The Measurement of Educational Inequality: Achievement and Opportunity1 ´ re Francisco H. G. Ferreira and Je ´ mie Gignoux Two related measures of educational inequality are proposed: one for educational achievement and another for educational opportunity. The former is the simple vari- ance (or standard deviation) of test scores. Its selection is informed by consideration of two measurement issues that have typically been overlooked in the literature: the implications of the standardization of test scores for inequality indices, and the possi- ble sample selection biases arising from the Program of International Student Assessment (PISA) sampling frame. The measure of inequality of educational opportu- nity is given by the share of the variance in test scores that is explained by pre- determined circumstances. Both measures are computed for the 57 countries in which PISA surveys were conducted in 2006. Inequality of opportunity accounts for up to 35 percent of all disparities in educational achievement. It is greater in (most of ) conti- nental Europe and Latin America than in Asia, Scandinavia, and North America. It is uncorrelated with average educational achievement and only weakly negatively corre- lated with per capita gross domestic product. It correlates negatively with the share of spending in primary schooling, and positively with tracking in secondary schools. JEL codes: D39, D63, I29, O54 Educational inequalities have long been a matter of significant policy concern, in both developed and developing countries. Some view educational achieve- ment as a dimension of well-being in its own right, or at least as a fundamental input into a person’s functionings and capacity to flourish (Sen, 1985). Education is also a powerful predictor of earnings, as we have known since the early days of work on human capital (for a review see, e.g., Psacharopoulos, 1994). More recent research has also found that inequality in educational 1. Francisco Ferreira (corresponding author) is a Lead Economist with the Development Research Group at the World Bank and a Research Fellow at the Institute for the Study of Labor (IZA); email: fferreira@worldbank.org. Je ´ mie Gignoux is an Assistant Professor at the Paris School of Economics ´ re (PSE) – French National Institute for Agricultural Research (INRA); email: gignoux@pse.ens.fr. We are grateful to the Editors, three anonymous referees and to Gordon Anderson, Markus Ja ¨ ntti, Maria Ana Lugo, John Micklewright, Alain Trannoy and participants at conferences and seminars in Barcelona, Buenos Aires, Chicago, Oxford and St. Gallen for helpful comments on earlier drafts. We are solely responsible for any remaining errors. The views expressed in this paper are those of the authors, and should not be attributed to the World Bank, its Executive Directors, or the countries they represent. THE WORLD BANK ECONOMIC REVIEW, VOL. 28, NO. 2, pp. 210– 246 doi:10.1093/wber/lht004 Advance Access Publication February 20, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 210 Ferreira and Gignoux 211 achievement and earnings inequality are correlated, both over time within the United States and across countries (see, e.g., Blau and Kahn, 2005; and Bedard and Ferrall, 2003). Education is also correlated with health status, and in some cases with political participation in the democratic process, so that inequalities in the former may translate into undesirable gaps and gradients in other dimen- sions as well. For all of these reasons, people care about the distribution of education. Those concerned about fairness and social justice care also about the distribu- tion of opportunities for acquiring a good education and, in particular, about the degree to which family background and other pre-determined personal characteristics determine a person’s educational outcomes. Nevertheless, there is much less agreement on how those concepts—inequality in educational out- comes, and inequality of opportunity to a good education—should be mea- sured. Constrained by data availability, early work comparing inequality in education across countries focused on educational attainment: the number of years of schooling a person had completed or, in some cases, broader ‘levels’ of education, such as primary, secondary, or higher. Thomas, Wang and Fan (2001) compiled a set of Gini coefficients for years of schooling for 85 coun- tries, over the period from 1960 to 1990. Castello ´ and Dome ´ nech (2002) and Morrisson and Murtin (2007) also examine inequality in years of schooling across a large number of countries. Interesting though those comparisons were, there is widespread agreement that a year of schooling is a problematic unit with which to measure “educa- tion.” Does a student learn the same amount in 6th grade in Zambia as in Finland? Is the value of one year of schooling the same even across different schools in a single country or city? The growing availability of data on student performance in comparable tests has confirmed what one already suspected: that the answer to these questions is generally ‘no’. The quality—and hence the ultimate value—of education varies considerably, both within and across countries. Over the last decade, different projects have compiled school-based surveys that administer cognitive achievement tests to samples of students across a number of countries, as well as collecting (reasonably) comparable information about the students’ families and the schools they attend. The OECD’s Program of International Student Assessment (PISA) and the International Association for the Evaluation of Educational Achievement’s Trends in International Mathematics and Science Study (TIMSS) are perhaps the best known, but the Progress in International Reading Literacy Study (PIRLS), which is applied to younger students, shares a number of common features.1 But performance in a test, while probably preferable to a simple indicator of enrollment or attendance, is not a perfect measure of learning either. For one 1. There is also an International Adult Literacy Survey (IALS), which is applied to adults long after they have left school. 