligible
for membership based on the attributes that she or he possesses.
Activated ethnic identities are those ethnic categories in which she or he
professes membership or to which she or he is assigned by others as a
member. All individuals have a repertoire of nominal ethnic identity categories.
This consists of all the meaningful membership rules that can be
fashioned from an individual’s given set of descent-based attributes, with
each rule corresponding to a nominal category. The ethnic identity that an
individual actually activates is chosen from this repertoire.
The key innovations in this definition are its specification of the role of
descent and in its emptying the definition of all features other than descent.
In specifying the role of descent, we make a distinction between ethnic
identity categories and the descent-based attributes, which are necessary to
qualify individuals for membership in those categories. These two concepts
520 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
have typically been conflated in the literature, but they are distinct in an
important way. The properties of descent-based categories are derived from
the properties of descent-based attributes but are different from them. In
particular, descent-based attributes have the property of changing over the
long term but being fixed in the short term. An individual’s repertoire of
nominal ethnic identities is also fixed in the short term because it is generated
from a fixed set of descent-based attributes. By contrast, the categories
that we activate based on these attributes can change even in the short term
(although they need not). This difference is critical to how we should think
about ethnic identities.
Consider, for instance, a woman living in New York with attributes such
as dark skin, birth in Trinidad, and descent from parents of African origin
(based on characters described in Waters, 1990). In the short term, we can
take these attributes as being fixed. If she has dark skin now, she is likely to
have dark skin 10 years from now. Because these attributes are fixed in the
short term, so is her nominal repertoire of descent-based categories. This
repertoire includes the categories West Indian, Black, Trinidadian, and so
on, but it does not include the categories Asian or German, for which she
does not have the requisite descent-based attributes. But the categories that
she activates from this fixed set can change, often quite rapidly. She may
well switch back and forth between the identities Black, West Indian,
Trinidadian, and others, depending on the incentives that she faces without
any change in her underlying set of attributes (based on Waters, 1990).
In the long term, even her descent-based attributes might change and with
it, her repertoire of nominal ethnic identity categories. Take, for instance, the
attribute of dark skin. The commonsense interpretation of some shade of skin
color as dark and another shade as light is the product of a humanly created
system of interpretation. These commonsense frameworks can indeed change,
although such changes are typically slow and take place over generations.
The principal justification for this definition is that it captures the conventional
classification of ethnic identity categories in comparative politics
to a greater degree than the alternatives. We substantiate this claim in some
detail elsewhere (Chandra, 2007) by considering the match between each of
the definitions proposed above, including ours, and a representative sample
of sets of identities chosen from across continents that are classified as ethnic
by at least three of four main cross-national counts: Atlas Narodov Mira
(Miklukho-Maklai Ethnological Institute, 1964), Horowitz (1985), Fearon
(2003), and Alesina et al. (2003). This includes Black and White in the
United States; Serb and Croat in the former Yugoslavia; Muhajir, Punjabi,
Pathan, and Baluch in Pakistan; Flemish and Walloon in Belgium; Aymara
Chandra, Wilkinson / Measuring the Effect of Ethnicity 521
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
and Quechua in Bolivia; Yoruba, Ibo, Hausa, and Fulani in Nigeria; and
Zulu, Xhosa, and Coloured in South Africa.
The purpose of constructing a definition by proposing criteria that justify
a conventional classification—rather than stipulating some objective
meaning of the term—is to use it to construct cumulative rather than selfstanding
propositions about ethnic identity. A self-standing proposition
about ethnic identity takes the simple form “Ethnic identity is associated
with Y.” A cumulative proposition takes the more complex form “Consistent
with or contrary to previous research, ethnic identity is associated with Y.”
For self-standing propositions, one can simply stipulate a definition—“By
the term ethnic identity, we mean A”—based on some conception of the
objective essence of the term, independent of its usage or its etymological
structure, or even on entirely arbitrary criteria. As long as the proposition is
consistent with the stipulated definition within a single piece of research,
nothing more is required. But a proposition that builds on or rebuilds the
foundation of previous research must either use the term ethnic identity in
the same way or, when there is a departure, specify the relationship between
past and present usage. Suppose that in previous usage, the term ethnic
identity meant A and the present proposition takes the term to mean B. Then
we cannot make a statement of the form “Consistent with or contrary to previous
work, ethnic identity is associated with Y,” unless we know what the
relationship between A and B is and can be confident that the statement is
consistent with that relationship. Thus, constructing a definition that captures
previous usage as far as possible is essential to moving ahead in a cumulative
fashion.
Given, however, that we indicate that the basis of this classification is
arbitrary—that the restrictions that separate the subset of ethnic identities
from the set of descent-based identities is arbitrary—should we continue to
use the concepts of ethnic and nonethnic identities or discard them in favor
of the concepts of descent-based and non-descent-based identities? The
answer depends on how we trade off the benefits of greater precision with
the cost of theorizing about and collecting data on the new concepts that
would replace it. Descent is correlated to a greater degree with ethnic identities
than with nonethnic identities: Even though all descent-based identities
are not ethnic, all ethnic identities, we have argued, are based on
descent, whereas nonethnic identities are based on many attributes other
than those generated by descent. Thus, the distinction between ethnic and
nonethnic identities does indeed capture a meaningful distinction between
descent-based and non-descent-based identities, although this is an imprecise
one. If we seek to make general claims about average trends across
522 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
multiple countries and over long periods, the gain in precision by using a
new concept may not be worth inventing a new term and reinterpreting the
results of previous research using this term. However, we should recognize
that there is likely to be a lot of unexplained variance, even in claims about
average tendencies, and we should interpret our claims accordingly. If we
seek to make point predictions about a single outcome in a single country
in a single period, then the gain in precision may be large enough to make
replacing the concept worthwhile.
Breaking Down the Concept of Ethnicity
In Figure 1, we diagrammatically outline one way of breaking down the
concept of ethnicity. Each node represents a concept, and each branch connects
this concept to the narrower concepts contained within it. We focus
selectively here on those concepts most relevant to the discussion that
follows. A full representation of this tree would include additional branches
at each node.
At the broadest level, we can imagine the term ethnicity as encompassing
two families of concepts—the structure of ethnic identities and the
practice of ethnic identification. Ethnic structure refers to the distribution
of descent-based attributes—and, therefore, the sets of nominal identities—
that all individuals in a population possess, whether they identify with them
or not. Ethnic practice refers to the act of using one or more identities
embedded in this structure to guide behavior. In other words, it refers to the
set of activated identities that individuals employ in any given context. The
set of activated ethnic identities for any given country is typically a subset
of the identities contained in the ethnic structure.5
The distinction between
ethnic structure and ethnic practice is more precise than but is consistent
with distinctions that others have made between commonsense real and
mobilized ethnic identities, latent and active identities, and so on.6
But
importantly, there is no basis on which to assume, as others have in the past,
that activated identities belong to any single dimension—they could consist
of attributes from several dimensions.
The ethnic structure of a population tends to be fixed in the short term
because it is based on the distribution of descent-based attributes; however,
because ethnic practice is based on activated categories, it can change. For
instance, Little (1998) shows us how even though the ethnic attributes of
members of Kenya’s Il Chamus community have remained largely fixed
Chandra, Wilkinson / Measuring the Effect of Ethnicity 523
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
over the past century, community leaders have activated membership in
different ethnic categories over time, identifying as Chamus in some periods
and as members of the larger ethnic categories of Samburu or Masaai
at others (Little, 1998). Such changes in activated categories, a large constructivist
literature has shown us, can be triggered by incentives created by
political institutions, employment opportunities, social interactions, and so
on. The possibility of change in the identities activated in practice need not
be realized—we might often see stability in the activated ethnic categories
that describe a population for long periods. But such stability should be
524 Comparative Political Studies
Figure 1
Ethnic Structure and Ethnic Practice
Structure Practice
Ethnicity
Activated in
Private
Activated in
Politics
Non
Institutionalized
Politics
Institutionalized
Politics
Nonelectoral
Politics
Electoral
Politics
Party
Behaviour
Voter
Behaviour
Military
(e.g., Wilkinson’s
ECI measure)
Bureaucracy
Other institutions:
(e.g., courts,
educational system,
privatesector)
Implicit
Explicit
(e.g., Chandra’s
EVOTE measure)
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
demonstrated rather than presumed to exist, and where it exists, it should
be treated not as natural fact but as something to be explained.
Over the long term, however, ethnic structure can change, too. Indeed,
there may be a relationship between change in ethnic practice in the short
term and ethnic structure in the long term. Imagine, for instance, that at some
initial point in history, political entrepreneurs work to transform the interpretation
of some descent-based attribute in the ethnic structure—for instance,
taking a shade of skin color commonly interpreted as dark and arguing that
we should see it in a differentiated way, separating light brown, dark brown,
and black. If successful, this enterprise would make a piecemeal change in
ethnic structure in the future. Today’s structure, then, can be in part the product
of the ethnic practice of a distant yesterday, and today’s ethnic practice
can affect the ethnic structure of a distant tomorrow. But at any given time,
we can distinguish between structure and practice.
The ethnic identities activated in practice differ according to whether
they are activated in private life, and the set of ethnic identities that are
activated in political life. By private life, we mean that aspect of life that
concerns individuals alone or their immediate family and friends. By political
life, we mean that aspect of life that concerns collective action by individuals
who are not bound by immediate personal ties. Caste identity in Sri
Lanka (e.g., Goyigama) is an example of an identity that informs private
actions, such as the choice of a marriage partner. But religious and linguistic
identities (e.g., Buddhist or Sinhala) rather than caste are the principal
identities invoked in collective action (Rajasingham-Senanayake,
1999; Tambiah, 1986).
Among the set of identities that are politically activated, we can distinguish
again between identities that are activated in institutionalized politics,
parliament, party politics, the legal system, and so on, and identities that are
activated in noninstitutionalized contexts, such as civil war, riots, and social
movements. In many countries, the set of identities activated in both contexts
may be identical, but in others—especially in states that outlaw certain
types of political participation—they can diverge. In Indonesia in the
1992 elections, for instance, institutionalized participation by political parties
activates religious identities (Muslim and Christian), whereas regional
identities are more likely to be found in the arena of noninstitutionalized
politics. And in the Belgian Congo, until shortly before independence in
1960, the Belgians banned colony-wide political parties—but not tribal and
regional associations, of which there were 317 by 1956—which led to the
structuring of identities in formal politics along quite different lines from
those in associational life (Willame, 1972).
Chandra, Wilkinson / Measuring the Effect of Ethnicity 525
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Among identities that are activated in institutionalized politics, we can
distinguish further between identities activated in electoral contexts, party
politics, and voting behavior and those in nonelectoral contexts, in the corridors
of parliament, the military, the judiciary, and the bureaucracy. In
Uganda, for instance, the identity of Nubian was an identity activated principally
in the military and the corridors of the bureaucracy of Idi Amin’s
regime, whereas the identities Baganda and Catholic have frequently been
activated in the course of electoral politics (Kasfir, 1976, 1979).
Among the identities activated in electoral politics, we can distinguish
between identities that drive voter behavior and those that drive party strategy.
In principle, we should expect there to be some connection between the
identities that parties activate and those that condition voter behavior. But
the two concepts are analytically distinct and may sometimes diverge. In
South Africa in the 1994 elections, for instance, the African National
Congress activates voters based on multiethnic appeals targeted to all South
Africans and in some contexts, on appeals based on the racial category
Black. But voters often voted for it, not on the basis of being South Africans
or Blacks, but on the basis of their tribal identities (Xhosa, Zulu, etc.).
Among identities activated in party politics, we can distinguish between
identities activated using implicit and explicit appeals. The Willie Horton
advertisements used by the Republican Party in its 1988 presidential campaign
in the United States are an example of an implicit appeal to race
(Mendelberg, 2001). By contrast, the election campaign run by Slobodan
Milosevic in 1992, in the immediate aftermath of communist rule, made an
explicit appeal to Serbian identity.
We could go on breaking down ethnic practice further at each node and
conducting a similar breakdown for ethnic structure. But the discussion
above should be sufficient to situate the discussion that follows. By abandoning
the use of the term ethnicity in favor of these narrower concepts, our
goal is not to argue that any one of them works in isolation. They may well
work together. But separating these concepts makes it possible to identify
precise relationships and test for them with some confidence about what we
are testing. Furthermore, we can ask questions about the effect of one
aspect of ethnicity on another, which we would not be able to do if we did
not draw these distinctions in the first place. For instance, once we separate
ethnic imbalance in colonial government from ethnic imbalance in postcolonial
governments, we can ask if one has an effect on the other. If we
separate ethnic imbalance from ethnic mobilization of voters, we can ask if
imbalance is a cause of such mobilization. If we separate the implicit mobilization
of ethnic identity from explicit mobilization, we can ask whether
each type of mobilization has different consequences.
