Education Inequalities and Conflict Database

Transcription

1 Education Inequalities and Conflict Database Technical annex to the global study on horizontal inequalities in education and violent conflict FHI 360 Education Policy and Data Center April 2015 Washington, DC

2 United Nations Children s Fund Peacebuilding Education and Advocacy Programme Education Section, Programme Division Three United Nations Plaza New York, New York April 2015 This document was produced through a partnership between UNICEF and FHI 360, as part of UNICEF s Peacebuilding Education and Advocacy Programme (PBEA), Learning for Peace Initiative. This is a technical annex to the research report exploring the relationship between horizontal education inequality and violent conflict, commissioned by UNICEF Peacebuilding, Education and Advocacy Programme and completed by the FHI 360 Education Policy and Data Center research team: Carina Omoeva, Elizabeth Buckner, Charles Gale, and Rachel Hatch. Ania Chaluda developed the back projection module used in the construction of the EIC dataset. Research interns Elyse Sadeghi, Kemi Oyewole and Khaled Al-Abbadi provided invaluable assistance in the construction of both EIC and SEIC datasets. The team is grateful to Bosun Jang and Hiroyuki Hattori of UNICEF PBEA for their continuous guidance and support throughout this research project. NOTE: For inquiries and to obtain the datasets for research and replication contact FHI 360 Education Policy and Data Center at epdc@fhi360.org

3 Contents Overview... 1 Part I. Education Inequality and Conflict Dataset (EIC)... 2 Educational Attainment Data... 2 Data Sources... 3 Educational Attainment Variables... 4 Calculating Educational Attainment... 4 Age Groups... 7 Extraction Process... 7 Social Group Markers... 8 Gender... 8 Identity Groups... 8 Treatment of Missing Ethnic and Religious Data Calculating Group-Level Attainment Back Projections of Educational Attainment Data Interpolation Calculating Inequality Indices Minority Groups Educational Inequality Indices Conflict Measures Calculation of Conflict Indicators Part II: Subnational Educational Inequality and Conflict (SEIC) Dataset Educational Attainment Alignment of Subnational Regions Education Inequality Measure at Subnational Level Interpolation Conflict Measure: GED Wealth Index References Appendix A: Overview of Included Datasets Appendix B: MICS Datasets and Subnational Representation Appendix B: Codebook for the EHID, EIC and SEIC Group Educational Attainment Dataset Variable List Educational Inequality and Conflict Dataset (EIC) Subnational Educational Inequality and Conflict (SEIC)... 33

4 Overview The purpose of this document is to describe the methodology, data sources, and procedures used in the construction of the Education Inequalities and Conflict (EIC) and the Subnational Educational Inequality (SEIC) Datasets. More information can be obtained through direct inquiry: This dataset was compiled as part of the comprehensive research project carried out by FHI 360 Education Policy and Data Center (EPDC) within the framework of a consultancy for UNICEF Peacebuilding in Education and Advocacy Programme (PBEA). In line with the focus of the research project, the database contains measures of horizontal inequality in education, across ethnic, religious, and subnational divisions. In addition, all measures are disaggregated by gender. The focus on ethnic, religious, subnational, and gender group identity markers is driven by the conceptual and empirical foundations of horizontal inequality theory. The theory states that inequality among groups defined by strong social identities generates group grievances and has a higher potential of translating into violent conflict than inequality among individuals (see Literature Review completed by EPDC as part of this project for more information). The database consists of two main datasets: 1. Educational Inequality and Conflict (EIC) dataset, which contains measures of inequality in educational attainment at the country level, disaggregated by level of education, gender, and type of group identity (i.e. inequality across ethnic, religious, and subnational groups), and data on conflict onset, type of conflict, and conflict incidence. The EIC dataset is constructed on the basis of the Group-level Educational Attainment Dataset (GEA), which is not used in our analysis but provides the source data on educational attainment at the group level that was used to calculate indicators of inequality. 2. Subnational Education Inequality and Conflict (SEIC) dataset, which contains data on educational attainment by subnational unit, along with inequality measures aggregated by type of inequality, gender, and level of education. This dataset also contains a group disadvantage measure, which compares its relative educational attainment with the national mean (see above). This dataset contains subnationally disaggregated conflict data from UCDP Georeferenced Events Dataset (GED). The group-level data extracted for this study include data on educational attainment for 111 countries, including 84 religious group disaggregation, 73 with ethnic group disaggregation, and 109 countries with subnational disaggregation. The historical time span covered by the dataset is years ; however, time series vary across countries depending on data availability. On average, the time span for a given country is 43 years, with the minimum of 1 year (for some subnational disaggregations) and maximum of 52 years. Geographically, the dataset covers all regions of the world, and is unique in providing data on types of intrastate inequality at this breadth and historical coverage. For the Horizontal Inequality and Violent Conflict research project, we do not use grouplevel data, however we provide the data for replication and analysis purposes in the Group-level Educational Attainment (GEA) dataset. It is the largest and longest-spanning dataset created to date to measure educational attainment disaggregated by identity group. In addition, the GEA dataset provides disaggregation by gender, offering attainment levels for males, females, and both genders combined. 1