212 THE WORLD BANK ECONOMIC REVIEW thing, tests and test items (i.e. questions) vary in difficulty. The final result is known to measure scholastic ability or learning achievement only imperfectly. For this reason, all of the aforementioned surveys present scores constructed from the raw results by means of Item Response Theory (IRT) models, which attempt to account for “test parameters,” so as to better infer true learning. This process generates an arbitrary metric for test scores, which are then typi- cally standardized to some arbitrary mean and standard deviation. Using these standardized test scores, a number of studies have attempted to provide international comparisons of educational inequality on the basis of achievement, rather than attainment. Micklewright and Schnepf (2007) and Brown et al. (2007) examine the robustness of measures of central tendency and dispersion in the distribution of student achievement obtained using differ- ent surveys, by comparing the measures and country rankings across them. Marks (2005), Schultz, Ursprung and Wossmann (2008), and Macdonald et al. (2010) examine the question of intergenerational persistence in educational achievement, which is closely related to that of inequality of opportunity, and present cross-country comparisons of measures of the association between student achievement and certain family characteristics. This paper seeks to contribute to that literature by proposing two simple and closely-related measures of inequality—one for educational achievement and another for opportunity to acquire education—and reporting them for all countries that participated in the 2006 wave of PISA surveys. To measure in- equality in achievement, we propose simply using the variance or the standard deviation of test scores. But we arrive at this simple proposal by considering the implications of two issues specific to the distribution of test scores for the measurement of inequality. These two issues are: (i) the fact that many common inequality indices are not ordinally invariant in the standardization to which IRT-adjusted test scores are generally subjected; and (ii) the fact that PISA student samples are likely to suffer from non-trivial selection biases in a number of countries. The choice of the variance (or the standard deviation) ad- dresses the first issue. We also propose two alternative two-sample non- parametric procedures to assess the robustness of the inequality measure to the sample selection biases, and implement them in the four countries for which PISA sample coverage (as a share of the total population of 15 year-olds) is smallest. The proposed measure of inequality of educational opportunity draws on the recent literature on inequality of opportunity in the income space, but is also adapted to the specificities of educational data and the resulting choice of measure for inequality in achievement. It also utilizes information on student background more comprehensively than all previous studies we are aware of, and is additively decomposable both across circumstances and population sub- groups. The measure is also isomorphic to (inverse) measures of educational mobility. Ferreira and Gignoux 213 We report our measures of inequality in educational achievement and oppor- tunity for the 57 countries that took part in the PISA 2006 exercise. Each measure was computed separately for each of the three tests applied by PISA: mathematics, reading and science. But there was a good measure of agreement between their rankings, and we often refer only to the math results in the text. We find considerable variation in the standard deviation of test scores, from lows of around 80 (for Indonesia, Estonia and Finland) to highs near 110 (in Belgium and Israel).2 Similarly stark variation exists in our measure of inequal- ity of opportunity, from 0.10 – 0.15 for Macau (China), Australia, and Hong Kong SAR, China, up to 0.33 – 0.35 in Bulgaria, France and Germany. The paper is organized as follows. Section I describes the data sets we use. Section II considers the implications of test score standardization and of the PISA sampling frame for the measurement of inequality in educational achieve- ment, and reports the standard deviation in test scores for our sample of coun- tries. Section III proposes our measure of inequality of educational opportunity (IOp), discusses some of its properties, and presents results. Section IV applies the proposed measures by examining how they correlate with two educational policy indicators across countries. Section V concludes. I . D ATA Two broad kinds of data are used for the analysis in this paper. The first is the complete set of PISA surveys, for all 57 countries that participated in the 2006 round. The second is a group of four household surveys, for Brazil, Indonesia, Mexico and Turkey, which are used as ancillary surveys in the two-sample non-parametric sample selection correction procedures described in Section 3. We briefly describe each of these in turn. The PISA 2006 Data Sets The third round of the Program of International Student Assessment surveys was conducted in 57 countries between March and November, 2006. Two earlier rounds were collected in 2000/2002 (in 43 countries), and in 2003 (in 41 countries). A fourth round has since been collected in 2009. Most OECD countries were surveyed, as were a number of developing countries in Asia, Latin America, North Africa and the Middle East. Table 1 lists all participating countries in the 2006 round, as well as their sample sizes. In each country, fifteen year-olds enrolled in any educational institution, and attending grade 7 or higher, were sampled. All children surveyed took three tests: in reading, mathematics, and science.3 Their performance in these tests 2. But the low variance for Indonesia is a good example of the sensitivity of these measures to assumptions made about the nature of selection into the test-taking sample. Under our scenario of “extreme” selection on unobservables, the variance of math scores for Indonesia triples. See below. 3. The data for achievements in reading for the United States were not issued after a problem occurred during the field operations in that country. T A B L E 1 . Sample Statistics, Mean Scores and the Standard Deviation in PISA Test Scores Reading SD Math SD (SE Science SD (SE 214 # Obs. Coverage rate Reading Mean (SE of SD) Math Mean of SD) Science Mean of SD) Asia & North Africa Azerbaijan 5184 0.88 355.0 70.26 2.12 476.8 47.96 1.64 385.3 55.68 1.92 THE WORLD BANK ECONOMIC REVIEW Hong Kong SAR, China 4645 0.97 538.9 81.79 1.92 551.4 93.39 2.31 546.1 91.71 1.92 Indonesia 10647 0.53 383.9 74.79 2.39 380.7 80.01 3.18 384.8 70.06 3.26 Israel 4584 0.76 441.3 119.34 2.79 443.3 107.33 3.20 455.6 111.45 1.92 Japan 5952 0.89 409.5 102.38 2.34 389.2 91.01 2.06 427.1 100.12 2.01 Jordan 6509 0.65 500.2 94.09 2.24 525.6 83.71 1.95 533.7 89.86 1.89 Korea 5176 0.87 290.5 88.29 2.68 315.9 92.59 3.12 326.3 90.06 2.35 Kyrgyzstan 5904 0.63 556.1 102.10 2.51 547.2 86.98 2.03 521.9 83.86 2.03 Macao-China 4760 0.73 490.6 76.36 2.26 524.4 83.90 1.51 509.5 77.83 1.58 Qatar 6265 0.90 312.5 108.12 1.15 317.7 90.24 1.39 349.1 83.29 1.37 RussianFederation 5799 0.81 442.4 93.23 1.87 478.7 89.53 1.58 481.5 89.57 1.33 Chinese Taipei 8812 0.88 506.7 84.38 1.73 562.7 103.11 2.16 543.7 94.45 1.63 Thailand 6192 0.72 425.2 81.85 1.73 425.5 81.43 1.57 429.7 77.17 1.45 Tunisia 4640 0.90 379.0 97.30 2.49 363.9 91.95 2.34 384.2 82.38 2.05 Turkey 4942 0.47 452.9 92.90 2.75 428.2 93.24 4.32 427.6 83.20 3.14 Latin America Argentina 4339 0.79 383.9 124.22 3.63 388.1 101.14 3.48 398.3 101.24 2.62 Brazil 9295 0.55 389.2 102.46 3.34 365.6 92.02 2.65 385.3 89.28 1.93 Chile 5233 0.78 447.9 103.24 2.44 417.1 87.44 2.17 443.1 91.68 1.72 Colombia 4478 0.60 390.3 107.83 2.38 373.8 88.04 2.42 391.9 84.81 1.81 Mexico 30971 0.54 427.4 95.68 2.27 420.7 85.27 2.16 422.6 80.70 1.47 Uruguay 4839 0.69 424.7 121.22 2.03 435.5 