526 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
What Have We Learned From
Previous Work That Uses the ELF index
The principal measure of the effect of ethnicity in cross-national empirical
work is the ELF index, calculated according to the formula 1 – ∑si
2
, where
si is the proportion of the ith activated ethnic category, i = {1, 2, . . . , n}.7
The
ELF index is also the standard measure of the effect of ethnicity in empirical
research, not sampled here, in which ethnicity is not the main focus but is
used as a control variable (e.g., Przeworski, Alvarez, Cheibub, & Limongi,
2000). In both types of work, the index is typically calculated on one or more
sources of the following three sources of data: Atlas Narodov Mira, published
in 1964 by Soviet ethnographers at the Miklukho-Maklai Ethnological
Institute, a data set on ethnic groups in 190 countries published by Alesina
et al. in 2003, and a comparable count of ethnic groups in 160 countries published
by Fearon in the same year (Fearon, 2003).
It is difficult, in much cross-national research, to make inferences about
the effect of ethnicity on dependent variables from studies that employ ELF
and its variants for three reasons: (1) The lack of a match between data and
concept. (2) The lack of a match between measure and concept. (3) The
lack of a match between the interpretation of the analyses using these measure
and data and the concept of “ethnicity.” We discuss each point in turn.
Lack of a Match Between Data and Concept
Structure or practice? It is not clear whether the data sets on which the
ELF index is based capture ethnic structure, defined above as the set of ethnic
categories that are commonsensically real, or ethnic practice, defined
above as the ethnic categories that are activated. Consequently, we cannot
interpret an association between a measure using these data and the outcome
of interest.
The approach taken by Atlas Narodov Mira (Miklukho-Maklai Ethnological
Institute, 1964) is not spelled out. Alesina et al. (2003) refer to ethnic
groups as the product of “persistent identification” (p. 161), thus conflating
structure and practice. Fearon (2003), by contrast, makes an explicit distinction
between the two, aiming to code for ethnic structure by trying to
capture commonsensically real identities—”how people in the country
mentally divide the social terrain in ethnic terms” (p. 203)—and not ethnic
practice, at least in the political realm. But the criteria for operationalizing
this distinction in the coding procedures are not laid out, and if we look at
the data, we find that all three data sets veer inconsistently between collecting
data on ethnic structure and ethnic practice.
Chandra, Wilkinson / Measuring the Effect of Ethnicity 527
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Consider the example of Albania. Each of the three data sets code ethnic
groups in Albania as Albanian, Greek, and Macedonian. But why not
include in their count ethnic groups based on religion (Catholic, Orthodox,
and Muslim) or dialect (Gheg speakers, concentrated in the north, and Tosk
speakers, concentrated in the south)? These other identities also appear to
be commonsensically real. The principal distinction between the excluded
groups and the included ones seems to be that the Albanian, Greek, and
Macedonian groups have more political resonance in the present than do
the others. Inadvertently, then, the data sets appear to be coding practice
rather than structure in the case of Albania.
But in Italy, the pattern is reversed. The Atlas Narodov Mira (MiklukhoMaklai
Ethnological Institute, 1964) codes Italy as being 98% Italian, with
a range of smaller groups making up the remaining 2%, including
Austrians, French, Slovenians, and Albanians. Fearon (2003) codes Italy as
an almost entirely homogeneous country, with a 98% majority of Italians.
Alesina et al. (2003) code Italy in a comparable way, as 94.0% Italian, 2.7%
Sardinian, 1.3% Rhaetian, and 1.9% other. But if we look at politically
mobilized identities in Italy, at least in electoral politics, we see several
other identities mobilized by political parties, including regional identities
(North and South), subregional identities (Milan and Lombardy), and racial
identities (native versus immigrant Italians). Germany, similarly, is consistently
coded by all three data sets as having an overwhelming German
majority, ranging from 91.0% in Alesina et al. to 98.8% in Atlas Narodov
Mira, with several exceptionally small minorities. But why only count
groups on the basis of nationality or race in Germany and not on region—
East Germans, for instance, or Bavarians—which our study of election
campaigns in these countries, discussed in the fourth section, reveals are
identities activated by political parties? By excluding these identities, the
data sets all appear to be attempting to capture some undefined notion of
structure rather than practice.
Structure. If the data refer to ethnic structure, then they are flawed in that
they ignore the problem of multidimensionality and level of aggregation. The
structure of ethnic identities in most countries is multidimensional, although
the number and type of dimensions can vary. The set of identities that are
commonsensically real in the United States, for instance, includes identities
based at least on the dimensions of race, nationality, region, religion, and
tribe. In India, it includes identities based on the dimensions of caste, language,
tribe, region and religion. In Zambia, it includes identities based on
the dimensions of tribe and language. In South Africa, it includes identities
528 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
based on the dimensions of race and tribe. In Malaysia, it includes identities
based on the dimensions of race, language, region, religion, and tribe.
Previous critiques have suggested that in inputting data into the ELF
index, we first need to make a decision about which of these dimensions is
the most relevant one (Posner 2004). But, indeed, why should we assume
that there is only one dimension to choose? If the set of commonsensical
identities in a country is multidimensional, we then need to faithfully
represent that multidimensionality.
Furthermore, categories on each dimension are arrayed at multiple levels—
how do we decide which ones to include and at which level of aggregation
(Laitin & Posner, 2001)? When faced with the dimension of tribe in India,
should we count categories at the highest level of aggregation (e.g., scheduled
tribe) or at the lowest level (Santhal, Munda, Bhil, etc.) or somewhere
in between? When faced with the dimension of religion in the United States,
should we count on the basis of metacategories (Christian, Muslim, Jewish)
or micro categories (Methodist, Baptist, Presbyterian, Shia, Sunni, Ismaili,
Hasidic, Orthodox, etc.)?
Alesina et al. (2003) and Fearon (2003) acknowledge these problems
and discuss them at some length, but they do not furnish the decision rule
that they employ to solve them, and we cannot infer the rule from looking
at their data. To illustrate, consider the coding for India across the three data
sets, displayed in Table 1.
The Atlas Narodov Mira (Miklukho-Maklai Ethnological Institute, 1964)
counts a cluster of groups on the basis of several dimensions—language (e.g.,
Hindi speakers and Tamils), nationality (Portuguese, English), tribe (Santals,
Munda), religion (Jews), caste (Gujars, Jats). But the choice of groups
included from each dimension is arbitrary, and we cannot discern the logic
to the level of aggregation chosen.
Alesina et al. (2003) include two groups from the highest level of aggregation
on the dimension of language (Indo-Aryan and Dravidian). But
given that they are concerned with a count of ethnic groups, it is not clear
why they choose the dimension of language rather than the dimensions of
caste, religion, region, and tribe, and it is not clear why they choose the
highest level of aggregation in this case, in contradiction of their intention
to collect data at as disaggregated a level as possible. (In the same data set,
Alesina et al., 2003, collect data separately on the dimensions of religion
and language but never make clear how these dimensions are separate from,
rather than contained within, the concept of ethnic identities.)
Fearon (2003) reports the ethnic structure as being made up of several
groups defined on the dimension of language—speakers of Hindi (.39),
Chandra, Wilkinson / Measuring the Effect of Ethnicity 529
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Bengali (.08), Telugu (.08), Marathi (.07), Tamil (.06), Gujarati, Malayalam
(.05), Kannada (.04), Oriya (.03), Punjabi (.03), Assamese (.01), and one
group on the dimension of religion, Sikhs (.02). But why not include other
groups on the dimensions of religion (Hindus and Muslims, e.g.) or tribe
(scheduled tribes and others) or caste (at the highest level of aggregation,
this would include upper castes, backward castes, and scheduled castes)?
And on the dimension of language, why not include groups at a higher level
of aggregation—for example, Indo-Aryan languages (including Hindi,
Bengali, Gujarati, Punjabi) and Dravidian languages (including Telugu,
Tamil, Kannada, and Malayalam)?8
In an improvement over the other two data sets, Fearon (2003) does provide
a conceptual justification for his count. He attempts to include groups
that fulfill as many of the following prototypical criteria as possible: First,
membership is reckoned primarily by descent. Second, members are conscious
of group membership. Third, members share distinguishing cultural
features. Fourth, these cultural features are valued by a majority of members.
530 Comparative Political Studies
Table 1
Different Codings for Indian Ethnicity Across Three Data Sets
Groups Included in Count ELF
Atlas Narodov Mira (Miklukho-Maklai Ethnological Institute, 1964)
Hindi-speaking peoples of North India (.25), Bihars, Maraths, Bengals, Gujarats,
Rajastans, Oriya, Panjabs, Assams, Kumaoni, Kashmirs, Bhils, Gujars, Sindhi,
Gurkhi, Pars, Jhats, Shina, Kho, Kohians, English, Jews, Pushtuns, Portuguese,
Telugs, Tamils, Kannara, Malayali, Gondi, Tulu, Oraoni, Kandhi, Kodagu,
Badaga, Irula, Urali, Maler, Mannans, Malavedans, Kurumba, Kadari, Paniabs,
Toda, Kota, Chenchu, Santals, Munda, Ho, Savara, Korku, Bhumidji, Kharia,
Gadaba, Djuangs, Minipuri, Naga, Garo, Balti, Lushei, Kachars, Tipera, Mikiri,
Kirats, Kuki, Tamangs, Ladahs, Bhoti, Thado, Miri, Abor, Mishmi, Dafla,
Limbu, Lepcha, Kanauri, Lahauli, Gurungs, Nevars, Magars, Sherps, Sunvars,
Burmese, Kachins, Chinese, Khamti, Khasi, Nikobars, Burishs, Andamanese .89
Alesina, Devleeschauwer, Easterly, Kurlat, and Wacziarg (2003)
Indo-Aryan (.72), Dravidian (.25), other (3.0) .42
Fearon (2003)
Speakers of Hindi (.39), Bengali (.08), Telugu (.08), Marathi (.07), Tamil (.06),
Gujarati, Malayalam (.05), Kannada (.04), Oriya (.03), Punjabi (.03),
Sikhs (.02), Assamese (.01) .81
Note: ELF = ethnolinguistic fractionalization.
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Fifth, the group has or remembers a homeland. Sixth, the group has a
shared history as a group that is “not wholly manufactured but has some
basis in fact.” Seventh, the group “is potentially stand alone in a conceptual
sense—that is, it is not a caste or caste-like group” (p. 201).
But we cannot tell how these conceptual criteria are operationalized in the
coding process. Determining whether the members of a group have a factual
rather than a fictitious history or whether they value distinguishing cultural
features or whether they are conscious of group membership is no easy task,
even for those who specialize in a particular country or a particular group.
How might a coder make these decisions, and how might others replicate
them? Nor is it clear how many prototypical criteria a group must satisfy to
be included or how a coder should decide between multiple candidate
groups on multiple dimensions that fit the prototypical criteria. Why, for
instance, was the category Jat (included in Atlas Narodov Mira [MiklukhoMaklai
Ethnological Institute, 1964] but not in Fearon’s work [2003]),
which appears to meet the first six criteria but not the seventh, not chosen
over the category Punjabi, which appears to meet the first, third, fourth, and
fifth criteria but not the second, sixth, and arguably, seventh? Some of the
groups included in Fearon’s count do not meet several of the conditions.
Hindi speakers, for instance, are not a group in which members are conscious
of group membership, share distinguishing cultural features that are
valued by a majority of members, and have or remember a homeland.
Furthermore, several groups excluded from this count appear to meet several
of the prototypical criteria, such as Indo-Aryans and Dravidians, Hindus and
Muslims, and scheduled tribes. These differences are consequential: The
ELF index jumps from .42 to .89 depending on the data chosen, and the size
of the largest ethnic group, also a common measure constructed on the basis
of these data sets, ranges from .25 (the size of Hindi speakers in the Altas
Narodov Mira data) to .39 (the size of Hindi speakers in the Fearon data) to
.72 (the size of Indo-Aryans in the Alesina et al. data [2003]).