5 EIC: Using the disaggregated group-level educational attainment data, we constructed the EIC dataset, which includes education inequality measures by type of inequality (ethnic, religious or subnational), level of schooling, and gender, collapsed at the country-year level. Each measure is intended to capture the disparity in education, and the presence of several metrics allows for comparisons and validation across them. The inequality measures include the Group Gini (GGini), Group Theil (GTheil) index, group-level coefficient of variance (GCOV), and a modification of the Lineq ratio measure proposed by Cederman et al. (2011). The construction of these indices is described further in this document. The education inequality measures in the EIC were merged with conflict data at the country-year level from the UCDP Conflict Onset dataset. The conflict data include measures of conflict incidence, new conflict onset, and the intensity of conflict. The selection of conflict data sources and the process for merging education and conflict data is described in this document. This document only details on the education and conflict variables that were generated as part of the data compilation process; other variables merged in from the conflict dataset have been kept intact in the dataset, but are not described in-depth here. In such cases, refer to the original sources cited in the footnotes. SEIC: To explore the effects of disparity on conflict at the subnational level, we constructed the SEIC dataset, where the unit of analysis is the subnational region. The conflict measure in SEIC comes from the UCDG Georeferenced Event Dataset, and because of the limitations of the conflict variable the SEIC is limited to the time period of and the geographic scope is restricted to 24 sub- Saharan African countries. In SEIC, the measure of inequality is not an inequality index as in EIC, but the size of the disparity between that region s mean educational attainment and the mean educational attainment for the country. Part I. Education Inequality and Conflict Dataset (EIC) The main focus of this dataset is inequality in education among identity groups defined by ethnicity and/or religion and disaggregated by gender, as well as subnational inequality. This section describes the sources and the extraction and disaggregation process used in the development of education inequality measures. It includes the description of the source data on education, the group markers used for disaggregation, and the treatment of missing data. The overall coverage of education inequality data is shown in Table 1 below. Table 1 Inequality Measures Inequality Measures # Countries Educational Inequality 111 Ethnic Groups 73 Religious Groups 84 Subnational Regions 109 Educational Attainment Data The education inequality measures in this dataset are based on data on educational attainment, by level, disaggregated by the relevant group markers (see below). The choice in favor of educational attainment, as opposed to literacy or resource inputs was made due to the availability and coverage of educational attainment information in surveys and censuses, and the possibility of disaggregating attainment by subgroup. No other indicator currently used in education neither access, nor quality or resources can rival the coverage and depth of educational attainment information. 2

6 In addition to this basic practical reason, there is a theoretical basis for examining educational attainment. It provides a gauge of the level of human capital stored in a particular subpopulation, and hence is relevant for the study of social inequality as a predictor of conflict. Educational attainment was used throughout the literature on horizontal and vertical inequality in education (see Literature Review). Data Sources The three primary sources of data for the dataset come from individual-level records of educational attainment, demographic information and household assets, drawn from three datasets: the Integrated Public Use Microdata Series International (IPUMS International), Demographic Household Surveys (DHS), and the Multiple Indicator Cluster Survey (MICS). Because the focus of the research is on group-based inequalities, we limited the analysis to DHS and MICS datasets that have information available on individuals religious or ethnic group identity. Because IPUMS also offers large sample sizes, we also used IPUMS datasets that identify individuals subnational region. Table 2 - Summary of Data Sources Source Countries Datasets Year of Median N Sample type Extraction IPUMS ,126 Random, nationally representative samples of census micro-data DHS ,811 Nationally representative survey MICS ,918 Nationally representative survey IPUMS: The IPUMS International is a project housed at the University of Minnesota that collects, preserves and harmonizes data from censuses around the world. Its data are drawn directly from the National Statistical Offices and relevant government agencies. The IPUMS International data repository includes data from 74 countries, drawn from 238 censuses and 544 million personrecords. Because IPUMS samples are drawn from the micro-data of national censuses, it is conducive to drawing large samples, which can provide more accurate information on the educational attainment of smaller minority groups. We typically drew a sample of 150,000 observations, but increased the samples for extremely large nations when possible. Additionally, in many countries, IPUMS International contains census data spanning multiple decades, often back to the 1960s. DHS: The USAID-funded DHS program contains data from over 300 surveys in over 90 countries. The first DHS surveys were conducted in the 1980s, and the first phase ( ) produced smaller n datasets which were focused mainly on women of reproductive age. Since then, the countries and topics covered by DHS has grown to cover a broader range of issues. There have been a total of six waves of DHS surveys, conducted at roughly five year intervals. 1 The goal of the DHS is to collect and disseminate accurate, nationally representative data on fertility, family planning, maternal and child health, gender, HIV/AIDS, malaria, and nutrition. The survey also collects important educational and demographic information on all individuals living in a household, making it appropriate for our purposes. MICS: The third source of data we use comes from the Multiple Indicator Cluster Survey (MICS). The MICS is a UNICEF-aided project that aims to fill data gaps on the situation of children and women in 1 See About Us - Demographic and Health Survey (DHS). Available at: 3