99.30 1.77 437.7 94.44 1.73 North America & Oceania Australia 22646 0.87 508.7 96.25 1.43 516.3 85.79 1.03 523.1 94.19 1.14 Canada 14170 0.87 512.3 93.79 1.00 517.4 88.03 1.09 522.5 100.23 1.02 New Zealand 4823 0.84 522.7 105.21 1.58 523.8 93.27 1.20 532.7 107.30 1.36 United States 5610 0.85 474.7 89.75 1.90 488.3 106.07 1.68 Eastern Europe Bulgaria 4498 0.83 406.8 117.51 4.00 417.4 101.10 3.65 439.1 106.72 3.20 CzechRepublic 5932 1.01 509.6 111.21 2.90 536.0 103.14 2.08 537.6 98.41 2.00 Estonia 4865 0.94 502.4 85.19 1.87 516.8 80.68 1.54 533.7 83.75 1.09 Croatia 5213 0.85 477.6 88.83 2.12 467.3 83.31 1.50 493.7 85.72 1.44 Hungary 4490 0.85 488.1 94.39 2.37 496.2 91.04 1.94 508.7 88.20 1.53 Lithuania 4744 0.93 469.3 95.54 1.51 485.6 89.80 1.73 486.5 89.99 1.52 Latvia 4719 0.85 484.9 90.70 1.69 491.2 82.81 1.51 493.8 84.38 1.30 Montenegro 4455 0.84 388.2 89.41 1.64 395.8 84.45 1.80 408.8 79.69 1.19 Poland 5547 0.94 512.6 100.22 1.48 500.9 86.52 1.13 503.3 89.87 1.11 Romania 5118 0.66 392.0 91.86 2.93 415.0 83.97 2.85 416.6 81.16 2.37 Serbia 4798 0.83 402.9 91.84 1.69 436.6 91.76 1.77 436.9 85.15 1.56 SlovakRepublic 4731 0.95 470.6 105.08 2.51 495.1 94.53 2.47 491.2 93.15 1.79 Slovenia 6595 0.88 468.6 87.97 2.47 482.2 89.25 1.36 494.2 98.11 1.35 Western Europe Austria 4927 0.92 494.0 108.16 3.16 509.5 98.06 2.29 513.9 97.83 2.41 Belgium 8857 0.99 507.1 110.02 2.81 526.9 106.13 3.31 516.3 99.70 2.00 Switzerland 12192 1.02 496.6 94.07 1.71 528.3 97.44 1.60 508.0 99.31 1.61 Germany 4891 0.95 496.5 111.95 2.67 504.3 99.08 2.53 516.2 99.98 1.99 Denmark 4532 0.85 493.8 89.30 1.63 512.2 84.85 1.53 494.7 93.13 1.42 Spain 19604 0.87 479.5 88.84 1.14 501.7 88.92 1.09 504.5 90.54 0.97 Finland 4714 0.93 547.1 81.23 1.08 549.0 80.87 1.01 563.4 85.62 1.00 France 4716 0.91 488.7 103.95 2.75 496.4 95.58 1.96 496.1 101.57 2.09 United Kingdom 13152 0.94 495.6 101.92 1.69 497.3 88.92 1.31 514.3 106.79 1.50 Greece 4873 0.90 461.9 102.61 2.92 462.0 92.30 2.37 476.6 92.12 2.03 Ireland 4585 0.94 518.6 92.39 1.86 502.3 81.99 1.50 509.5 94.35 1.50 Iceland 3789 0.96 485.0 97.09 1.23 505.6 88.08 0.89 491.0 96.87 0.95 Italy 21773 0.90 477.0 108.76 1.74 473.6 95.82 1.66 487.2 95.56 1.31 Liechtenstein 339 0.84 510.7 95.14 2.93 524.9 93.05 2.17 522.3 96.96 2.10 Luxembourg 4567 1.03 480.1 99.85 0.72 490.5 93.15 0.73 486.8 96.53 0.67 Ferreira and Gignoux Netherlands 4871 0.96 513.9 96.62 2.47 537.4 88.60 2.18 530.8 95.63 1.64 Norway 4692 0.97 484.4 105.15 1.92 489.8 91.58 1.38 486.9 96.12 1.98 Portugal 5109 0.78 476.8 98.82 2.28 470.9 90.65 1.97 479.0 88.56 1.71 Sweden 4443 0.97 509.0 98.21 1.77 503.2 89.66 1.37 504.2 94.21 1.40 Notes: The standard deviation (S.D.) of test scores is used as an ordinal measure of inequality in achievement, as discussed in the text. Standard errors reported in the columns next to the S.D. are bootstrapped. Source: Authors’ analysis based on data from PISA 2006. 215 216 THE WORLD BANK ECONOMIC REVIEW forms the basis for the assessment of their learning or cognitive achievement. Yet, educationalists seem agreed that raw, unadjusted test scores are of little value. Test questions (or ‘items’) vary in their degree of difficulty, and simply adding up correct answers, or weighing them arbitrarily, does not correctly measure the latent variable of interest—cognitive achievement. Instead, the ed- ucational community in charge of international tests such as PISA, TIMSS, PIRLS and IALS processes raw scores through statistical techniques known as Item Response Theory (IRT). In essence, an item response model consists of an equation of the form: pðsju; aÞ: ð1Þ Equation (1) gives the probability of scoring s in a given test, conditional on in- dividual latent cognitive ability u and test item parameters a (such as their diffi- culty). Given an additional assumption about the distribution À Á latent ability of in the population (usually a normal law such as u N mu ; s2 u ) and an ob- served distribution of raw scores, F(s), the IRT model is used to back out a dis- tribution of the latent variable u.4 Unfortunately, the inference of statistics summarizing distributions of unob- served latent variables such as u is subject to a specific small-sample measure- ment error problem: Each pupil answers a limited number of items so that it is not possible to estimate individual abilities accurately. In this situation, the dis- tribution of estimates for individual abilities obtained with traditional methods (such as maximum likelihood estimates) does not converge to the population distribution of these abilities as the number of examinees increases (Mislevy et al. 1992). These estimates of parameters of this distribution are thus incon- sistent (although the asymptotic bias decreases with the number of items per examinee). The standard solution to this measurement error problem in psychometrics is to draw a number of plausible values for the latent variable for each individ- ual. The marginal distribution of ability for each pupil, conditional on his or her answers and a set of observables, is estimated, and a number M of draws from this distribution is obtained. These M draws are known as plausible values of a pupil’s score. To estimate a given statistic s, each of the M datasets containing one plausible value per pupil should be used separately to obtain a sm . The final estimate ^ set of estimates ^ s of the statistic s is given by the average of the M estimates ^ sm . For PISA, M ¼ 5, which implies that five “data sets” are used separately to compute sample statistics (e.g. means and standard devia- tions). In conformance with the advice in the PISA Data Analysis Manual (OECD, 2009), all of the estimates presented in this paper are computed as av- erages of the summary statistics estimated separately for each of the five data 4. See Baker (2001) for a general introduction, and OECD (2006) for a description of how the IRT method is applied to PISA surveys. Ferreira and Gignoux 217 sets (rather than as summary statistics of the distribution obtained by first aver- aging across plausible values).5 This use of Item Response Theory involves a number of functional form as- sumptions which are not innocuous. Brown et al. (2007) have shown, for in- stance, that the final distribution of test scores can be sensitive to differences in the specification of the model used to estimate equation (1).6 Here, however, we are concerned with the standardization that happens after the IRT adjust- ment (and the appropriate treatment of the distributions of plausible values generated in the process). Once that procedure is complete, and a new distribu- tion of ‘adjusted’ test scores (which we denote by x) has been generated, this latter variable is standardized, according to a simple formula such as: ^À s Á yij ¼ m ^þ xij À m ð2Þ s In equation (2), xij denotes the ( post-IRT, pre-standardized) test score for indi- vidual i in country j. m and s denote their original mean and standard devia- tion across all countries in the sample (the world, or the OECD, for example). ^ (s m ^ ) is the new