Practice. If the data refer to ethnic practice—the set of identities that are
politically or otherwise activated—then they are flawed because they overlook
the problem of overlap and incompleteness. Take, first, examples of
overlap. We have no reason to expect that the ethnic categories that individuals
activate in practice should be mutually exclusive. Indeed, in many of the
cases at which we have looked, the mobilized categories are overlapping. In
a count of politically activated ethnic categories in India in the 1991 elections
(described at more length in the fourth section), political parties activated
the following ethnic categories (proportion of the population in
Chandra, Wilkinson / Measuring the Effect of Ethnicity 531
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
parentheses): Hindus (.82), Muslims (.12), Sikhs (.02), other backward
classes (.52), scheduled castes (.16), Jharkhandis (.03), Assamese (.26), and
Tamils (.07). The categories of Hindu, Muslim, and Sikh are mutually
exclusive in relation to one another but overlap with the categories other
backward classes, scheduled castes, Jharkhandis, Assamese, and Tamils.
Sometimes, there are cases of total overlap so that one category is entirely
nested within another. In Belgium, for instance, among the categories
mobilized by political parties are native Belgians, who constitute 91% of
the population, and French speakers, who constitute 42% of the population
and who are largely contained within the category of native Belgian.
Now take examples of incompleteness. There is no rule that individuals
in a population should all activate their ethnic identities exclusively. Indeed,
it is only a few very polarized countries at particular points in time, such as
Yugoslavia in 1992, where almost the entire population lines up behind an
ethnic identity—but even in such countries, the ethnic identification may
not be complete. The data on political parties that we discuss below shows
that 86% of the population in Yugoslavia voted for ethnic parties in 1992,
leaving a minority of voters who voted for other types of parties. In other
countries, we typically see several types of identities activated in practice. In
the 2004 U.S. presidential elections, for instance, some voters activated their
class identities (e.g., middle class), others their party identities (Republican
and Democrat), others based on age (e.g., pensioners), and still others their
racial identities (e.g., Black). But the point is not confined to identities activated
by political parties. If we surveyed individuals in any country about
which identities they were activated on, our guess is that it would be a rare
country in which 100% of the population indicated an ethnic affiliation.
When our data sets produce mutually exclusive and exhaustive counts that
add up to 100% of a population, then what exactly do these data mean?
Lack of a Match Between Measure and Concept
Recall the formula for the ELF index: 1 – ∑si
2
, where si is the proportion
of the ith activated ethnic category, i = {1, 2, . . . , n}. This formula requires
the ethnic categories to be mutually exclusive (i.e., if you are in Ethnic
Category 1, then you are not in Ethnic Categories 2–n) and exhaustive
(every member of the population is in some ethnic category). Given mutual
exclusiveness and exhaustiveness, this index measures the probability that
two randomly chosen individuals from a country’s population belong to different
groups. Thus, a society with two groups, a majority of 80% and a
minority of 20%, would have an ELF index score of 1 – (.64 + .04) = .32.
532 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
A society with several small groups of 25% each would have a higher ELF
index score of 1 – (.0625 + .0625 + .0625 + .0625) = .75.
But as we have noted above, the multidimensionality of ethnic structures
and the overlap and incompleteness associated with ethnic practice violate
one or both assumptions. Thus, we cannot take the ELF index to be a reasonable
measure of either ethnic structure or ethnic practice.
To illustrate the problem that the multidimensionality of ethnic structures
poses for the ELF index as a measure of ethnic diversity, compare
India and Zambia. We note above that at least five dimensions of ethnic
identity are commonsensically real in India, with approximately two to
seven categories arrayed on each, if we confine ourselves to the highest
level of aggregation. In Zambia, only two dimensions of identity are salient:
tribe and language (Posner, 2005). The dimension of language currently has
four groups arrayed on it at the highest level of aggregation: {Bemba speakers,
Nyanja speakers, Tonga speakers, Lozi speakers}. The dimension of
tribe has roughly 70: {Chewa, Tembuka, Bemba, . . . }. Which country is
more diverse given variation in the number of dimensions and categories in
each? The ELF index cannot tell us, because it cannot incorporate these data.
To illustrate the problem that the overlap and incompleteness of ethnic categories
activated in practice poses for the ELF index as a measure of ethnic
diversity, let us return to the example of India. As we note above, in 1991
political parties activated the following categories in the Indian parliamentary
elections: Hindus (.82), Muslims (.12), Sikhs (.019), other backward classes
(.52), scheduled castes (.16), Jharkhandis .031), Assamese (.26), and Tamils
(.066). Given that these categories are neither exhaustive nor mutually exclusive,
putting them into the ELF index yields a nonsensical number.
Lack of a Match Between Concept
and Interpretation of the Measure and Data
Even if we assume that the data and the measure are conceptually justified,
the interpretations of the results using that data and measure are not,
for one or more of three reasons: First, they are inconsistent with our understanding
of what ethnic identity is; second, they do not make a distinction
between the effect of some feature intrinsic to ethnic identity and the effect
of a feature of the context in which ethnicity is activated; and, third, they
treat the ELF index as an all-purpose measure of ethnicity rather than a
measure of one narrow concept related to it.
Interpretations that are inconsistent with our conceptual understanding.
Much of the work that uses the ELF index interprets it as a measure of
Chandra, Wilkinson / Measuring the Effect of Ethnicity 533
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
something exogenous to individual choice. Sometimes, this interpretation is
explicitly stated. Ordeshook and Shvetsova (1994), for instance, make the
remarkable assertion that ethnic heterogeneity “is not a product of individual
choice—rather, it is better portrayed as an exogenously determined
social state” (p. 108). More often, it is buried in the assumptions informing
the design of the study. It is conventional in time-series analyses, for
instance, to treat the ELF index as a constant for the entire period, in sharp
contrast to other variables. Take Przeworski et al. (2000) as one influential
example. The authors are meticulous in constructing annual measures for
democracy and on economic development over a 50-year period and taking
the possibility of endogeneity between the two into account. But the effect
of the ELF index on democracy is calculated and interpreted using the same
data across this 50-year period.
According to the conceptualization of ethnic identity introduced here,
some aspects of ethnic identity can indeed be treated as being constant in
the short term. Ethnic structure, for instance, can certainly be treated as
being exogenous in the short term, whereas ethnic practice can change
endogenously. And even among the ethnic categories activated in practice,
some may be stable for long periods, whereas others change. But for the
reasons outlined above, we have no way of knowing whether our data on
ethnic heterogeneity measure structure or practice. And if it is indeed the
categories activated in practice that are counted, we have no way of knowing
whether they are stable or not. Thus, we cannot support interpretations
based on the claim or the assumption that ethnic heterogeneity is not a
product of individual choice. It may be—but then again, it may not.
Similarly, the relationship between the ELF index and the outcome of
interest is often interpreted as an indication of the relationship between cultural
differences and that outcome. Fearon and Laitin’s (2003) work is simply
one influential example of the body of work that treats ethnicity and culture as
interchangeable concepts. But if we look at the identities that we classify as
ethnic, they are often distinguished only by descent-based attributes—not by a
common culture, however defined (Chandra, 2006). Indeed, many of those
whom we put into the same ethnic category are put there despite their cultural
differences with other members, not because of them. At the same
time, many of the identities that we do not classify as being ethnic (e.g.,
class identities, professional identities) are often distinguished by a common
culture (Chandra, 2006). Take, for instance, those categorized as Black and
as White in New York. The cultural vocabularies of many of those categorized
as Black are incomprehensible to one another (Malcolm X, 1964);
yet, they would classify themselves and be classified by others as members
534 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
of the same ethnic category. And although a middle-class Black may have
been able to perfectly comprehend a White neighbor from her or his class and
educational background, this shared cultural vocabulary would hardly lead
one or others to classify these two individuals as being part of the same ethnic
category. We cannot therefore reasonably interpret a relationship between
the ELF index and some outcome as an indicator of the effect of culture.
Interpretations that do not distinguish between the effect of features
intrinsic to ethnic identity and contextual features. An example may help to
illustrate this point. In an examination of the relationship between different
types of identity in India and violence, Wilkinson (in press) found that religious
identities in India were more likely to be associated with violence
than linguistic identities. Based on what he knew about the data and the
context, he interpreted the finding as one not about the effect of religion per
se but about the effect of institutions. The institutional context in India
provides legalized channels for linguistic mobilization but not religionbased
mobilization. Consequently, religious identities were diverted into
violent channels of protest. In a different institutional context, the argument
implied, the same association might be found between class-based identities
and violence, or language and violence, or ideological identities and
violence, and so on.
This is an instance in which a correlation between a measure of an ethnic
identity and the dependent variable—in this case, a religious identity—
could not be straightforwardly interpreted as proof of the effect of a feature
intrinsic to ethnic identity. Rather, it was the effect of the context in which
that ethnic identity was institutionalized. However, much cross-national
statistical work using ethno-fractionalization indices does not consider in
practice—even if it does so in theory—the distinction between these two
effects, automatically assuming that an association between ELF and the
outcome of interest tells us something about the relationship between
something intrinsic to ethnicity and that outcome.
Interpretations that treat the effect of the ELF index as an indicator of the
effect of ethnicity. Ethnicity, we have argued throughout, is simply a label for
many concepts. The ELF index can at best be seen as a measure of ethnic
diversity (based on structure or practice) but not a proxy for the effect of ethnicity
as a whole. Yet, that is how it is often interpreted. If the ELF index—
or some other individual measure—does not matter, then we argue that
ethnicity does not matter. And if the ELF index—or some other individual
measure—is shown to matter, then we read the results as indicating that
ethnicity matters.
Chandra, Wilkinson / Measuring the Effect of Ethnicity 535
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Among these works, Fearon and Laitin (2003) are perhaps the most
careful in the measures that they use and in interpreting their results. They
use a battery of measures to operationalize several different concepts
related to ethnic identity, including ethnic grievances, linguistic and religious
discrimination by the state, and the size of the largest ethnic group.
But even they sometimes assume that this handful of concepts stands in for
ethnicity broadly defined. In their words, the fact that there is no association
between the ELF index and civil war onset “runs contrary to a common
view among journalists, policy makers, and academics, that holds ethnically
divided [italics added] states to be especially conflict-prone due to
ethnic tensions and antagonisms” (p. 75). In our view, a test of ethnic structure
is not a test of the ethnic division hypothesis. Rejoinders to this work
use the same general language. As one example, consider Cederman and
Girardin (2007), who argue that their N*
index “has a strong effect which is
highly significant, thus casting doubt on the tendency to ignore ethnic politics
as an explanation of civil wars” (p. 173; italics added).
Using the ELF index or any other single measure as an all-purpose measure
of the effect of ethnicity or ethnic divisions or ethnic politics is as
illogical as using the index of effective number of parties—calculated in a
similar way, using the proportion of the vote won by political parties rather
than the proportion of the population made up by ethnic groups—to operationalize
the effect of politics. Suppose that to measure the effect of politics
on economic growth, we specified a model that examined the association
between the effective number of parties and a growth measure. In evaluating
this model, the reader would immediately ask—why is the effective number
of parties the right concept and measure to use? And suppose that we found
that the effective number of parties had no association with that growth
measure. We would hardly accept this as evidence of the claim that politics
does not matter for growth. We might be able to rule out the effect of the
effective number of parties, but there are surely many other ways of measuring
the effect of political competition on growth. By the same logic, the
fact that the ELF index or any other individual measure matters or does not
matter does not tell us whether ethnicity matters or does not matter.
ECI and EVOTE:
Two Alternatives to the ELF index
In recent years, those interested in the effect of ethnicity have begun to
develop new indices and concepts that measure some specific aspect of ethnicity.