7 the areas of health, education, child protection and HIV/AIDS. Since the mid-1990s, there have been five waves of MICS surveys: MICS1 (1995); MICS2 (2000); MICS3 ( ); MICS4 ( ); MICS5 ( ). After reviewing the possible MICS surveys, we limit our analysis to datasets from Wave III and Wave IV, which provide the most reliable estimates of ethnicity, religion and education. 2 MICS questionnaires are modular tools that are meant to be adapted to the needs of a country; this means that the specific questions asked vary, although wording of questions and implementation methods are largely standardized across nations. MICS surveys are typically carried out by government organizations, with the support and assistance of UNICEF and other partners. Our dataset benefits from the fact that DHS and MICS survey administrators work closely together to harmonize survey questions and modules and coordinate implementation, to ensure comparability across surveys and countries. Educational Attainment Variables We use two variables to measure an individual s level of formal education: highest level of education attained and total years of schooling completed. Our dataset only includes surveys for which an indicator for highest level of educational attainment was available or could be proxied through level attended and years completed at that level. We also include total years of schooling attended as an indicator of educational attainment; however, data on total years of schooling was not available in all IPUMS surveys. Our primary measures of education, therefore, are: 1) the mean years of schooling among year olds of a given ethnic or religious group, or living in a certain subnational region, and 2) the percentage of those individuals that has completed a given level of education. We standardize the existing measures of education from each survey into three categories of educational attainment: 1. Mean Years of Schooling: Mean Years of Schooling 2. Primary incomplete: No education or some primary (but less than primary complete) 3. Primary complete: Primary education complete with some secondary possible 4. Secondary complete: Secondary education complete and above Individuals having only some secondary education are classified as primary complete, but not secondary complete. We classify anyone having higher education or some post-secondary in the category secondary complete. Calculating Educational Attainment Mean Years: The key measure of educational attainment is the group s mean years of schooling, which is calculated as the average number of years of school among all individuals in each identity group. In some IPUMS datasets, information on years of schooling completed is not available. In these cases, we estimate mean years of schooling by assigning a certain number of years of schooling to proportions of the population in each school level (e.g., 50% of those with some primary incomplete are assigned 3 years, etc.) and then calculating the group s average. Although the precise value for mean years is not accurate, it is an estimate based on real data extracted from the data on attainment, and is therefore likely offers a reliable estimate of inter-group differences in mean years. Attainment by Level: Because we draw on various sources, the datasets included in our analysis differ from one another in terms of how they determine educational attainment. In IPUMS, questions 2 See About UNICEF Data and Analytics for more information in MICS: 4

8 are not necessarily standardized, as they originate directly from census questionnaires. In DHS and MICS, questions determining educational attainment typically follow a standard sequence, namely: What is the highest level of schooling attended? What is the highest grade completed at that level? To determine an individual s level of educational attainment, a decision must be made about the grades that form the cut-off points at which someone can be considered to have completed a given level of schooling. This can be problematic when durations of schooling have changed over time. Illustratively, among a group of countries studied in an analysis of out-of-school children data, 28% had considered changes to primary school duration over the period analyzed (EPDC, 2013). Consider a country that changes its official length of primary from 5 years to 6 years. Older respondents who completed primary under the former system (5 years) would be classified as incomplete primary, if 6 years of primary were considered the cut-off to have completed the level. The UNESCO Institute for Statistics (UIS) has a classification system for standardizing school level durations in order to make levels comparable across countries and over time, based on things like level of specialization in the curriculum. However this system can gloss over nationally-defined school level durations, as captured in UNESCO International Bureau of Education World Data on Education. Moreover, neither the ISCED nor nationally-defined school structures may capture the variation within countries in terms of how to code individuals who have attended non-formal or madrassa schooling, as examples. Some surveys and censuses have methods for coding variations such as these, while others do not. The way we define school durations has important implications for the degree to which measures such as educational attainment reliably capture the level of schooling in a population. In part for this reason we also use the continuous years of schooling variable from each source, when it is available. For the purposes of this analysis, we used internationally standardized variables when possible. When it was not possible, we determined individuals attainment by consulting information on the structure of the national education system. The following is a discussion of how school level durations were determined depending on each data source. Table 3 Overview of Educational Measures Source Sample Frame Educational attainment categories EIC/ SEIC categories IPUMS Census No schooling or less than primary completed Primary incomplete (individual) Primary completed Primary complete Secondary completed Secondary complete Higher education completed DHS Household No education Primary incomplete Primary incomplete Primary complete Primary complete Secondary incomplete Secondary complete Secondary complete Higher education MICS Household No education, non-formal schooling, pre-school Primary incomplete Primary incomplete Primary complete Primary complete Secondary incomplete Secondary complete Secondary complete Higher education, post-secondary 5