arbitrary mean (standard deviation) for the standardized dis- tribution. In the PISA procedure, it has a value of 500 (100). It is the distribu- tions of yij that are used in computing means and inequality indicators for each country j in the PISA data set. As we will see in the next section, the operation described by equation (2), even if the IRT procedure that precedes it is taken as given, poses serious issues for inequality measurement. In addition to standardized test scores, the PISA data set contains information on a number of individual, family and school characteristics for each test-taker. The presence of these covariates accounts for a large part of the interest of the research community on the PISA data. For the analysis of inequality of opportu- nity in education, we focus on a subset of these covariates that are informative of the family background and other inherited circumstances of the child. Ten such variables are used: gender, father’s and mother’s education, father’s occupa- tion, language spoken at home, migration status, access to books at home, 5. The sampling variance of population parameter estimates are computed using the Balanced Repeated Replication (BRR) weights provided within the data (PISA 2006). BRR is a replication method for multistage stratified sample designs similar to the Jackknife. The particular variant of the BRR known as Fay’s method was used. For PISA it consisted in forming pairs (called strata) of schools (the primary sampling units) and drawing a number of replicates of the sample (using a so-called Hadamard matrix). 80 replicates were performed. Each of these replicate attributes weight 1.5 to one of the school and weight 0.5 to the other in each strata, the selection being different for each replicate. The BRR weights are then computed as the product of students’ original sampling weights and the school weight (1.5 or 0.5) for each particular replication. See Mislevy (1991) and Mislevy et al. (1992) for a more detailed discussion. 6. Brown et al. (2007) investigated this question by applying the IRT model used in the 1999 TIMMS sample retrospectively to the 1995 sample, which had used a different specification for (1). Although changes were small for most developed countries, there were some non-trivial re-rankings among developing countries. 218 THE WORLD BANK ECONOMIC REVIEW durables owned by the households, cultural items owned, and the location of the school attended (used as an indicator or a rural or urban upbringing).7 Parental education is measured by the highest level completed and is coded using ISCED codes into four categories: a) no education or unknown level; b) primary education (ISCED level 1); c) lower secondary education (ISCED level 2), upper secondary (ISCED level 3), or post-secondary non-tertiary education (ISCED level 4); and d) college education (ISCED level 5)). Father’s occupation is classified using ISCO codes. We aggregate occupations into three broad cate- gories: a) legislators, senior officials and professionals, technicians and clerks; b) service workers, craft and related trades workers, plant or machine operators and assemblers, and unoccupied individuals; and c) skilled agricultural and fishery workers, elementary occupations or unknown occupation. The variable for language spoken at home is a dummy identifying a language other than the language of the test. The migration status variable is a dummy identifying a first or second generation migrant as an individual who was, or whose parents were, born in a foreign country.8 The number of books at home variable, an indicator of parental human capital, is a categorical variable coded into four categories: a) 0 to 10 books; b) 11 to 25 books; c) 26 to 100 books; and d) more than 100 books. Ownership of durables, an indicator of family wealth, is captured by six dummy variables indicating the ownership of a) a dishwasher; b) a DVD or a VCR player; c) a cell phone; d) a television; e) a computer; f) a car. Ownership of cultural possessions is captured by three dummy variables indicating the ownership of a) books of literature; b) books of poetry; and c) works of arts ( paintings are mentioned as an example of such works in the formulation of the question). School location is a proxy for the person’s inherited spatial en- dowment and we recode it using three categories: a) villages or small towns (less than 15,000 inhabitants); b) towns (between 15,000 and 100,000 inhabi- tants); and c) cities (larger than 100,000 inhabitants). School location informa- tion was not collected in France; Hong Kong SAR, China; and Liechtenstein. A final data issue worth highlighting is that of sample coverage and represen- tativeness. PISA samples were designed to be representative of the population of 15 year-olds who are enrolled in grade 7 or higher in any educational insti- tution. The samples are not, therefore, representative of the total population of 7. School-level variables are not used in this analysis deliberately, for reasons which should become clear in Section 4. 8. Naturally, non-random measurement error in these covariates would be undesirable. In particular, one might be concerned that information on family background elicited from children might be systematically misreported. To assess the seriousness of this problem, we use supplementary information on parental education asked directly of parents, which is available for sixteen countries in PISA 2006. For these countries, attainments reported in the child and parent questionnaires for both the mother and the father match exactly for 70 percent of children. Moreover there is no evidence that children tend to systematically report higher than actual attainments: The shares of children reporting higher and lower attainments than the parents are close to 15 percent for both fathers and mothers. Ferreira and Gignoux 219 15 year-olds in each country: Children who dropped out of school before they turned fifteen, as well as those who are so delayed that they are in grade 6 or lower at age fifteen, are purposively excluded. In addition, sampling flaws induce an additional under-coverage of enrolled 15 year olds. PISA documenta- tion suggests that this arises from the fact that their sampling frame (a listing of schools and sampling weights) is established in the year preceding the surveys, on the basis of current school enrollment on that year. But some schools close down between the two years, and new ones are not included in the sample. Changes in the enrollment of 15 year-olds arising from this process are not taken into account. The PISA sample coverage rate, defined as the ratio of the covered student population (using PISA expansion factors) to the total population of 15 year- olds, varies considerably across countries, and is reported in column 2 of Table 1. Although coverage is typically high in OECD countries, it is low in many developing ones: Coverage rates are as low as 47% for Turkey, 53% for Indonesia, 54% for Mexico, and 55% for Brazil. Overall, coverage is less than 