Proceeding in chronological order, the Minorities at Risk Database
536 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
now provides data on a range of concepts for over 200 ethnopolitical groups
(see the project’s Web site at http://www.cidcm.umd.edu/mar/). ReynalQuerol
(2002) proposes a new index of ethnic polarization. Scarritt and
Mozaffar (2003) propose an alternative measure of politically relevant
groups that recognizes that there are multiple levels of ethnic identity operating
simultaneously and thus makes precise distinctions between commonsensically
real groups, politicized groups, and particized groups. Posner
(2004) published an index of politically relevant ethnic groups for Africa,
which was designed as an improvement over the ELF index when the variable
of interest is the effect of political competition on growth. Cederman
and Girardin (2007) recently developed N*
, a measure of ethnonationalist
exclusiveness intended to replace the ELF index in “assessments of the role
of ethnicity in internal conflicts” (p. 173). And Bossert, d’Ambrosio, and
La Ferrara (2006) have recently developed the generalized index of fractionalization,
an indicator that can simultaneously measure similarity and
dissimilarity of multiple identities (class, income, ethnicity, race), depending
on which ones we think are theoretically most important. None of these
indices is a one-size-fits-all replacement for the ELF index, nor do any
escape all the problems with it identified above. But each improves on the
ELF index for some class of problems.9
In this section, we review two alternative cross-country measures from our
own work that capture concepts related to ethnicity: First, ECI is a measure
of the imbalance between the representation of ethnic groups in the army,
bureaucracy, and civil services of colonial states. Second, EVOTE is a measure
of the aggregate percentage of the vote obtained by ethnic parties across
countries in 1996. Both ECI and EVOTE capture aspects of ethnic practice:
the representation of ethnic groups in nonelectoral politics (in the case of the
ethnic imbalance measure) and the explicit mobilization of voters by political
parties (in the case of EVOTE). Returning to the diagram in Figure 1, note
that ECI captures an aspect of how ethnic identities are activated in nonelectoral
politics by looking at one institution—the military. EVOTE, by contrast,
captures an aspect of how ethnic identities are activated in electoral politics
by looking at the explicit behavior of political parties.
Ethnic Imbalance (ECI)
In the study of ethnic conflict, many scholars have focused on the relative
degree of inequality among groups—especially in terms of groups’ disproportional
access to or control of the state, the military, the police, education,
and the economy—as an important driver of conflict. The distinction
Chandra, Wilkinson / Measuring the Effect of Ethnicity 537
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
between backward and advanced groups along these dimensions, for
instance, is central to Horowitz’s influential theory of conflict (1985),
which draws on several decades of research by social psychologists on
group categorization and discrimination. Backward groups, Horowitz shows,
“have frequently exhibited severe anxiety about threats emanating from
other groups” (pp. 175-176). One form that this anxiety takes is that of
apprehension about being dominated and turned into
hewers of wood and drawers of water. Conciliatory leaders are being depicted
as being excessively generous in granting concessions to ethnic strangers,
whose intentions are, allegedly, to take control of the country and subordinate
the backward group. Every issue can then become a survival issue. (p. 179)
The imbalances between groups that see themselves as advantaged and
disadvantaged and whose leaders regard their windows of opportunity to
act as being relatively small before their groups face permanent exclusion
from power form a constant theme in much of the comparative literature
on ethnic conflict.
In what follows, we introduce one possible way to directly measure the
imbalances among ethnic groups in key state institutions, derived from
Wilkinson’s ongoing study of the long-term impact of colonialism on ethnic
conflict. Wilkinson is interested in measuring the extent to which the colonial
state left a highly unequal and imbalanced ethnic distribution of power in
government, education, business, the administration, and the army at the
point of independence. It seems likely that directly paying attention to such
ethnic imbalances will be much better than using the ELF index or another
off-the-shelf ethnic data set in explaining why colonies with apparently similar
inheritances in terms of political institution building and the presence of
trained personnel (such as that in India and Pakistan) have done so differently
in terms of their postindependence stability and levels of ethnic conflict.
The most important imbalance in terms of causing conflict, and the one
we focus on here, was that of ethnic imbalance in the military. The overrepresentation
of some groups in the army and the underrepresentation of
others—even if the underrepresented groups were large and influential in
politics or the economy—have been identified by many scholars as an
important driver of ethnic violence, democratic instability, and civil wars in
the postindependence era (Decalo, 1990; Gutteridge, 1962; Horowitz,
1985). Among the rank and file, this ethnic imbalance was often of long
standing—for instance, that in Pakistan, where as a result of recruitment
patterns set in the late 19th century, East Pakistanis (52% of the state’s population)
accounted for only 1% or less of the positions in the army and
538 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
higher administration at independence in 1947 whereas West Pakistanis (in
particular, people from the province of Punjab) had a majority of positions
in the army and administration. In Sudan, just to provide another example,
the overwhelming proportion of positions in the administration and in the
army at independence were occupied by Arabs from the north, rather than
Southerners (Johnson, 1992; Shepherd, 1966).
Ethnic imbalances could also be severe among the officer corps, and the
overrepresented ethnic groups among the officer corps were different from
those groups overrepresented in the ranks. This was the result of the colonial
states’ rapid attempt to localize their previously European-dominated
officer corps in the run-up to independence by employing well-educated
recruits, who tended to come from different ethnic groups than those that
dominated the rank and file (Gutteridge, 1962; Horowitz, 1985; Miners,
1971). In Nigeria, for instance, as in West Africa more generally, many of
the new officers were educated men from the south and port cities, whereas
the bulk of the enlisted men had long been from tribes living in interior districts
in the north (Horowitz, 1985; Miners, 1971).
How might we measure such imbalances? The most appropriate measure
in terms of not only reflecting the theoretical mechanisms in which we
are interested but also simply being easy to calculate and understand is a
slight modification of the ECI, developed by Gayl Ness (1967) in the 1960s
to measure ethnic imbalance in Malaysia.10
As we can see in Table 2, to calculate
a country’s ECI, we simply sum the differences between each
group’s share of the population and its share in the army, government, or
whatever sphere in which we are interested. A score close to zero indicates
a state where there is little ethnic imbalance, whereas a country with an ECI
score of 1.0 or more would be highly imbalanced. Table 2 shows how the
ECI score is calculated: The first half depicts a relatively high-imbalance
country, in which the ECI is 0.96, whereas the second shows a low-imbalance
country, in which the imbalance is 0.10.
Although there are good reasons to think that ethnic imbalances in politics,
the economy, the civil service, and the army are all significant for
explaining the outcomes in which we are interested, reasonable data coverage
is unfortunately available (and with some difficulty) for only the last
two of these.11
However, the army and civil service are clearly key institutions:
They were often the major focus of ethnic competition during the
colonial period, and severe imbalances in either have the potential to lead
to much wider conflicts, as we have seen in Sri Lanka, Pakistan, Uganda,
Kenya, Nigeria, and elsewhere. In fact, it is partly because of the key importance
of these services that we have reasonable data on ethnic balances:
Chandra, Wilkinson / Measuring the Effect of Ethnicity 539
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Colonial legislators asked for breakdowns to redress perceived ethnic balances;
colonial administrations kept records of ethnic balances to make sure
that they were recruiting from “loyal” groups; and postindependence scholars
have, of course, paid attention to the ethnic breakdown of these services
because of their continuing political importance.
One important issue, however, involves how we use data to calculate the
index. How should we identify the relevant ethnic group classifications on
the basis of which to count? Wilkinson uses the ethnic group and territorial
classifications provided in the original colonial data on the military. Using
the colonial classifications provides detailed information on group and
intragroup identities made salient by colonial policy and boundaries that
may be politically meaningful but may not match at all the ethnic categories
that are used to generate indices such as the ELF index. They are likely to
track the identities and imbalances that are actually related to conflict rather
than those that are simply meaningful to anthropologists or linguists, even
if these salient identities are also endogenous to state policy.12
Calculating the ECIs for this article provides some striking data on the
extent of the ethnic imbalances in many countries’ armed forces at independence.
In Kenya, for instance, as a result of the British recruitment from
groups regarded as martial, the ECI at independence was quite high: 0.96.
This reflected the heavy recruitment from groups such as the Samburu
540 Comparative Political Studies
Table 2
Ethnic Concentration Index (ECI)
Proportion of Proportion of
Population Occupation ∆
High-imbalance country
Group A 0.50 0.10 0.40
Group B 0.25 0.50 0.25
Group C 0.13 0.05 0.08
Group D 0.12 0.35 0.23
Sum ∆ 1.00 1.00 ECI = 0.96
Low-imbalance country
Group A 0.25 0.30 0.05
Group B 0.25 0.20 0.05
Group C 0.25 0.25 0.00
Group D 0.25 0.25 0.00
Sum ∆ 1.00 1.00 ECI = 0.10
Note: After Ness (1967).
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
(11% of the population, 36% of the army) and Kalenjin (11% of the population,
26% of the army) and the very low levels of recruitment from the
Kikuyu, Embu, and Meru (27% of the population, 3% of the army; Parsons,
1999).13
In Singapore, the imbalances were even more striking (ECI = 1.31)
because of the heavy recruitment of the minority Malay community to the
security forces, in which 20% of all Malay men were employed, compared
to only 4% of the general population (Bedlington, 1981).
Another state that was profoundly imbalanced during the colonial period
was Iraq (Table 3). When the British organized the Iraqi military in the
1920s, they decided to preserve the Sunni bias of the Turkish provincial
forces from which many of the Iraqi soldiers and officers were drawn. As we
can see from the table, which displays the composition of the Iraqi officer
corps in the colonial period, the recent struggle between a Sunni-dominated
military state and a non-Sunni majority is emphatically not just a product
of the Saddam Hussein regime. Iraq’s officer corps in the colonial period
had an ECI of 1.46, reflecting the overwhelming Sunni dominance in the
army. This massive imbalance in the officer corps was largely replicated in
the ranks, prompting Shia protests throughout the later colonial period over
discrimination against their community (Tarbush, 1982).
We provide some comparative statistics in Table 4 to show how the ECI
rankings for various colonial armies at independence compare to those in the
more widely known ELF index, as well as the improved index of ethnic fractionalization
(EF) data set and index of cultural difference (CULDIF), developed
by Fearon (2003; see also Fearon’s Web site, http://www.stanford
.edu/~jfearon/). The EF index uses improved data on group proportions to
recalculate the EF data, and the CULDIF index is Fearon’s effort to measure,
using data on language, the degree of cultural difference in a population.
Chandra, Wilkinson / Measuring the Effect of Ethnicity 541
Table 3
Ethnic Imbalance in the Colonial Iraqi Military
Proportion of Share in
Population Military ∆
Kurds 0.18 0.05 0.13
Sunni 0.16 0.90 0.75
Shia 0.63 0.05 0.58
Turcomans 0.02 0.00 0.02
Sum ∆ 1.00 1.00 ECI = 1.46
Source: Wilkinson database (see Chandra, 2006, 2007, in press), including Tarbush (1982).
Note: ECI = Ethnic Concentration Index.
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
The ECI and these other indexes are clearly measuring different things,
as we can see from the scatter plot between ECI and EF in Figure 2, as well
as from the correlation coefficients between ECI and these other indices:
The correlation coefficient between the ECI data and the EF index is .43
and between ECI and CULDIF is –.39. The potential value of the ECI as a
likely predictor of conflict is suggested when we compare Australia and
Burundi, which are both at the low end of the range on the EF index (.14
and .33) and ELF index (.31 and .03)—though not on Fearon’s CULDIF
indicator—but which look completely different on the ECI data, with
Australia at .16 and Burundi at a very high 1.00.
The comparison between India and Pakistan on these three measures is
also revealing. Why has India done better than Pakistan since independence
in terms of internal levels of conflict? The conventional measures do not help
us here, given that India is more fractionalized and has a higher CULDIF
level than does Pakistan. But ECI does seem to help explain the difference.
In Pakistan (ECI = 1.30), the imbalance between the Punjabi-dominated east
542 Comparative Political Studies
Table 4
A Comparison Between the Ethnic Concentration Index (ECI)
and Other Ethnic Fractionalization (EF) Indices
ECIa
ELFb
CULDIFc
EFd
Australia 0.01 0.32 0.15 0.15
Nigeria 0.20 0.87 0.66 0.80
Malawi 0.42 0.62 0.29 0.83
Syria 0.58 0.22 0.24 0.58
Sri Lanka 0.59 0.47 0.39 0.43
Ghana 0.77 0.71 0.39 0.85
Kenya 0.96 0.83 0.60 0.85
India 0.99 0.89 0.67 0.81
Burundi 1.00 0.04 0.04 0.33
Lebanon 1.03 0.13 0.19 0.78
Burma 1.23 0.48 0.42 0.52
Pakistan 1.30 0.64 0.29 0.53
Singapore 1.31 0.42 0.39 0.39
Uganda 1.53 0.90 0.65 0.93
a. Ethnic imbalance, as measured by a modification of the ECI (Ness, 1967), based on
Wilkinson’s data (Chandra, 2006, 2007, in press).
b. Ethnolinguistic fractionalization, as measured by the ELF index (Atlas Narodov Mira;
Miklukho-Maklai Ethnological Institute, 1964).
c. Cultural difference, as measured by Fearon’s (2003) index of cultural difference.
d. For these data, based on the EF index, see Fearon’s Web site: http://www.stanford.edu/~jfearon/.