9 IPUMS: We use the internationally harmonized categories of educational attainment. It is important to note that in the vast majority of IPUMS datasets, the categorization follows a generalized (primary-lower secondary-upper secondary) schooling structure; however, this classification was not possible in all datasets given differences across national educational systems. 3 In fact, IPUMS states that its categorization does not necessarily reflect any particular country's definition of the various levels of schooling in terms of terminology or the number of years of schooling. Instead, the internationally harmonized variable is an attempt to merge -- into a single, roughly comparable variable -- samples that provide degrees, ones that provide actual years of schooling, and those that have some of both (IPUMS 2014). Therefore, individuals may be coded as primary complete regardless of whether or not they have completed primary schooling as defined in their respective school system. IPUMS categorizes educational attainment into four levels: no schooling or less than primary completed, primary completed, secondary completed, and higher education completed, which we standardize to three levels of primary incomplete, primary complete, and secondary complete. We also use the number of years of schooling a person completed to calculate mean years of schooling. DHS: For both DHS and MICS, the highest level of educational attainment is determined through the combination of two related questions the highest level of education attended and the last year successfully completed at that level. These two questions are then matched to the national educational system to determine whether respondents completed primary or secondary school in their nation. In the majority of DHS datasets, a single year determines the categories primary complete and secondary complete. However, particularly in older phases of DHS, a respondent may be considered to have completed primary at multiple years of schooling, if the official durations changed over time. The mutually exclusive categories used for educational attainment are no education, primary incomplete, primary complete, secondary incomplete, secondary complete and higher education. We recode this variable to align to our three levels of primary incomplete, primary complete, and secondary complete. We also use the number of years of schooling a person completed to calculate mean years of schooling. MICS: Unlike DHS and IPUMS, MICS does not provide standardized categories of educational attainment. We thus were left with the task of making a determination about school level durations for these countries and years in order to re-code educational attainment, taking into consideration types of schooling (non-formal, madrassa, etc.) and different school levels to code individuals in similar fashion to the DHS and IPUMS datasets. For each case, a determination was made based on information supplied by UIS/ISCED and UNESCO IBE, as well as MICS reports and level durations for other country-years. This is documented in the Technical Annex. We also use the continuous years of schooling variable where it is available. There are important differences between IPUMS, DHS, and MICS in terms of how individuals are classified into levels of educational attainment. IPUMS, in its international harmonization process, tends to align individuals to UNESCO s international norm of 6 years primary, 3 years lower secondary, 3 years upper secondary (6-3-3). In some cases, IPUMS classifies an individual as having completed primary schooling if that individual has 6 years of primary schooling, regardless of the actual length of primary in a given nation. This means that in some cases, individuals who are classified as having completed primary school will actually not be eligible for secondary school or received any of the benefits of primary school completion. Alternatively, countries with 4-years of primary are classified as Less than Primary in EDATTAN, regardless of the fact that these students are eligible for lower secondary in their nations. Like DHS, MICS takes a two-stage process to 3 See: Comparability General, for extensive explanation of EDATTAN recoding. Available at: Note that exceptions are made for countries where primary school is shorter than 6 years, including Bangladesh, Belarus, Colombia and Spain

10 determining educational attainment by asking both highest level attended and highest year completed at that level. This means that the data can be aligned to the educational system in the specific country. A number of countries in our dataset revised their educational systems at some point in the past, meaning that individuals from different age cohorts may have experienced varying lengths of primary or secondary education. Datasets are standardized to the duration of schooling levels at the time the survey data were collected. We choose not to use datasets from countries (e.g., Georgia) that changed the structure of their educational systems multiple times, making attainment in different cohorts too difficult to reliably estimate. Where we were working with original datasets (i.e. DHS and MICS), we chose not to change the duration of levels of school and allow for cross-national variability of durations (as opposed to standardizing to 6-3-3). We believe that keeping the national definitions of level durations is more meaningful in interpretation. Variability of durations across countries does not interfere with the validity of the inequality measures, because they capture inequality across groups or units within each country, regardless of whether or not it aligns with the definitions in other countries. Age Groups Although the focus of our analysis is on young people, aged 15-24, we take advantage of the data available to back project estimates of educational attainment in prior decades. To carry out the back projections, we group individuals into 10-year intervals (15-24, 25-34, etc), which allows us to examine the educational attainment of year olds at the time of the survey, and also back project attainment levels for older cohorts. Individuals younger than 15 were excluded from the analysis because they are not the focus of the analysis. In almost every country examined, we have the individual s exact age at the time of the survey however, in a few countries, individuals are grouped into five year intervals by age. The specific samples that provide ages grouped into intervals are: Ireland, Israel, Italy, Netherlands, State of Palestine, and Slovenia. In these cases, the data are coded to the mid-points of the intervals (rounding down). 4 Extraction Process Our data comes from both census-level information and population surveys, which means that the calculation of population estimates differs by dataset. In general, we use the sample weights suggested by survey administrators in all extraction estimates. Because IPUMS data comes from micro-level census data, the majority of datasets from IPUMS are self-weighted at the individual level. Nonetheless, there are some exceptions to this rule, and we follow the sample weighting suggested by IPUMS in their sample description information. 5 DHS and MICS are both population samples, and therefore, both require population weights to accurately reflect population averages. Although DHS and MICS have individual weights for the specific individuals surveyed, our data education, ethnicity, age and gender come primarily from the household rosters, which are questions asked to everyone in the household. As such, we use 4 More detailed documentation from IPUMS is available here: ( 5 See Sample Information for detailed description of sampling procedures in IPUMS surveys. Available at: 7