80% of the total population of 15 years-olds in fifteen countries. Table 2 pro- vides a sense of the sources of exclusion for the four countries in our dataset with the lowest coverage rates, by decomposing those selected out of the sample into children no longer in school, children with excessive delays, and those missed due to PISA sampling issues. It should be obvious from these magnitudes that any international comparison of countries with vastly different coverage rates must seek to address the problem in some way, and we suggest two alternatives in Section 3. Ancillary Household Survey Data Sets Our proposed procedure to examine the sensitivity of inequality measures to sample selection, which is described below, relies on using information on fifteen year-olds from general-purpose household surveys. While these surveys may have their own sampling issues, these are not dictated by school enroll- ment or delay status, or by school closures, openings and reforms. We obtained such household surveys for the four countries with the lowest coverage rates in the 2006 PISA sample: those reported in Table 2. For Brazil, we used the Pesquisa Nacional por Amostra de Domicı ´lios (PNAD) 2006. For Indonesia, we used the SUSENAS 2005. For Mexico, the Encuesta Nacional de Ingresos y Gastos de los Hogares (ENIGH) for 2006 was used. For Turkey, the Household Budget Survey (HBS) 2006 was used. All four are large-sample household surveys with national coverage and rep- resentative down to the regional level, which are fielded on an annual basis by each country’s national statistical authority. The PNAD 2006 collected infor- mation from a sample of about 119,000 households and 410,000 individuals; SUSENAS 2005 from 257,900 households and 1,052,100 individuals; the ENIGH 2006 from 20,900 households and 83,600 individuals; and the HBS 2006 from 8,600 households and 34,900 individuals. We restrict the samples 220 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . PISA Sample Coverage: Analysis for Four Developing Countries Brazil Indonesia Mexico Turkey Expanded 15 year-old populations, using PISA data and weights Total population of 15-year-olds 3 390 471 4 238 600 2 200 916 1 423 514 Total enrolled population of 15-year-olds at grade 7 or above 2 374 044 3 119 393 1 383 364 800 968 Weighted number of students participating to the assessment 1 875 461 2 248 313 1 190 420 665 477 Coverage rate of the population of 15-year-olds, from PISA (%) 55.3 53.0 54.1 46.7 Total missed children (%) 44.7 47.0 45.9 53.3 Composition of those not covered by PISA samples Out-of-school children (%) 10.2 25.5 24.1 21.6 Delays of more than two years (%) 19.8 0.9 13.1 22.2 PISA sampling issues (%) 14.7 20.6 8.8 9.5 Source: PISA 2006 surveys; PNAD 2006 for Brazil, Susenas 2005 for Indonesia; ENIGH 2006 for Mexico, and HBS 2006 for Turkey. The share of fifteen year-olds who are not enrolled in school comes from the ancillary household surveys. Those delayed by more than two years come from household surveys, and are checked with PISA administrative records. The last row is derived as a residual. Ferreira and Gignoux 221 to children aged 15, for which we have 7,626 observations in the PNAD 2006; 22,600 in the SUSENAS 2005; 1,921 in the ENIGH 2006; and 683 in the HBS 2006. Although some children in boarding schools and other institutions are likely to be out of the sample frame, those samples should otherwise be repre- sentative for the total population of 15 year-olds. In these four countries, these are the staple surveys for assessing the distribu- tion of household income and, in some cases, consumption expenditures. But they also collect information on other topics, including labor supply, education and migration. We use information on parents’ characteristics for estimating the total population of 15 year-olds in groups defined by similar gender, mother’s education and father’s occupation. The classification of the family background variable can be made comparable with the ones in the PISA by ap- propriate aggregation of coding categories. Parental characteristics are missing for orphans, children who do not live with their parents, or whose parents did not report their education. For instance, the information on mother’s education is missing for about 15.0% of 15 year-olds in the PNAD 2006, 8.7% in the SUSENAS 2005, 11.9% in the ENIGH 2006, and 3.8% in the HBS 2006. When comparing the two surveyed populations, children with missing parental background information in the household surveys are not dropped, but associ- ated with those with the same information missing in the PISA survey. II. MEASURING INEQUALITY IN E D U CAT I O N A L AC H I E V E M E N T Measures of inequality in educational achievement are based on distributions of standardized test scores ( yij ), constructed from the IRT-adjusted scores (xij ) by means of a transformation such as equation (2). In the case of PISA, the transformation is given by (2) exactly, with m ^ ¼ 100. That oper- ^ ¼ 500, and s ation involves both a translation of the original distribution (by the difference between the new arbitrary mean and the original mean, re-scaled) and a rescal- ing (by the ratio of the new to the original standard deviations). In the field of inequality measurement it is usual to impose axioms, or desir- able properties, that individual indices should respect. Three common such axioms are: (i) symmetry: which requires that the measure be insensitive to any permuta- tion of the y vector; (ii) continuity in any individual income; (iii) and the transfer principle: which requires that the measure should rise (strong axiom) or at least not fall (weak axiom) as a result of any sequence of mean-preserving spreads. In addition, inequality indices often satisfy either one of two invariance axioms: 222 THE WORLD BANK ECONOMIC REVIEW (iv-a): scale invariance: which requires that the index be insensitive to any re-scaling of the y vector: IðyÞ ¼ IðlyÞ; l . 0, where y is the vector of inter- est, and l is a positive scalar. (iv-b): translation invariance: which requires that the index be insensitive to a translation of the y vector: IðyÞ ¼ Iðy þ aÞ; a = 0, where a is a non-zero constant vector of the same dimension as y. An important result, due to Zheng (1994), is that no inequality index that sat- isfies axioms (i)-(iii)—known as “meaningful” inequality measures—satisfies both (iv-a) and (iv-b). This impossibility result, in other words, states that no meaningful inequality index can be both scale- and translation invariant. A direct implication of Zheng’s result for the measurement of inequality of educa- tional achievement using standardized data is stated below as our Remark 1: Remark 1: No meaningful inequality index yields a cardinally identical measure for the pre- and post-standardization distributions of the same test scores. Note that the remark derives from the standardization procedure (equation 2), rather than from the much more complex item response theory adjustments. It refers, therefore, to the measurement of