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
and Bengali west was at the heart of the Bangladeshi campaign for independence
and the 1970-1971 civil war, as well as in subsequent conflicts in
West Pakistan between Punjabis and the smaller ethnic groups: Rizvi
(2000) points out that even today, although
Pakistan officially discarded the British concept of martial races for recruitment
to the Army and somewhat expanded the recruitment base . . . ethnic
imbalances persist in the Army. The Punjab continues to provide the bulk of
officers and other ranks; unofficial estimates put its share as 65 and 70 per
cent respectively. (p. 240)
India, however, had a considerably lower ECI figure at independence
(0.988 compared to 1.303), a reflection of the fact that although India, like
Pakistan, inherited a heavily Punjabi army (the state of East Punjab
accounted for 42% of the Indian military in 1947), the Punjabi proportion
was nowhere near a majority of the Indian army. Moreover, the category of
Punjabi in the Indian army, unlike that in Pakistan, was cross-cut by an
Chandra, Wilkinson / Measuring the Effect of Ethnicity 543
Figure 2
Scatter Plot of the Indices of Ethnic Fractionalization (EF)
and Ethnic Concentration (ECI)
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
important religious cleavage between Punjabi Hindus (19% of the army)
and Punjabi Sikhs (23%) and by the fact that the Punjabi presence in the
army was balanced by a broader representation of other states than that in
Pakistan.14
The dominant groups in such major states as Uttar Pradesh and
Bombay, for instance, were represented in the army more or less in proportion
to their population.
In other countries too the ECI measure appears to make more sense than
the other indices when we think about the relationship between ethnic practice
and conflict. Malawi’s low level of ethnic imbalance (ECI = 0.42) in its
army helps to explain its relative stability since independence despite its
high level of ethnic fractionalization (EF = 0.83). Burundi’s, Sudan’s, and
Lebanon’s high levels of ethnic imbalance in their armies at independence,
however, may help us to understand why they have been so unstable since
independence.
Perhaps the most puzzling cases in the above list—which go against the
general expectation that conflict rises with ECI—are those of Singapore,
Sri Lanka, and Nigeria. In Singapore, the explanation for low conflict
despite a high imbalance at independence seems to be that the imbalance
was immediately recognized as a potential major problem by the new,
Chinese-dominated government, which rapidly reduced the Malay proportion
in the Singapore security forces during the 1965-1971 breathing space
provided by the staged British withdrawal from its military base in
Singapore (Bedlington, 1981). Sri Lanka and Nigeria both have relatively
low imbalances in the army at independence but high postindependence
conflict. The explanation in both cases seems to be that the ECI, useful
though it is, gets at only some of the most salient aspects of ethnic practice.
In Nigeria, the relatively mild northern imbalance in the rank and file
fails to take account of the early dominance of southerners in the officer
corps; in 1960, only 14% of the officers were from the north. It was a mismatch
that led to several coup attempts in the 1960s as southerners sought
(unsuccessfully) to take action to prevent the likelihood of northern dominance
(Horowitz, 1980; Miners, 1971).
In Sri Lanka, some of the answer to the puzzle comes from the fact that
the Tamils have a significant intragroup cleavage between Sri Lanka
Tamils and the separate Indian Tamil community (11% at independence)
who were brought over in the century before independence to work the
plantation economies in the central highlands and who had virtually no
representation in the army and civil services. By counting both groups as
part of the single category Tamils, we somewhat underestimate the ECI for
the Sri Lankan military. But the more important part of the puzzle is that
544 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
the imbalance in the military that Wilkinson is able to capture is only one
part—though an important part—of the overall ethnic imbalances that
drove the conflict. As Tambiah (1986) has pointed out, much of the conflict
in Sri Lanka, whose education system has produced many graduates, has
focused on the threat (from the Sinhalese perspective) of Tamil overrepresentation
in medicine, administration, and the professions, not just in the
military and civil services. It was discrimination against Tamils in government
service and higher education in the 1960s and 1970s that created the
environment in which the movement for Tamil independence flourished,
especially among alienated Tamil high school and college graduates.
EVOTE: Aggregate Vote Obtained
by Ethnic Parties Across Countries
EVOTE is a time-sensitive measure of the aggregate vote obtained by
ethnic parties across countries. It is one of over a hundred variables from
the Constructivist Dataset on Ethnicity and Institutions (CDEI) project,
constructed by Chandra in collaboration with graduate students at MIT and
New York University. The following discussion is based on a crosssectional
version of the data set, which includes that on EVOTE for over a
100 countries in the lower-house legislative election closest to but before
1996. We chose these 100 countries on the basis of whether they held
competitive elections—measured by Keefer’s (2005) index of legislative
competitiveness—but are now expanding the coding across space to cover
any country which held an election, competitive or not, and over time.
EVOTE is constructed as follows: First, we classify each political party
in each country for which we can obtain data as ethnic, multiethnic, or
nonethnic on the basis of its campaign in the legislative election closest to
but before 1996. Then, we add up the percentage of votes obtained by all
ethnic parties in a given country. Thus, EVOTE for Country A is constructed
as follows: EVOTE96 (Country A) = Vote for Ethnic Party 1
(Country A) + Vote for Ethnic Party 2 (Country A) + Vote for Ethnic Party
3 (Country A) + . . . Vote for Ethnic Party n (Country A). In principle, the
value of EVOTE can range between 0% (for countries with no ethnic parties)
to 100% (for countries in which all votes are captured by ethnic parties).
In reality, EVOTE for the year 1996 ranges from 0.00% (e.g., in
Greece) to 85.63% (in Yugoslavia), with a mean value of 12.95%.
The classification of parties, which is the foundation of EVOTE, is based
on the definitions proposed in Chandra (2004), according to which an ethnic
political party is
Chandra, Wilkinson / Measuring the Effect of Ethnicity 545
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
a party that represents itself to voters as the champion of the interests of one
ethnic category or set of categories to the exclusion of another or others, and
makes such a representation central to its strategy of mobilizing voters. (p. 3)
The key aspects of this definition are as follows: exclusion, an ethnic
party must make an appeal on behalf of some ethnic group (or groups) that
excludes others; explicitness, the appeal must be open; and centrality, the
appeal must be central to its mobilizing strategy. A multiethnic party also
makes an open appeal related to ethnicity central to its mobilizing strategy
but assumes a position of neutrality and equidistance toward all relevant
groups. In other words, it differs from an ethnic party only in its inclusiveness.
A nonethnic party is one that does not make an ethnic appeal central
to its mobilizing strategy. Note that these definitions classify parties on the
basis of their messages, and because messages can change across elections,
they are time sensitive: A party classified as an ethnic party in one election
need not be classified the same way in subsequent elections.
The concept of EVOTE attempts to measure an aspect of ethnicity related
to the behavior of political parties: the use of an explicit ethnic appeal. It does
not capture implicit behavior. By restricting EVOTE to explicit appeals, we
are not taking the position that implicit appeals are unimportant—simply
that they are different. Separating implicit and explicit appeals makes it
possible to test for the effect of one on the other and of each on other outcomes,
separately. Indeed, several other variables in CDEI, not discussed
here, capture the implicit ethnic appeals made by political parties.
Even among explicit appeals, EVOTE is only one of the many ways of
measuring the explicit behavior of political parties. Others in this family
include variables that code the aggregate vote captured by nonethnic and
multiethnic parties, names and sizes of the ethnic groups explicitly mobilized
by political parties, the types of identities explicitly mobilized by
political parties, the number of identity types mobilized in each country, the
proportion of an explicitly mobilized ethnic group that votes for its ownparty,
and so on. These other measures of explicit appeals are all separately coded
in CDEI, but here we focus simply on EVOTE.
The coding of the parties is based on a content analysis of the election
campaign of the party in question using four sources: the Europa World
Yearbook, the Political Handbook of the World, news sources from the
Foreign Broadcast Information Service, and Lexis-Nexis searches. For
each party, we obtain a sample of campaign materials (speeches at election
rallies, policy pronouncements, and so on) as reported in Foreign Broadcast
Information Service and Lexis-Nexis for a period up to 3 months before
the election date. These include reports from the international media and
546 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
translations of local news reports from newspapers, radio, and television.
These samples have three advantages: First, they are primary sources that
report what parties are actually saying to voters rather than what they print
in their manifestos. Second, many of these sources are translations of what
parties say to voters in local languages. Third, they are time-sensitive
sources that report party statements for the year of election. Where the
samples are too small to permit reliable codings, we turn to local newspapers
and secondary sources as a last resort. These sources give us a sample
of articles for the election platform of each party. We archive the materials
for each party for each country after completing the coding. This archive,
composed of a uniform set of source materials for each observation
(country or party), makes it possible to doublecheck old variables as we
proceed and to construct new variables as they become important.
The coding is based on a protocol that establishes rules for the identification
of an appeal as being ethnic, explicit, or central. The content analysis is
qualitative rather than quantitative. Thus, rather than establish centrality simply
by counting the number of times that an issue is mentioned, the protocol
identifies rules of interpretation for centrality—an ethnic appeal can be central
on the basis of an issue’s frequency but also the way in which an ethnic
appeal is used. For instance, a party that associates an ethnic category
with ownership of the state would be coded as an ethnic party even if statements
to this effect were not frequent, on the basis of the reasoning that
once such a statement is made, it colors the interpretation of other statements.
If, given the content analysis, we find that a political party makes an
open and exclusive appeal to some ethnic category or set of categories and
that such an appeal is central to its campaign, we code it as an ethnic party.
If we find that a political party makes an open and inclusive appeal to all
ethnic categories that define a population and makes such an appeal central
to its election campaign, we code it as a multiethnic party.15
And if we find
that a political party does not make an open appeal or a central appeal to an
ethnic category, whether exclusive or inclusive, we code it as a nonethnic
party. We document each coding for each party on each country, compare
it with codings in other data sets where available, check for consistency
across coders, and assign a reliability score to the coding (1 = high certainty,
2 = moderate certainty, 3 = low certainty) based on the quality of
information in the sample. Because countries can sometimes have over a
hundred parties, each with a separate sample, this level of documentation
adds considerably to the time required to construct this data set. But it is
important if other researchers are to replicate our efforts and estimate the
bias and error in the data.
Chandra, Wilkinson / Measuring the Effect of Ethnicity 547
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Consider the case of India as an example of our coding procedures.
Hundreds of parties competed in India in the 1991 parliamentary elections
(the elections closest to but before 1996), but most of them obtained a
miniscule percentage of the vote. We obtained disaggregated data on all
parties that obtained at least 0.01% of the vote, thereby including 66 parties
in our data set. We then coded each of these 66 parties on the basis of a content
analysis of its party platform. Of the 66 parties, we coded 13 parties,
accounting for 51.81% of the vote, as nonethnic and 18 parties (38.95% of
the vote) as ethnic, but we were not able to find sufficient articles on election
platform to code the remaining 35 parties, accounting for 10.24% of
the vote (these were small parties, with a mean vote of 0.14%).
Our purpose in constructing EVOTE and related variables is to investigate
two principal questions: First, what explains variation in the incidence
and performance of ethnic parties across countries? Second, what is the relationship
between the explicit political mobilization of ethnic identity and
political stability? However, this measure is also of value to social scientists
interested in exploring the effect of the politicization of ethnic identities, at
least as measured through the party system, on a range of outcomes. One
such set of questions is the following: What is the effect of the explicit
politicization of ethnic divisions (measured using EVOTE) on some outcome
of interest, including war, riots, economic growth, public policy, welfare
spending, and so on? Replacing ELF with EVOTE in this body of work
would be a meaningful test of whether one concept—namely, the degree to
which ethnic identities are explicitly politicized in the party system at a particular
point—matters in explaining any of these outcomes, while leaving
open the possibility that some other aspect of ethnicity might also matter in
ways not captured by EVOTE.