11 household sampling weights rather than individual weights, as the primary sampling unit is the household. Social Group Markers We know that the educational experiences and life opportunities of individuals may differ based on their demographic attributes; as such, we group individuals into social groups along four lines: gender, ethnicity, religion and subnational region. All educational attainment values were disaggregated across these lines, with inequality indices calculated separately by gender and type of social inequality grouping (ethic, religious or subnational). Gender Gender is an important variable in our analysis, as the prior literature has suggested that males grievances may be a more important driver of violent conflict. Information on gender was available all surveys with very little missing data. Individuals with missing or unknown gender information were dropped from the analysis (in all cases amounting to less than 5% of observations). Identity Groups We coded individuals into identity groups along two primary lines: ethnicity and religion. Additionally, we use subnational regions to code for an additional measure of subnational inequality based on geographic location. It is important to note that the specific definitions of ethnicity and religion vary substantially across years and countries. The number of groups and the primary marker of ethnicity differ across countries, source datasets, and years. Our ability to group individuals into identity groups is based on the existing datasets, and their definitions. The definition of identity group is important both for the disaggregation of educational attainment information as well as for the estimation of population proportions for each subgroup, which are later used in the calculation of population-weighted inequality measures (see below). We recognize that the DHS and MICS surveys may present challenges in examining group-based identities. As noted by Kuhn and Weidmann (2013) and Cederman, Weidmann and Gleditsch (2011), DHS and MICS are intended to be nationally representative surveys of individuals in certain age and gender categories, but they do not select a sample frame with representative samples of each ethnic group. This presents a problems because sample sizes tend to become very small when the identity group is further disaggregated by age and gender groupings), making estimates less reliable. Moreover, all of the datasets rely on a respondent's self-reported membership in a particular ethnic group, which may not always be accurate, or consistent within groups. Nonetheless, ethnic and religious categories, population weights, and education and economic measures disaggregated along these social lines are routinely and widely used across the conflict literature, particularly the horizontal inequality literature. As such, DHS and MICS, in addition to IPUMS, constitute the most important and comprehensive data sources available to examine horizontal inequalities in education, and provide the most reliable estimates on educational attainment available presently. It is important to note that subnational regions in our datasets may not reflect the actual administrative regions of a country. In the subnational dataset, administrative units were grouped into larger regions, following sampling approaches used by DHS and MICS, or larger units offered by IPUMS. In other cases, subnational regions were realigned in order to ensure that the same regions are measured across time (as internal borders have frequently changed). The user should also note that the number of subnational regions may differ between the national and subnational dataset. Subnational datasets are vertically aligned across years at the subnational unit level. For the national 8