inequality in IRT-adjusted test scores, and not to a comparison between adjusted and unadjusted scores. How important is Remark 1? Clearly this depends on whether or not in- equality indices applied to pre- and post-standardization distributions are ordi- nally equivalent—that is to say, whether they rank distributions in precisely the same way, regardless of cardinal differences in value. After all, standardization is just a change in metric. The ( post-standardization) mean score in each country j, for example is simply: ^ x s  my j ¼m ^þ mj À m ð3Þ s where mx j is the pre-standardization mean in country j, and other notation is as in equation (2). Since every other term in (3) is a constant, my x j and mj are ordi- nally equivalent. One is a monotonic (and in this case, affine) transformation of the other. Country ranks based on either would be identical. The only effect of standardization on country mean scores is a change in metric. Since this was the point of the process in the first place, there seems to be no cause for concern. The same is true for percentile-based measures of dispersion, such as the inter-quartile ratio, or the absolute difference P95-P5 used by Micklewright and Schnepf (2007) to compare dispersion across 21 countries and three differ- ent surveys. Equation (2) is itself a monotonic, and therefore rank-preserving, transformation. Since each score yi occupies precisely the same rank in its dis- tribution as the original score xi did in its distribution, rank- or percentile- Ferreira and Gignoux 223 based measures—be they ratios or differences, will be cardinally different, but ordinally equivalent. Yet this is not true of inequality measures in general. The post- y standardization Gini coefficient in country j (Gj ) for example, can be straight- forwardly shown to relate to the pre-standardization Gini (Gx j ) as follows: mx js^ x Gy j ¼ y Gj : ð4Þ mj s Unlike in equation (3), the terms multiplying Gx j are not all constants. In par- ticular, the post-standardization Gini is a function of the ratio of pre- to post- standardization means, which varies with mx j (see equation 3). The existence of a second argument in (4) implies that the post-standardization Gini coefficient is not ordinally equivalent to its pre-standardization analogue. Most other common meaningful inequality measures do not share the linear- ity of the Gini, so their post- and pre-standardization formulae cannot be related as straightforwardly. Nevertheless, substitution of equations (2) and (3) into the formulae for the Generalized Entropy or the Kolm-Atkinson classes of inequality measures yield expressions that are functions of both the central dis- tance indicators of the measure in question, and of the ratio of pre- to post- y standardization means (mx j =mj ). For the Generalized Entropy (GE) class, for example: " À Á!a x !a # 1 1X m ^ =s xij À m ^ þs mj GEy ¼ 2 À1 : ð5Þ j a À a nj iej x mj my j These results give rise to our second remark: Remark 2: A number of well-known inequality indices are not even ordinally equivalent when applied to pre- and post-standardization distributions. Ordinal equivalence with respect to standardization is clearly a desirable proper- ty for an index used for measuring inequality in educational achievement. The standardization operation given by (2) is meant merely to adjust an arbitrary metric. It is not intended to fundamentally alter our judgment of how countries compare with one another in substantive terms. Yet, when indices such as the Gini or Theil index are applied to these standardized distributions, we cannot be confident that the original rank in post-IRT adjusted inequality is preserved.9 What then are the options for those interested in the distribution of educa- tional achievement? One could, of course, rely on rank-based measures such as the inter-quartile range or percentile differences which, as noted above, are 9. Gamboa and Waltenberg (2012), for example, report Theil-L indices of post-standardized PISA test scores. 224 THE WORLD BANK ECONOMIC REVIEW ordinally equivalent. However, these measures do not satisfy the transfer prin- ciple: a progressive transfer (from above) to the income recipient on the 95th percentile will, for example, cause the p95-p05 measure to indicate an increase in inequality. And of course, because such indices are insensitive by construc- tion to any chances in incomes that do not affect those on the percentiles of reference, they also violate continuity. A possible alternative would be to use an absolute measure of inequality— such as the variance, or the absolute Gini coefficient10—which are ordinally in- variant in the standardization. The variance of a post-standardized distribution (Vyj ), for example, is a monotonic (linear) function of the pre-standardization variance (Vx j ), and does not depend on any other moment of the pre- standardization distribution:  2 y s ^ Vj ¼ Vjx : ð6Þ s The variance is seldom used as an inequality measure because it is scale- dependent: It increases with the mean. It also fails the transfer sensitivity axiom, by placing greater weight on transfers higher up the distribution than to those lower down. While these are not trivial concerns, it appears to us that in the context of distributions of educational achievement, they are less severe than violating either the transfer principle itself (like the percentile based mea- sures) or ordinal invariance in the standardization, which allows an apparently innocuous operation to fundamentally alter distributional rankings. The vari- ance (and the standard deviation, of course) is a meaningful measure of in- equality in the precise sense that it satisfies axioms (i)-(iii) above. The variance is also additively decomposable, and shares of the variance obtained from some such decompositions can be shown to be cardinally invariant to stand- ardization, as discussed in the next section.11 These properties will prove in- strumental in adapting an intuitive measure of inequality of opportunity to the context of education. It should be noted that ordinal invariance is not a mere theoretical curiosity. For the countries listed in Table 1, country rankings for inequality in achieve- ment in Mathematics differ considerably if one uses the variance (which pre- serves the original ordering) or the (relative) Gini coefficient (which does not). As an example, consider the positions of Mexico and Germany: Mexico is ranked the 15th most unequal country by the Gini, but only 44th most unequal by the variance. Germany has the 8th highest variance, but 22nd highest Gini.12 10. The absolute Gini coefficient, of course, is the standard (relative) Gini index scaled up by the mean. 11. An alternative ordinally invariant measure is the ratio of the within-country variance to the overall variance in the pooled sample of countries. This measure would also preserve cardinality. We are grateful to an anonymous referee for pointing this out. 12. The detailed comparison is available from the authors upon request. Ferreira and