Other questions that can be addressed using EVOTE include the following:
Is the politicization of particular types of ethnic divisions (e.g., region,
religion, language, tribe) associated with particular types of outcomes? This
would entail using the percentage of votes won by regional, religious, linguistic,
or tribal parties across countries as measures of the political salience
of these types of divisions across countries? What determines the size of the
coalition that an ethnic party is likely to mobilize? To assess this, we would
need to use the party as the unit of analysis and take the proportion of the
ethnic party’s target category as the dependent variable. Are we more likely
to see the ethnification of politics in new democracies? This would entail
treating EVOTE (or one of its substitutes) as the dependent variable and
using the age of the democracy (measured in the Przeworski et al. data set
[2000]) or the presence of founding elections as an independent variable. Is
548 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
there a link between colonial history and the degree of ethnic politicization?
This would entail regressing EVOTE (or its substitutes) on the range of variables
on colonial history being collected by Wilkinson and others.
EVOTE addresses several of the problems identified with the ELF index,
and the data that inform it, in the following way: First, it codes variables on
the basis of ethnic practice, consistently separating this from ethnic structure.
Second, it has a clear criterion for how to identify the relevant categories
and dimensions. In the realm of ethnic practice, the categories and
dimensions counted in CDEI are those that are named by the parties in
question. Thus, the CDEI count of explicitly activated ethnic categories in
India is based on all the categories explicitly mobilized by political parties
in the 1991 parliamentary election campaign (the national legislative election
closest to but before 1996). Third, it does not impose any assumptions
about mutual exclusiveness and exhaustiveness; rather, it allows us to
observe such completeness in the data.
In countries in which all individuals activate ethnic identities in their
voting behavior, the value of EVOTE would be 100%. In countries in which
only some individuals activate ethnic identities in their voting behavior, the
value of EVOTE would be less than 100%. In the Indian case, the value of
EVOTE is .39. This proportion is unaffected by whether the ethnic parties
in question mobilize mutually exclusive or overlapping categories. In India,
it so happens that the parties activate overlapping categories. But the value
of EVOTE would be the same even if the parties in question activated mutually
exclusive categories. Similarly, EVOTE also does not impose the
requirement that the categories activated by political parties be complete,
because it is the votes won by the parties that mobilize each category that
are added, not the proportion of the population made up by the categories
themselves.
How is EVOTE related to ELF? There appears to be no relationship.
Table 5 compares the data on some individual observations. As we see,
countries with similar levels of EVOTE (Canada, India, and Guatemala)
vary widely in their measures of ELF no matter which data set we use. And
India and Sri Lanka, with a comparable value on ELF (at least using the
Alesina et al. data [2003]) have different values on EVOTE.
Taking all observations together, the correlation between EVOTE and
the ELF index based on the Atlas Narodov Mira data (Miklukho-Maklai
Ethnological Institute, 1964) is only .07. The correlations between EVOTE
and the ELF index, calculated from the Fearon data (2003) and the Alesina
et al. data (2003) are similarly weak: respectively, .11 and .08. Figure 3
summarizes the relationship between the two measures. Figure 3 uses
Chandra, Wilkinson / Measuring the Effect of Ethnicity 549
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
the data on ethnic groups provided by Fearon (2003), for which there is the
greatest overlap in the number of observations, but the pattern remains the
same using the ELF indices based on the ELF measures using the other two
sources of data.
550 Comparative Political Studies
Figure 3
Scatter Plot of the Index of Ethnolinguistic
Fractionalization (ELF) and EVOTE
Table 5
A Comparison Between EVOTE and Fractionalization Indices
Year of Election
EVOTE for EVOTE ELFa
ELFb
ELFc
CULDIFd
Canada 34.4 1993 .75 .71 .60 .50
Guatemala 41.9 1995 .64 .51 .49 .49
India 38.9 1991 .89 .42 .81 .67
Sri Lanka 6.9 1994 .47 .41 .43 .39
Note: ELF = ethnolinguistic fractionalization; CULDIF = cultural difference.
a. Index of ethnolinguistic fractionalization based on Atlas Narodov Mira (Miklukho-Maklai
Ethnological Institute, 1964).
b. Index of ethnolinguistic fractionalization based on Alesina, Devleeschauwer, Easterly,
Kurlat, and Wacziarg (2003).
c. Index of ethnolinguistic fractionalization based on Fearon (2003).
d. Index of cultural difference based on Fearon’s data (see his Web site: http://www.stanford
.edu/~jfearon/).
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
What Might We Learn From ECI and
EVOTE About the Effect of Ethnicity?
To illustrate, we revisit the argument made by Fearon and Laitin (2003)
about the effect of ethnicity on the onset of civil war. The authors find that
“it appears not to be true that a greater degree of ethnic or religious diversity—
or indeed any particular cultural demography—by itself makes a country
more prone to civil war” (p. 75). The principal piece of evidence supporting
this claim is an empirical model demonstrating that there is no statistically
significant relationship between the ELF index, calculated using
Fearon’s (2003) data on ethnic groups, and the likelihood of civil war onset.
They find a similar absence of association when they use three other measures
for the same concept: the size of the largest ethnic group (using the
same data), the number of distinct languages spoken by groups exceeding
1% of the country’s population (using data from the Ethnologue database),
and a measure of religious fractionalization (using data from the same
sources used to construct the data set on ethnic groups; e.g., CIA World
Factbook). We conduct a comparison here between what we learn from
adding ECI or EVOTE to the core model in Fearon and Laitin’s work for
the limited observations for which we have data. We reproduce the database
of Laitin and Fearon not because we differ with their central findings in the
role of capacity in explaining civil wars but because they make a clear argument
on the basis of tests using what we would call ethnic structure variables.
More pragmatically, we also choose their 2003 article because they
have been especially good at making their data easily available for others to
replicate.16
The general points we make here would apply to any other
cross-national empirical study using the ELF index.
In Table 6, we report the results of this preliminary comparison. Column
1 estimates the broad empirical results in Fearon and Laitin 2003 (Table 1,
Model 1).17
With annual data on 160 countries between 1945 and 1999, the
Fearon and Laitin model is based on a total of 6,327 country–year observations.
Columns 2 and 3 compare this core model with a model that adds
the ECI variable. Column 2 replicates the model in column 1 for the limited
observations for which we have data on ECI. Because Wilkinson has
data on ECI for less than 40 countries, this yields a truncated data set of
1,265 country–year observations over the same period. Column 3 adds ECI
to the model in column 2. Column 4 replicates the model in column 1 for
the 846 observations for which it is feasible to run the model using EVOTE.
Column 5 adds EVOTE to the model in column 4. The observations that we
use in these two columns include the 91 countries for which there was an
Chandra, Wilkinson / Measuring the Effect of Ethnicity 551
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
overlap between countries in the Fearon and Laitin data set and Chandra’s
data set for the years 1991 to 1999. Therefore, the sample is truncated by
time in addition to space. We choose this period because EVOTE, unlike
ECI and ELF, is a time-sensitive measure. It reports the total vote share won
by ethnic parties in legislative elections between 1991 and 1996 in countries
that held competitive elections. This may well change in different elections.
552 Comparative Political Studies
Table 6
New Measures of Ethnic Practice and Civil War Onset
Original Modela
Adding ECI Adding EVOTE
1 2 3 4 5
Prior war –0.961 –1.715 –2.143 –2.134 –2.330
(0.315)***
(0.635)***
(0.646)***
(1.151)*
(1.195)*
Per capita income –0.346 –0.420 –0.544 –0.403 –0.441
(0.072)***
(0.156)***
(0.206)***
(0.211)*
(0.226)*
Log (population) 0.265 0.250 0.385 0.315 0.239
(0.073)***
(0.173) (0.184)**
(0.246) (0.270)
Log (% mountains) 0.213 0.496 0.056 0.270 0.349
(0.084)**
(0.203)**
(0.229) (0.261) (0.283)
Noncontiguous state 0.438 –0.169 –0.549 1.218 1.020
(0.275) (0.550) (0.650) (1.032) (1.103)
Oil exporter 0.829 1.330 1.098 1.771 2.263
(0.275)***
(0.515)***
(0.528)**
(0.985) *
(1.100)**
New state 1.726 –0.054 –0.190 3.107 3.336
(0.338)***
(0.821) (0.846) (0.937)***
(1.003)***
Instability 0.616 –0.081 –0.071 0.247 0.363
(0.235)***
(0.466) (0.474) (0.722) (0.720)
Democracy 0.020 0.045 0.054 0.054 0.062
(0.017) (0.034) (0.036) (0.066) (0.067)
ELF 0.219 –0.386 –2.097 1.554 1.765
(0.363) (0.747) (0.987)**
(1.445) (1.501)
ECI 2.091
(0.680)***
EVOTE 0.032
(0.016)**
Constant –6.647 –6.137 –7.138 –8.025 –8.068
(0.722)***
(1.675)***
(1.809)***
(2.420)***
(2.534)***
Observations 6,327 1,265 1,265 846 846
Note: Standard errors in parentheses. ELF = ethnolinguistic fractionalization; ECI = ethnic
concentration index (ethnic imbalance).
a. Fearon & Laitin’s Model 1, excluding fractionalization variable.
*
p < .10. **
p < .05. ***
p < .01.
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
We restrict the sample to the years 1991-1999 on the reasoning that EVOTE
can at best be seen as a measure of the political activation of ethnic identities
in the years proximate to this election—but it would be a stretch to
think of EVOTE as a measure of the political activation of ethnic identities
in more distant periods.
As we see in the original Fearon and Laitin model (2003) reported in column
1 of Table 6, the ELF index is not significantly associated with the likelihood
of onset of civil war in the complete data set. It continues not to be
significant when we run a regression (column 2) using only the limited sample
(1,265 observations) for which the ECI variable is available. But in column
3 we find that when we include the ECI variable in the regression, it is
positive and significant in explaining civil war onset (regardless of whether
we include ELF, as we do here, or exclude it). This association is not just the
result of the much smaller number of observations, because ELF continues
not to be significant for the same set of observations. Note too that in column
3 the sign on ELF changes direction and becomes significant. Similarly,
in column 5, we find that when we include the EVOTE variable in the
regression, it is positive and significant in explaining civil war onset (regardless
of whether we include ELF, as we do here, or exclude it). Because the
ELF index is not significant for the same set of observations in column 4, we
know that this finding is not just the result of the much smaller number of
observations.
Interpreting the Association
We cannot say whether these results will hold up with a more complete
data set. But taking them as provisional data, we can indeed identify more
about what we have learned and what we have not from these measures
than we can from the ELF index. We discuss below a substantive interpretation
of the effect associated with all three measures.
Fearon and Laitin (2003) interpret the lack of a statistical association
between the ELF index and the likelihood of the onset of civil war as evidence
that what they term ethnic demography (and what we term ethnic
structure) does not have a direct effect on civil war. This is an especially
careful interpretation in several ways. First, Fearon and Laitin link their
measure with a concept: They note explicitly that the ELF index and the
concept of ethnic diversity refer to ethnic structure and not ethnic practice.
Second, it is based on several measures of ethnic diversity. Third, their
interpretation of the lack of association between these measures and the
onset of civil war is nuanced. They take this to mean only that there is no
Chandra, Wilkinson / Measuring the Effect of Ethnicity 553
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
direct relationship between ethnic structure and the onset of civil war, noting
that there may well be an indirect relationship.
But given the lack of a conceptual justification guiding the design of the
data sets and measures, even this careful interpretation is difficult to support.
Take the effect associated with the ELF index. As we have argued in the body
of this article, we simply do not know whether the data used to compute the
ELF index reflect ethnic practice or ethnic structure. Similarly, we do not
know what the index captures. The second measure—the size of the largest
ethnic group—is more justifiable because it does not employ the assumptions
that drive the ELF index. But here too, we have no way of knowing
whether the data used in constructing this measure describe the largest
group based on ethnic structure or ethnic practice. Third, take the measure
of ethnic diversity based on linguistic distance. Without a discussion on
how the boundaries between language groups are drawn, we do not know
if these data are reliable measures for linguistic structure or linguistic practice
for the countries in the data set. But ethnicity, as we note earlier, is not
defined simply by language. Thus, even the most justifiable data set on linguistic
structure cannot tell us what the effect of ethnic diversity is, broadly
defined. The same points apply to the data and measures on religious fractionalization.
Because we do not know what these data and measures mean,
we also do not know how we might improve on them to generate more precise
conclusions.
Consider now what we learn and what we do not learn from the use of ECI
and EVOTE. ECI and EVOTE capture particular aspects of ethnic practice—
the representation of ethnic groups in nonelectoral politics, in the case of the
ethnic imbalance measure, and the explicit mobilization of voters by political
parties, in the case of EVOTE. Furthermore, they capture ethnic practice at
different points in time. ECI is a measure of ethnic practice at the time of
independence from a colonial power. The imbalance reflected in the ECI
measure may or may not be reflected in the period in which a civil war breaks
out. As such, it is a measure of the long-term effects of historical ethnic practice.