12 dataset, the vertical alignment of subnational regions is not as critical as inequality is captured at the national regions. In other words, to measure inequality between regions for Country A, it matters less if the subnational regions are A, B, and C in Year 1, or D, E, and F in Year 2, as long as the number of units is relatively stable (plus or minus one) year-to-year. In the subnational dataset, the subnational unit itself is the unit of analysis, and therefore it must be stable across years. Ethnicity Ethnicity coding varied across datasets, depending on the definitions used by the survey and the options available to respondents in each survey. In most cases, the types of ethnicity variables available from a given dataset reflected the priorities placed by census or survey design specialists and therefore may miss other ethnic groups potentially present in the population. Proxy variables were sometimes used by the surveys to identify ethnicity (in DHS), and imputation was done for members of the household based on the ethnicity of their household head in some datasets (MICS). In all cases, individuals below the age of 15 were asked their ethnic belonging. The following table gives an overview of ethnic group coding information: Table 4 Overview of Ethnicity Coding Source Indicator(s) Coding Level # Datasets IPUMS a) ethnicity Individual 30 b) race Individual 17 c) indigenous status Individual 3 DHS ethnicity Individual where available, otherwise imputed 94 based on household head MICS ethnicity Household head 29 a. IPUMS Ethnicity Coding Indicators of ethnicity and religion drawn from IPUMS data have not been standardized crossnationally. In order to draw as comprehensive a dataset as possible, our dataset draws from all IPUMS datasets that include information on 1) ethnicity; 2) race; or 3) indigenous status. We do not use language or mother tongue as a marker of ethnicity in IPUMS, primarily because other variables that ask about ethnicity directly are available. The ethnic group was coded according to the following rules: if an IPUMS dataset includes an indicator for an individual s ethnicity, we use that indicator. If it does not, we search for an indicator of race. If the dataset does not include either an ethnicity or race indicator, we use a binary variable indicating indigenous status. This procedure was adopted to standardize coding across a wide variety of country contexts. If a country includes both an indicator identifying indigenous status and an ethnicity marker, our procedure defaults to ethnicity. 6 6 Note: the ordering of variable selection makes little difference for the majority of countries as some countries tend to use the term ethnicity and others race, and it is rare for countries to have both indicators. However, in certain countries, primarily in Latin America, the ordering of variables may matter. In these countries, individuals were first asked if they were indigenous or not, then if they were indigenous, were asked to specify which indigenous group they belonged to, and this group was coded as their ethnicity. We use ethnicity in these cases, as non-indigenous are grouped together as non-indigenous. 9

13 Because IPUMS datasets come from micro-level census data, ethnicity in all IPUMS data is coded at the individual level. From the total 52 IPUMS datasets with some form of ethnicity information, 30 use ethnicity, 19 use race and 3 use indigenous status. b. DHS Ethnicity DHS provides ethnicity of respondent information in the individual and (if available) male datasets. It should be noted that in contrast to IPUMS, DHS will sometimes use the respondent s language as an indicator of ethnicity. The individual dataset contains records for females of reproductive age (15-49) and a subset of males who are husbands of females of reproductive age (usually years of age). In order to obtain the most detailed information on educational levels of household members, ethnicity information must be merged from these separate files onto the household dataset. The procedure we follow for merging datasets is provided in the DHS User Forums on the Program s website, by analysts working with the program. Once the ethnicity data in the male and female datasets are merged onto the main household dataset (which contains information for every person in the households interviewed), there are still a large number of missing values. This is because only a subset of females and males are interviewed. To obtain larger sample sizes, we decided to impute missing values based on other observed values in each household. The procedure for constructing our ethnicity variable is as follows; 1) prioritize observed values where they are available, 2) if observed values are missing, replace with the value of others in the household when all values agree, 3) if from a household where there are differing observed religion values, replace missing values with the value for the household head, and 4) if from a household where there are differing observed religion values and the household head has a missing value, replace with the value for the person who has line number 2 in the household (typically identified as a spouse). The level of detail contained in DHS datasets allows for flexibility in how our ethnicity (and religion) variables are coded from these sources, and we take advantage of this to produce a reliable statistic at the largest possible sample size. c. MICS Ethnicity The majority of MICS datasets include an indicator for both ethnicity and religion; our dataset relies on these two indicators. The primary question used to classify individuals into ethnicity in MICS comes directly from the household questionnaire, and is phrases as: To what ethnic group does the head of this household belong? The definition of an ethnic groups or categories in each country was determined by MICS survey administrators in advance. In MICS, the ethnicity of the household head is applied to all individuals in the household, with the exception of Trinidad and Tobago, where it is determined at the individual level. This practice is the same across all countries, which allows for some degree of standardization; however, it also introduces possible bias if children or other members of the household do not belong to the same ethnicity as that of the household head. Religion With respect to religion, we use the most detailed version of religion that can be obtained. Worldwide, identity-based differences often emerge along sectarian lines even within the same major religion (e.g., Protestant-Catholic, or Sunni-Shia divides). For this reason, we want a variable that groups individuals not only into major religious group, but also into denominations or sects of the same religion. As such, we use the most detailed religion variable possible as our primary classification for religion. In IPUMS, this is an important distinction; however, in DHS and MICS, there is only one religion variable, and so we use the provided variable. 10