Gignoux 225 For these reasons, we adopt the variance and the standard deviation as our basic measures of inequality of educational achievement. Because users of this kind of data are generally more comfortable with the standard deviation than its square, this is the variable we report. Columns 3-11 in Table 1 present the mean and standard deviation (S.D.) of the standardized test scores in reading, math and science, in that order, for all 57 countries in the 2006 PISA surveys. The column immediately to the right of each S.D. column reports its bootstrapped standard error. Among the countries with higher inequality in math scores are Western European countries such as Austria, Belgium, France, Germany, and Italy; East European ones such as Czech Republic and Bulgaria, Latin American countries such as Argentina and Uruguay, but also Israel and Taiwan, China. Among the ones with lower inequality in achievements are other European coun- tries such as Croatia, Denmark, Estonia, Finland, Ireland, and Latvia, but also Asian countries such as Indonesia, Thailand and Jordan. Countries such as the UK, Japan, and the United States take intermediate rankings.13 Sample Selection Issues Although we have established that the country ranking that can be derived from Table 1 is ordinally equivalent to the pre-standardization ranking, the issue of PISA sample selection remains a potential problem. As noted in Section 2, coverage rates range from a low of 0.47 in Turkey, to 1.02 in Switzerland.14 Selection would not be a problem if one were interested exclu- sively in the performance of 15 year-olds that are in school, and within a rea- sonable range of their expected grade of attendance. But this is likely to be an excessively narrow prism through which to assess a country’s educational system and—even more so—to make international comparisons. Consider the example of two hypothetical “educational strategies,” illustrated by countries A and B, which have identical distributions of school and family characteris- tics, as well as of underlying ability in the population of 15 year-olds. Country A seeks to be inclusive, and allocates resources towards retaining as many stu- dents as possible in school, and towards promoting learning by those with the lowest demonstrated achievement. Country B, on the other hand, actively dis- courages enrollment by those with lower ability, and seeks to retain only the top half of performers in school by age 15. Looking only at the test scores for the samples of enrolled fifteen year-olds will naturally suggest that Country B has both a higher mean and a lower variance than country A, and thus a supe- rior educational system altogether. 13. The inequality measures obtained for Azerbaijan seem particularly small and place the country as an outlier in all the analyses. It is unclear how much of this is due to the data collection procedures in this country, but such a different pattern is not likely due to real differences only. 14. One presumes that coverage rates in excess of 1.00 must be due either to statistical discrepancies in the estimates of 15 year-olds in the total population, or to errors of inclusion in the sample of test-takers. 226 THE WORLD BANK ECONOMIC REVIEW This is not to suggest, of course, that Brazil, Indonesia, Mexico, Turkey, or any of the other countries with low coverage rates in Table 1 actively pursue an exclusionary strategy like that of hypothetical country B. But dropping out and lagging behind are, nevertheless, extremely likely to be selective processes, in the sense that they are correlated with family and student characteristics that also affect test scores. If one is interested in comparing the educational achieve- ment of the population of fifteen year-olds across countries, therefore, the PISA samples suffer from selection bias. Correcting for such biases is never simple, and even less so when non- participants are not observed at all in the sample (unlike, say, when seeking to correct for labor force participation on the basis of surveys that contain infor- mation on both earners and non-participants). While we do not offer a sample selection bias correction procedure for all countries in the PISA sample in this paper, we propose a simple two-sample non-parametric mechanism for assess- ing the sensitivity of our inequality measures to alternative assumptions about the sample selection process. Denote the (density of the) distribution of test scores y in a particular country j by fj ðyÞ. Consider a vector of covariates X that is observed both in the PISA sample and in an ancillary household survey, which is representative of the full population of 15 year-olds. Note that the density of test scores in the PISA sample can be written as: ððð ððð fj ðyÞ ¼ Fj ðy; XÞdX ¼ gj ðyjXÞfj ðXÞdX: ð7Þ In (7), F denotes the joint distribution of y and X, g denotes the conditional distribution of y on X, and f denotes the joint density of the covariates in the vector X.15 If the joint density of the observable covariates X in a particular survey for country j is written fj ðXjs ¼ surveyÞ, then our first proposed esti- mate for a test-score distribution (density) corrected for sample selection on ob- servables is given by: ððð fjSO ðyÞ ¼ gj ðyjXÞcj ðXÞfj ðXÞdX ð8Þ where fj ðXjs ¼ HHÞ cj ðXÞ ¼ : ð9Þ fj ðXjs ¼ PISAÞ 15. The triple integral notation is short-hand for integrating out every element of X, so that there are as many integrals as there are elements in the vector of covariates common to both surveys. As it happens, in our application that dimension is three. Ferreira and Gignoux 227 Equation (9) is simply the ratio of the density of fifteen year-olds whose ob- served characteristics X take certain values, in the ancillary household survey (HH), to the density of fifteen year-olds with the exact same observed charac- teristics in the PISA survey. cj ðXÞ is a re-weighting function exactly analogous to that used by DiNardo, Fortin and Lemieux (1996) to construct counterfac- tual income densities in their study of inequality in the US. Whereas DiNardo et al. use the ratio of densities across different years (of the same survey), we use the ratio of densities across different surveys (for the same year). To the extent that test-taking (i.e. being in the PISA sample) is correlated with ob- served covariates in X, the counterfactual distribution in (8) should correct for the corresponding selection bias.16 In practice, this procedure was implemented by partitioning both the PISA and the ancillary household survey into cells with identical values for three observed covariates: gender, mother’s edu- cation, and father’s occupation, with the latter two variables classified as in Section 2.17 The ratios of densities in each cell in these partitions were used to construct the reweighting function (Equation 9), and both the S.D. and the IOp measures were computed over the counterfactual density of scores given by (8). This procedure assumes that selection into the PISA sample is fully explained by observable variables, such as