By contrast, EVOTE is a measure of political mobilization around ethnic
identities in the period proximate to a civil war. In the model above, the
election for which EVOTE is coded occurs just before, just after, or in the
same year as the year in which a civil war broke out.
We can take these results to suggest that ethnic practice has some effect
on the likelihood of violence. The precise effect remains to be elaborated,
but we can map out an agenda for ascertaining the nature of this effect on
the basis of how we conceptualize and measure these variables. Take ECI
first: ECI can be taken as one measure of interethnic disparity presumed to
554 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
link ethnic diversity to the likelihood of violence. Suppose we continue to
find an association between ECI and the onset of civil war when we have a
more complete data set. We know that ECI is a measure of historical practice.
To identify whether it does indeed increase the likelihood of civil war,
we would have to theorize about and measure the link between historical
interethnic imbalances and the actual outbreak of civil war. Might the
shadow of colonial inequities affect the likelihood of civil war regardless of
postindependence events? One way to test this hypothesis would be direct,
using data on postindependence ethnic imbalances; but given the difficulty
of collecting data on these, we could also simply reduce the magnitude of
the ECI indicator as we move further away from the year of independence.
Furthermore, we should find a link between ECI in some period and the
activation of ethnic identities in the period proximate to violence. This
could be EVOTE (to the extent that institutionalized activation increases the
likelihood of violence) or a measure of noninstitutionalized ethnic activation
(e.g., protests, riots, strikes). If we did not find a link between ECI and
a measure of ethnic mobilization, we would have to identify a different
chain of mechanisms leading to civil war onset.
Consider now what we learn from the association between EVOTE and
the likelihood of civil war. The positive association between EVOTE and
the onset of civil war for the cases for which we have data tells us that in
countries with competitive elections, the mobilization of ethnic identities
by political parties and the onset of civil war are positively related. But we
cannot say at this point that one causes the other. In the model above, the
election for which EVOTE is coded occurs just before, just after, or in the
same year as the year in which a civil war broke out. We do not know if
EVOTE was similarly strong in these countries in elections preceding civil
war. Therefore, we cannot say that the activation of ethnic identities by
political parties caused the civil war—it may well have been its product.
Alternatively, and more likely, the explicit activation of ethnic identities by
political parties and civil war (or lesser forms of violence) may well coevolve.
A proper test would require time-series data on EVOTE on a broader
set of countries (and not just those that hold highly competitive elections),
which we are in the process of collecting.
Just as important, we can also clearly identify what these results do not
suggest. They do not tell us anything about the relationship between ethnic
structure and the likelihood of civil war. At this point, we cannot tell
whether ethnic practice as measured by ECI or EVOTE is a product of ethnic
structure or independent of it. Thus, we cannot tell if ECI and EVOTE
have an effect on the likelihood of civil war or whether they are simply
Chandra, Wilkinson / Measuring the Effect of Ethnicity 555
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
intervening variables in a causal chain triggered by the presence of a particular
form of ethnic structure. All that we do know is this: There is a weak
bivariate association between ELF and ECI and between ELF and EVOTE
(reported in the preceding section) and no statistically significant association
between ELF and the likelihood of civil war. If we assume that ELF
captures the concept of ethnic structure, we can interpret this to suggest that
ethnic structure has a weak relation to ethnic practice. This weak association
might lead us to conclude that ethnic practice affects the likelihood of
civil war independently of ethnic structure. But what we know about ELF
hardly warrants such an informative conclusion. A more reliable conclusion
about the relationship between ethnic structure, ethnic practice, and the
likelihood of civil war requires the collection of new data and more appropriate
ways of measuring ethnic structure than the ELF index.
Even without a precise measure of ethnic structure, we suspect that, by
looking closely at EVOTE and ECI in individual cases, it may not determine
ethnic practice. Take the case of India. Although we do not have a
measure, we might guess that there are at least 5 times as many ethnic identities
embedded in India’s ethnic structure as were actually explicitly activated
by political parties in politics. We could also reasonably say that the
ethnic structures of Canada, Guatemala, and India are distinct. The number
and type of ethnic identity dimensions relevant in each country and the categories
arrayed on each appear, at least at first glance, to be different. Yet,
each country had an equivalent percentage of voters voting for ethnic parties
our data: 35% in Canada, 39% in India, and 41% in Guatemala.
Conversely, the ethnic structures of India and Sri Lanka are similar: Both
countries, although of very different sizes, are divided by language, religion,
region, and caste. But the degree to which explicit ethnic rhetoric is
used in both countries—and the proportion of voters who are voting for ethnic
parties in both countries—is quite distinct. Only 7% of voters voted for
explicit ethnic parties in Sri Lanka, compared to 39% in India.
The same can be said of the gap between ethnic structure and ethnic
imbalance. Take Pakistan and Burundi, countries of dramatically different
sizes and structures. Pakistan’s ethnic divisions are based on tribe, region,
language, religion, and caste, with no majority group on any except the
dimension of religion. Burundi, by contrast, appears to be a simpler ethnic
structure, with one principle division based on tribe (for want of a better
word), with a dominant majority and two smaller minorities. Yet both had
very high ECIs at independence: 1.3 and 1.0, respectively.
556 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Evaluating and Improving the Measures
Neither ECI nor EVOTE is free of bias and measurement error. In some
cases, neither has been able to escape some of the problems that characterize
the ELF index and previous data. In other cases, there are new biases
and errors. But to the extent that these measures and the concepts that they
capture are transparent, they help us in deciding how to assess the reliability
of our findings and how to improve them.
Take the case first of ECI, which relies on census classifications for its
count of ethnic categories. Our previous data sets also rely heavily on
national censuses. But ethnic identity is, of course, a dependent variable as
well as an independent variable, so the census data that underlie this indicator
will often themselves be highly politicized. These data are likely to be
particularly politicized and unreliable in high-conflict states where (a) governments
seek to make groups appear larger or smaller than they are or to
ignore some identities (e.g., religious data are not collected in the French
census, linguistic data are not collected in the Belgian census); (b) individuals
within the groups strategically declare themselves to be members of
one group rather than another; and (c) conflict and violence can rapidly
change the demographic composition of a population because of migration
and deaths, as well as government preferences and individual preferences
on group identity. Thus, we should not use ECI as an independent variable
to predict the degree of interethnic conflict in a society without thinking
about how to address this problem of endogeneity.
EVOTE has other biases and errors. The data reported above are characterized
by a selection bias: As we note, they include only those countries
that held competitive elections in the 1990s—based on Keefer’s index of
legislative competitiveness (2005). Autocracies and democracies in which
elections were not judged as being highly competitive by the Keefer index
are only now being added to this initial data set. But even when expanded
over space and time, an important selection bias will remain. EVOTE can
be collected only for countries that have elections (a measure of the vote
would hardly be sensible for countries that do not have elections). The
exclusion of countries that do not hold elections in each period for which
we have data will have to be taken into account in an interpretation of association
between EVOTE and the likelihood of civil war (and several other
dependent variables). Furthermore, there is measurement error in the data
that may well be a source of bias. One major source of measurement error
comes from the quality of our sources. This error can be a source of bias for
some purposes. For instance, we have better information for some regions
Chandra, Wilkinson / Measuring the Effect of Ethnicity 557
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
(Europe) than we do others (Africa), for larger parties than smaller parties,
and for incumbent parties than opposition parties. Suppose now the activation
of large numbers of small ethnic categories is less likely to produce
violence than the activation of a small number of large categories. Suppose
also that small parties are disproportionately likely to be associated with
small ethnic categories. In this case, the true effect of EVOTE may be distorted
by the systematic undercounting/error associated with small parties.
Several other instances of bias and error are documented in the coding
protocol accompanying this variable. But the strength of the new measures
that we have proposed here may lie precisely in the transparency of their
weaknesses. Our principal critique of previous work has been not that the
conclusions generated from that work are wrong but that we do not know
what they mean, how we can assess their degree of reliability, and how we
can improve on them. The advance in knowledge that we associate with the
new data and measures that we have proposed—and several of the more
recently developed alternatives—is not that they generate correct or even
different conclusions but that we can interpret conclusions generated by
them, estimate the level of uncertainty and the degree of bias in those conclusions,
and formulate strategies for how to improve on them without having
to start from a new foundation. This, ultimately, is what we need to
progress in a cumulative way rather than go around in circles.
Notes
1. See, for instance, Easterly, Alesina, and Baqir (1999); Easterly and Levine (1997);
Fearon and Laitin (2003); Miguel (2004); Miguel and Gugerty (2005); Neto and Cox (1997);
Ordeshook and Shvetsova (1994); Posner (2004).
2. Horowitz (1985), for instance, counts Hindus and Muslims in India, Christians and
Muslims in Lebanon, and Creoles and Indians in Guyana and Trinidad as ethnic categories
even though they do not possess his primary defining characteristic of a myth of common
ancestry. Fearon (2003) counts Hindi speakers in India, Blacks in the United States, and
Mestizos in Nicaragua as ethnic groups, even though those classified as Hindi speakers in
India do not meet his definitional criterion of having a distinct history as a group or a shared
culture valued by the majority of members, even though those classified as Blacks in the
United States do not often share the definitional criterion of having a shared culture, and even
though many of those classified as Mestizos in Nicaragua do not share the definitional criterion
of the descent rule. Chandra’s previous work (2004) counted categories based on region
as ethnic, even when it was not clear whether these groups met her definition of ethnic groups
as ascriptive groups.
3. See Horowitz (1985) for a comparable but not identical list. The modifications of this
list (e.g., the inclusion of identities based on dialect) are based on examining his actual classification
of identities as ethnic.
558 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
4. Although several of the authors whom we discuss here use the word group, the entities
that they call groups are simply categories. We cannot avoid the use of the word group here,
because so many of those whose work we discuss use it. But we suggest interpreting the word
group throughout as category because that would be the more accurate understanding of an
author’s usage.
5. In political science, one of the first books to raise the question of how we move from
many ethnic categories to only a few salient political identities was Brass’s Language,
Religion and Politics in North India (1974). In it, Brass highlighted the role of political competition,
institutional rules, and intrinsic aspects of the ethnic identities themselves in explaining
why out of many thousands of ethnic groups in India, only a few dozen identities had
become politically salient.
6. See Fearon (2003) and Posner (2004).
7. In addition to being cited to the aforementioned works, this statement is based on a representative
sample of research to review. Using the social science citation index, we surveyed
all articles with ethn in the title published since 1990 in the following journals: American
Journal of Political Science, American Political Science Review, Comparative Political
Studies, Comparative Politics, International Organization, Journal of Conflict Resolution,
Journal of Economic Growth, Quarterly Journal of Economics, and World Politics.
8. We put to one side for the moment a topic that we return to at the end of this article:
the basic reliability of the census statistics from which we typically derive these indicators. For
instance, Fearon’s total (2003) for Hindi speakers (.39) in constructing his index is presumably
higher than the Atlas Narodov Mira’s (.25; Miklukho-Maklai Ethnological Institute,
1964) because of a decision that the Indian government and census bureau took in the early
1970s to aggregate 48 categorical answers to the question “What language do you speak?”
under the single heading of Hindi in the published census tables. This includes large language
groups such as Bhojpuri (23 million speakers in 1971) and Chattisgarhi (10.6 million) that
have generated quite substantial political movements at various times (personal communication
to Wilkinson from former Indian census commissioner M. Viyayanunni, March 27, 2003).
For an excellent introduction to the many state, individual, and socioeconomic incentives that
can affect the quality and reliability of census statistics, see Kertzer and Arel (2002).
9. For instance, N*
is a measure that does not require the assumption of exhaustiveness
that the index of ethnolinguistic fractionalization does; thus, it can be used in instances where
the activated ethnic categories in a polity do not sum to 100%.
10. Ness (1967) uses percentages so that his ethnic concentration index runs from 0 to 200
rather than 0.0 to 2.0, as it is here in the form of ethnic imbalance.
11. Cruise O’Brien (1980) is a rare exception, and her study of Kenya clearly measures
the extent of the economic inequalities among that country’s ethnic groups, inequalities that
were important in driving conflicts in the 1960s and 1970s. She shows that two thirds of the
businesses with more than 50 employees in 1961 were under the control of Kenya’s 100,000strong
Asian minority, thereby creating a powerful incentive for postindependence leaders to
target this advanced community to solidify their support.