14 In all cases, we retained the original structure of the survey, including individuals who were categorized as having No religion or Unknown religion (distinct from being missing or not-inuniverse). In these cases, individuals are categorized as religious or ethnic group, or incorporated into the minority group. For our purposes, it was largely irrelevant whether the largest religious group was Muslim, Sunni, Buddhist or No Religion, etc. Nonetheless, we also extracted information on both the denomination and the percent of the population each denomination comprised in a meta-data file. This meta-data file allows for comparison of how a nation s religious composition has changed over time or across surveys. In calculating the percent of the population belonging to each religious group, the numerator was defined as all individuals identifying with each religion; the denominator was all individuals in the population including children and the elderly. Although we limited our educational calculations to those aged 15-54, we believed that it was important to calculate the distribution of religion and ethnicity affiliation according to population as a whole, including young people and those 55 and older to take into account varying birth rates. Table 5 Overview of Religion Coding Source Indicator Level of Coding # Datasets IPUMS Religion, detailed Individual 73 DHS Religion Individual where available, otherwise imputed based 70 on household head MICS Religion Household head 27 a. IPUMS Religion Coding In the IPUMS datasets, two variables are available for religious identification Religion and Religion, Detailed. Religion classifies individuals into the world s major world religions, while the detailed version of religion classifies individuals into smaller sects and denominations. 7 In our dataset, we rely on the detailed religion classification. This variable is not standardized cross-nationally or over time within the same nation. In some cases, changes to census questions over time means that the specific religious group that is classified as the majority religion will change over time, even if the major religious groups have not changed. This only occurred in IPUMS surveys, and reflects changes in the survey constructions such that one survey will disaggregate only to major religions (i.e., Christianity), while the next survey will disaggregate much further (i.e., Catholic, Protestant, or even within Protestant denominations). The resulting changes to the definition and size of groups means that it can be difficult to compare groups over time. b. DHS Religion Coding The same process used to code ethnicity for DHS datasets is used to code religion. The religion variable (v130 and mv130 for individual and male, respectively) is more often available in DHS datasets than the ethnicity variable. 7 Detailed information on the the religion variable across countries can be found from the IPUMS website: 11

15 c. MICS Religion Coding In the MICS survey, there is a specific question that asks about the religion of the head of household. The working of the question has changed slightly between various MICS samples, but can nonetheless be considered a good measure of religious affiliation. In MICS Wave 3, the question asked: What is the religion of the head of this household? In Wave 4, it asked: What is the religion the head of household practices? As with ethnicity, the religion of the head of household was applied to all members on the household roster. Treatment of Missing Ethnic and Religious Data Missing data creates potential for bias at various stages of the dataset compilation. Missing data on ethnic or religious identity means that we may miscalculate the percentage of the population belonging to each ethnic or religious group. Although subsequent problems posed by the incorrect estimation of group size are likely very small, in extreme cases, we may fail to identify a sizeable subgroup due to missing data, leading us to group them with other minority groups. In general, we believe the comprehensiveness of our datasets makes this a small concern to the reliability of our estimates. IPUMS In most surveys, we had information on all individuals ethnic or religious affiliation, with few missing observations. However, in a few survey samples from IPUMS, only adults were asked about religious affiliation. In these surveys, we applied the religion or ethnicity of the household head to all children with missing observations. The overall size of each group changed, but did not seem to alter the order of major religion groups in these countries (see Technical Appendix for more detail). DHS Missing values in the original individual and male datasets were not an issue, for the most part. Gabon (2000) and Mozambique (2003) had 10-20% missing values for ethnicity/religion in the original datasets, but almost all other datasets had less than 1% missing values for these variables. Once these values were merged onto the main household dataset (detailed in the sections on ethnicity and religion coding), missing values were imputed. The result is that in our constructed ethnicity and religion variables, roughly 10-20% of all observations in interviewed households had missing values. Surveys with particularly large numbers of missing values in our constructed variables include Pakistan (2006) and Ghana (2008). In both cases it appears that this is due to an abnormally small number of observed values compared to the total number of household records. For our study, only complete cases were used. MICS In MICS, very few samples had any missing ethnic or religious data. Because ethnicity and religion were coded at the household level, individuals with missing data had to be dropped, because we had no additional data from the household to apply to individuals. In Afghanistan, 3.14% of total observations had missing ethnic data, and in Cote d Ivoire, 17% had missing ethnicity data. In Vietnam 2006, 85% of individuals were missing data on religion, and given this significant percentage, the religion data was dropped entirely. Calculating Group-Level Attainment The number of identity-based groups in each country varies substantially, both due to demographic changes and changes in data collection procedures. Among the countries for which we have data, 12