12. An example of the partial endogeneity of ethnic identity to group categorizations used
by the state in military recruitment comes from research on colonial Nyasaland (Malawi) by
Parsons (1999). The use of ethnic criteria as an implicit indicator of recruits’ reliability by
colonial recruiters encouraged some strategic switching on the part of those who identified
themselves as members of Group X to be eligible for recruitment in the colonial army. In
Nyasaland, members of the nonmartial Lomwe identified themselves as Yao to enlist in the
King’s African Rifles. Colonial officers also categorized those who might have claimed several
Chandra, Wilkinson / Measuring the Effect of Ethnicity 559
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
identities in line with their ethnic stereotypes according to which groups were martial and
which not so that “if a Southerner [in Nyasaland] refused to enlist or was a bad soldier, he was
a Mpotola; if he did enlist and served with distinction, he was a Nguru or even a Yao” (p. 91).
13. See Table 2.1 in Parsons (2003).
14. These figures have been generated by an analysis of the ethnic composition of each
regiment in the unified Indian army that was divided between India and Pakistan in
August–September 1947. The basic data on the religious balance of the army are extrapolated
from the table titled “Class Composition of Indian Infantry 1947” in Appendix 7 of Tuker
(1950), supplemented with an analysis of regimental histories—where these are not clear from
the name of the regiment—that establishes the religious and provincial origin of recruits.
15. This criterion is somewhat easier to operationalize than the definition by Horowitz
(1985), who codes ethnic parties according to the distribution of their supporters, as well as by
whether they serve the interests of an ethnic group or a cluster of groups.
16. The Fearon and Laitin data set used in their 2003 article is available at Fearon’s Web
site: http://www.stanford.edu/~jfearon/.
17. Our results are slightly different because we do not include (because it was not
included in the particular replication dataset we downloaded) the variable “religious fractionalization,”
which was however not significant at the .05 level in any of Fearon & Laitin’s models.
References
Alesina, Alberto, Devleeschauwer, Arnaud, Easterly, William, Kurlat, Sergio, & Wacziarg,
Romain. (2003). Fractionalization. Journal of Economic Growth, 8(2), 155-194.
Bedlington, Stanley S. (1981). Ethnicity and the armed forces in Singapore. In Dewitt C.
Ellinwood & Cynthia C. Enloe (Eds.), Ethnicity and the military in Asia (pp. 127-139).
Buffalo: State University of New York.
Bossert, Walter, d’Ambrosio, Conchita, & La Ferrara, Eliana. (2006, July). A generalized
index of fractionalization (Working Paper No. 313). Milano, Italy: Innocenzo Gasparini
Institute for Economic Research. Retrieved December 19, 2007, from http://www.igier
.uni-bocconi.it/folder.php?vedi=3516&tbn=albero&id_folder=182
Brass, Paul. (1974). Language, religion and politics in North India. Cambridge, UK:
Cambridge University Press.
Brubaker, Rogers. (2004). Ethnicity without groups. Cambridge, MA: Harvard University Press.
Cederman, Lars-Erik, & Girardin, Luc. (2007). Beyond fractionalization: Mapping ethnicity
onto nationalist insurgencies. American Political Science Review, 101(1), 173-186.
Chandra, Kanchan. (2001). Cumulative findings in the study of ethnic politics. APSA-CP,
12(1), 7-11.
Chandra, Kanchan. (2004). Why ethnic parties succeed. NewYork: Cambridge University Press.
Chandra, Kanchan. (2006). What is ethnic identity and does it matter? Annual Review of
Political Science, 9, 397-424.
Chandra, Kanchan. (2007). Part I: A conceptual framework for thinking about ethnic identity.
In Ethnicity, Politics, and Economics. Manuscript in preparation.
Chandra, Kanchan. (in press). A constructivist dataset on ethnicity and institutions. In Rawi
Abdelal, Yoshiko Herrera, Ian Johnston, & Rose McDermott (Eds.), Identity as a variable.
Cambridge, UK: Cambridge University Press.
Chandra, Kanchan, & Boulet, Cilanne. (2006). A language for thinking about ethnic identity
change. In Ethnicity, Politics, and Economics. Manuscript in preparation.
560 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Chhibber, Pradeep, & Kollman, Ken. (1998). Party aggregation and the number of parties in
India and the United States. American Political Science Review, 92, 329-342.
Collier, Paul, & Hoeffler, Anke. (2001). Greed and grievance in Civil War (Report No. 28126).
Washington, DC: World Bank. Retrieved December 5, 2007, from http://worldbank.org/
programs/library
Cruise O’Brien, Rita. (1980). Factors of dependence: Senegal and Kenya. In W. H. Morris-Jones
(Ed.), Decolonisation and after: The British and French experience (pp. 283-309). London:
Frank Cass.
Decalo, Samuel. (1990). Coups and army rule in Africa: Motivations and constraints. New Haven,
CT: Yale University Press.
Easterly, William, Alesina, Alberto, & Baqir, Reza. (1999). Public goods and ethnic divisions.
Quarterly Journal of Economics, 114, 1243-1284.
Easterly, William, & Levine, Ross. (1997). Africa’s growth tragedy: Policies and ethnic divisions.
Quarterly Journal of Economics, 112, 1203-1250.
Fearon, J. (2003). Ethnic structure and cultural diversity by country. Journal of Economic
Growth, 8(2), 195-222.
Fearon, James D., & Laitin, David D. (2000, October). Ordinary language and external validity:
Specifying concepts in the study of ethnicity. Paper presented at the meeting of the
Laboratory in Comparative Ethnic Processes, University of Pennsylvania, Philadelphia.
Fearon, James D., & Laitin, David D. (2003). Ethnicity, insurgency and civil war. American
Political Science Review, 97, 75-90.
Gutmann, Amy. (2003). Identity in democracy. Princeton, NJ: Princeton University Press.
Gutteridge, William. (1962). Armed forces in new states. New York: Oxford University Press.
Hegre, Håvard, & Sambanis, Nicholas. (2001). Sensitivity analysis of empirical results on civil
war onset. Journal of Conflict Resolution, 50, 508-535.
Horowitz, Donald L. (1980). Coup theories and officers motives: Sri Lanka in comparative
perspective. Princeton, NJ: Princeton University Press.
Horowitz, Donald L. (1985). Ethnic groups in conflict. Berkeley: University of California Press.
Hutchinson, John, & Smith, Anthony D. (1996). Ethnicity. Oxford, UK: Oxford University Press.
Johnson, Douglas H. (Ed.). (1992). British documents on the end of empire: Sudan (Vol. 2).
London: Her Majesty’s Stationery Office.
Kasfir, Nelson. (1976). The shrinking political arena: Participation and ethnicity in African
politics. Berkeley: University of California Press.
Kasfir, Nelson. (1979). Explaining ethnic political participation. World Politics, 31, 365-388.
Keefer, Philip. (2005). DPI 2004: Database of Political Institutions: Changes and variable
definitions. Washington, DC: World Bank.
Kertzer, David I., & Arel, Dominique. (Eds.). (2002). Census and identity: The politics of race,
ethnicity, and language in national censuses. New York: Cambridge University Press.
Kymlicka,Will. (Ed.). (1995). The rights of minority cultures. Oxford, UK: Oxford University Press.
Laitin, David, & Posner, Daniel. (2001). The implications of constructivism for constructing
ethnic fractionalization indices. APSA-CP, 12(1), 13-17.
Little, Peter D. (1998). Maasai identity on the periphery. American Anthropologist, 100(2), 444-457.
Malcolm X. (1964). The autobiography of Malcolm X. New York: Ballantine Books.
Mendelberg, Tali. (2001). The race card: Campaign strategy, implicit messages, and the norm
of equality. Princeton, NJ: Princeton University Press.
Miguel, Edward. (2004). Tribe or nation? Nation building and public goods in Kenya versus
Tanzania. World Politics, 56, 327-362.
Chandra, Wilkinson / Measuring the Effect of Ethnicity 561
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Miguel, Edward, & Gugerty, Mary Kay. (2005). Ethnic diversity, social sanctions and public
goods in Kenya. Journal of Public Economics, 89, 325-368.
Miklukho-Maklai Ethnological Institute. (1964). Atlas narodov mira. Moscow: Author.
Miners, Norman J. (1971). The Nigerian army, 1956–1966. London: Methuen.
Mozaffar, Shaheen. (2007). Assessing The Political Effects Of Ethnopolitical Group
Fragmentation. Prepared for presentation at the annual meetings of the American Political
Science Association August 30-September 2, 2007, Chicago, Illinois
Mozaffar, Shaheen, Scarritt, James R., & Galaich, Glen. (2003). Electoral institutions, ethnopolitical
cleavages, and party systems in Africa’s emerging democracies.” American
Political Science Review, 97, 379-390.
Ness, Gayl. (1967). Bureaucracy and rural development in Malaysia. Berkeley: University of
California Press.
Neto, Octavio Amorim, & Cox, Gary W. (1997). Electoral institutions, cleavage structures and
the number of parties. American Journal of Political Science, 41, 149-174.
Ordeshook, Peter C., & Shvetsova, Olga V. (1994). Ethnic heterogeneity, district magnitude
and the number of parties. American Journal of Political Science, 38, 100-123.
Parsons, Timothy H. (1999). The African rank-and-file: Social implications of colonial military
service in the King’s African Rifles, 1902–1964. Portsmouth, NH: Heinemann.
Parsons, Timothy H. (2003). The 1964 army mutinies and the making of modern East Africa.
Westport, CT: Praeger.
Posner, Daniel. (2005). The institutional origins of ethnic politics in Africa. Cambridge, UK:
Cambridge University Press.
Posner, Daniel N. (2004). Measuring ethnic fractionalization in Africa. American Journal of
Political Science, 48, 849-863.
Przeworski, Adam, Alvarez, Michael E., Cheibub, José Antonio, & Limongi, Fernando.
(2000). Democracy and development: Political institutions and well-being in the world,
1950–1990. New York: Cambridge University Press.
Rajasingham-Senanayake, Darini. (1999). Democracy and the problem of representation: The
making of bi-polar ethnic identity in post/colonial Sri Lanka. In Joanna Pffaff-Czarnecka
(Ed.), Ethnic futures (pp. 99-134). New Delhi, India: Sage.
Reynal-Querol, Marta. (2002). Ethnicity, political systems and civil wars. Journal of Conflict
Resolution, 46(1), 29-54.
Rizvi, Hasan-Askari. (2000). Military, state and society in Pakistan. New York: St. Martin’s.
Scarritt, James R., & Mozaffar, Shaheen. (1999). The specification of ethnic cleavages and
ethnopolitical groups for the analysis of democratic competition in Africa. Nationalism
and Ethnic Politics, 5, 82-117.
Shepherd, George W., Jr. (1966). National integration and the Southern Sudan. Journal of
Modern African Studies, 4, 193-212.
Tambiah, Stanley J. (1986). Sri Lanka: Ethnic fratricide and the dismantling of democracy.
Chicago: University of Chicago Press.
Tarbush, Mohammed. (1982). The role of the military in politics: A case study of Iraq to 1941.
London: Kegan Paul.
Tuker, Sir Francis. (1950). While memory serves. London: Cassell.
Waters, Mary. (1990). Ethnic options: Choosing identities in America. Berkeley: University of
California Press.
Weber, Max. (1978). Economy and Society. (Guenther Roth and Claus Wittich, Eds.).
Berkeley: University of California Press.
562 Comparative Political Studies
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from
Wilkinson, Steven I. (in press). Which group identities lead to most violence: Evidence from
India. In Ian Shapiro & Stathis Kalyvas (Eds.), Order, conflict and violence. Cambridge,
UK: Cambridge University Press.
Willame, Jean-Claude. (1972). Patrimonialism and political change in the Congo. Stanford,
CA: Stanford University Press.
Kanchan Chandra is an associate professor of political science at New York University. She
is the author of Why Ethnic Parties Succeed: Patronage and Ethnic Head Counts in India
(2004).
Steven Wilkinson is an associate professor of political science at the University of Chicago.
He is the author of Votes and Violence: Electoral Competition and Ethnic Riots in India (2004).
Chandra, Wilkinson / Measuring the Effect of Ethnicity 563
at Masarykova Univerzita on October 25, 2016cps.sagepub.comDownloaded from