16 some are extremely heterogeneous, while others tend to be ethnically and religiously homogeneous. Additionally, we noted that some national censuses seem to allow write-in options for ethnicity, while others have standardized groups. The exact number of the ethnic and religious groups available, then, will differ by country and year. Moreover, even within nations, long-term demographic shifts have altered the composition of the population over time, and so we see new groups emerging and disappearing from different datasets. These shifts mean that it can be difficult to accurately estimate the size of horizontal inequalities in education over time. In calculating our indicators for inequality in group-level educational attainment, our interest was in relative advantage or disadvantage, not absolute educational attainment. For this reason, we did not attempt to trace the educational attainment of specific religious or ethnic groups over time, or make groups comparable across countries or even over time within the same country. To do so would be infeasible on a cross-national scale given the small number of surveys available in any one country. Instead, the goal was to successfully identify the major ethnic and religious groups in each country and use this information to calculate differences in educational attainment across identity groups. To calculate group-level attainment, we first identified the major ethnic and religious groups in each country. To do this, we calculated the percentage of the population that each ethnic group and religious denomination comprised. In the vast majority of surveys, we calculated each this included all individuals; however, in selected surveys this required coding children under 15 with the religion or ethnic group of the household head (see Technical Appendix). 8 The question of how to treat minority groups was one that required careful consideration. The meaning and experience of every minority group clearly varies cross-nationally. In some surveys, we found that ethnicity was disaggregated to incredibly small levels, such that only a few individuals were listed as having each ethnicity. Because our data analysis process required a minimum number of individuals in each ethnic and religious group, in order for us to successfully disaggregate educational attainment by group, age and gender, we defined a minority group as any group that composed less than 5% of the total population. All minority groups are categorized into a minority group (i.e., ethnicmin and religmin). The actual ethnic groups and religions included in the minority variable may vary across samples even within the same country, if the religious composition of the population changes over time or due to varied definitions of religion and ethnicity across samples. Minimum Observation Cut-Offs We also imposed a minimum number of observations (N) for each sub-group included in the analysis, in order to ensure that our educational attainment indicators are accurate and reliable. When the number of observations in each group becomes too low, it becomes susceptible to the effect of outliers, and may not accurately reflect the actual level of educational attainment in the population as a whole. To avoid problems associated with small samples, we imposed a minimum of 50 individuals in each sub-group (ethnic or religious group, by age cohort and gender). Any group that was smaller than 50 individuals was dropped from the analysis. The problem of a small subgroup was more acute in DHS and MICS surveys, which are nationally representative samples, as well as very small nations. We were also much more likely to lose observations on sub-groups among the oldest age cohorts in our samples. Back Projections of Educational Attainment Data EPDC back projections estimate the historical education levels of the cohort ages years old up to 40 years in the past. The back projections use the attainment data for ethnic, religious, and 8 Sample codes: relig1 = largest religious denomination; relig2 = second largest religious denomination 13

17 subnational groups extracted from IPUMS, DHS, and MICS data as baseline data. The back projections rely on International Institute for Applied Systems Analysis (IIASA) methodology (Lutz et al. 2007) and use gender-specific information on life expectancy and survival ratios based on data from the UN general model life tables. EPDC expands the application of the methodology to the social groups of interest to this study (ethnic, religious, and subnational) and to a greater number of developing countries. The back projections were only developed for ethnic and religious groups. Our assumption is that ethnic and religious groups remain stable over time, and estimates of educational attainment for today s elderly members of a given ethnic group provide a reasonably good proxy of the educational attainment for youth several decades prior to the source date. We do not apply back projection methodology on educational attainment levels disaggregated by subnational unit because internal migration that has taken place in most of the countries in our dataset makes it difficult to assume that education levels of the elderly population of a given subnational unit is a good reflection of its educational landscape in prior years. This is particularly challenging in post-conflict settings that severely disrupt the demographics of all subnational regions. Like IIASA, we adjust the life expectancy data, specific to each country, year, and gender, for education levels. We use the general expectancy at the age of 15 for those at the education level that is most dominant in that age group. The assumption regarding the relationship between life expectancy and the level of education is as follows: If people with no schooling have the life expectancy of X, then people with completed primary education have the life expectancy of X+1, and people with completed secondary education have the life expectancy of X+3 (one for primary, plus two for secondary resulting in three years on top of no schooling life expectancy). In other words, if no schooling is most dominant, then we assume that people with no schooling have the same expectancy as the general expectancy at the age of 15, and we add 1 year for completed primary education and an additional 2 years for secondary education completed. If primary completed is most dominant than we assume that people with primary education completed have the same expectancy as the general expectancy and we subtract 1 for no schooling and add 2 for secondary education completed. We assume that most though not all of the educational attainment in older groups was acquired by age 24. Specially, our back projections assume that 80% of year olds with completed primary education had achieved that level 10 years before, and that 60% of year olds with completed secondary education had reached that educational level 10 years earlier. These assumptions are based on empirical observation and comparison of different age cohorts across all datasets, and are intended to approximate the real levels of attainment in previous time periods. Our assumptions were developed and tested through country comparisons where census data was available from the same period as back-projected estimates. Based on our analysis, back projected estimates may be off from the true values by up to 12 percentage points, for the lowest level of educational attainment (no education for less developed countries or primary complete for developed countries). While the degree of error that our back-projections introduce varies by country, we assume that the bias affects all groups within a country equally, and therefore would have minimal if any effect on our inequality indices. Once the actual and back-projected education attainment values are compiled for all countries and ethnic groups, some countries may have duplicate sets of projection results from different sources. For example, if Country A has a census every ten years in 1990, 2000, and 2010, back projections for year 1980 will be available from each of the censuses, with the closest set (a 10 year back projection) 14