gender and family background. While such variables are likely to play a role in selection, it is also likely that other, unob- served variables do too. Within the set of girls with mothers with no formal ed- ucation and fathers who work in agriculture, for example, it is possible that a higher proportion of high-ability students than low-ability students stay in school long enough to enter the PISA sample. This kind of selection would imply that equation (8) may overstate the achievement of those students who are counterfactually “brought back into” the sample: Simple re-weighting ef- fectively assigns all those out-of-sample students the same scores obtained by students similar to them (in terms of the variables in X). If they are, in fact, likely to perform somewhat less well because of unobserved differences, the procedure overstates their true performance. By its very nature, of course, selection on unobservables is harder to account for. The ancillary household surveys used to construct the reweighting function do not contain information on test scores. To provide another sensitivity test for the possible magnitude of sample selection bias driven by unobservables, we consider the (rather extreme) assumption that all those students who are counterfactually “re-introduced” into the PISA sample by the above procedure— a proportion given by cj ðXÞ À 1, for each X—do no better than those who are actually in the sample. In practice, we ascribe to them the lowest observed score for their cell in the partition. 16. The superscript SO stands for selection on observables. 17. Surveys were thus partitioned into 24 cells. Given the sample sizes reported earlier, particularly for Turkey’s HBS and, to a lesser extent, Mexico’s ENIGH, it was not possible to further refine the partition by using additional covariates. 228 THE WORLD BANK ECONOMIC REVIEW In order to provide a sense of how sensitive our estimates of educational in- equality (reported in Table 1) might be to sample selection, Table 3 reports the results of both of the above scenarios for the four countries with the lowest PISA coverage ratios in Table 1. To economize on space, Table 3 reports the effects of these ‘selection correction’ procedures both on the standard deviation of test scores and on our measure of inequality of educational opportunity, which is in- troduced in the next section. The first three columns report these measures (and standard errors) for the uncorrected, original PISA sample, for reading, math and science respectively. The next three report estimates for the correction that assumes selection on observables only (equation 8), and the final three for the correction that assumes selection on unobservables (with no common support). The results in Table 3 provide a mixed message. Somewhat surprisingly, both inequality of achievement (measured by the standard deviation) and in- equality of opportunity seem to be quite robust to selection on observables, despite very low coverage rates (of approximately 50% in these four countries). While this is encouraging, the same cannot be said for the estimates for selec- tion on unobservables. Under these (admittedly extreme) assumptions, inequal- ity in achievement increases by between 44% in Turkey and 92% in Mexico. Inequality of educational opportunity also rises in all countries, except Mexico. It is possible to interpret these results as comforting, if one chooses to focus on the relative robustness of the measures to selection on observables, even in countries where PISA coverage is lowest. It seems likely that, if these observed variables account for most of the sample selection process, the estimates of ed- ucational inequality in Table 1 are robust for all countries. The fact that those estimates are sensitive to selection on unobservables can be minimized by the strength of the “no common support” assumption that assigns the very lowest grade in each cell to all those students counterfactually added to the sample. Yet, it would probably be wiser to interpret the results from Table 3 as pro- viding grounds for caution. We simply do not know how much selection into the PISA sample takes place on the basis of variables other than gender, mother’s education and father’s occupation. Until more is known about the composition of the group of fifteen year-olds that is excluded from the PISA sample, the possibility remains that inequality in countries with low coverage is underestimated. Investigation of that group of teenagers would seem like an important—but so far neglected—area of study for those interested in the dis- tribution of educational achievement, particularly in developing countries. III. A MEASURE OF INEQUALITY OF ED UCATIONA L OP PO RTUN ITY At least as important as the total level of inequality in educational achievement is the question of how much of that inequality is explained by pre-determined circumstances, which individuals simply inherit, rather than controlling. While many may find some inequality in achievement—that might reflect differences in effort, or perhaps even differences in innate ability—quite acceptable, it is Ferreira and Gignoux 229 T A B L E 3 . Inequality of Achievement and Opportunity in Low-Coverage Countries: Sensitivity to Different Assumptions on Selection into the PISA Sample PISA population without Correction assuming Correction assuming strong any correction selection on observables selection on unobservables Reading Math Science Reading Math Science Reading Math Science TURKEY Inequality (SD) 92.90 93.24 83.20 98.38 91.43 82.58 155.67 134.04 121.61 2.75 4.32 3.14 IOp 0.251 0.241 0.249 0.250 0.236 0.250 0.327 0.320 0.326 0.026 0.033 0.032 BRAZIL Inequality (SD) 102.46 92.02 89.28 102.86 90.44 86.75 179.82 146.68 146.17 3.34 2.65 1.93 IOp 0.268 0.318 0.286 0.265 0.309 0.262 0.404 0.404 0.385 0.020 0.005 0.021 MEXICO Inequality (SD) 95.68 85.27 80.70 95.63 85.02 79.18 196.85 162.79 136.99 2.27 2.16 1.47 IOp 0.278 0.261 0.271 0.267 0.242 0.255 0.256 0.250 0.228 0.024 0.002 0.024 INDONESIA Inequality (SD) 74.79 80.01 70.06 71.03 76.27 65.74 130.56 135.89 112.79 2.39 3.18 3.26 IOp 0.250 0.237 0.220 0.218 0.200 0.181 0.274 0.261 0.261 0.038 0.042 0.045 Notes: IOp denotes the measure of inequality of educational opportunity, defined in equation (13). It is the share of the total variance in test scores which is accounted for by the student’s pre- determined circumstance variables. Source: Authors’ analysis based on data described in the text. common to come across arguments against unequal opportunities among stu- dents. These are differences in achievement that do not reflect the choices or actions of today’s students, but only inherited circumstances beyond their control. That such inequalities are morally objectionable is today a dominant view among social justice theorists. See, for example, Cohen (1989), Dworkin (1981), Roemer (1998) and Fleurbaey (2008) for some of the classic references. There is also a positive argument against the inheritance of educational in- equality, namely that if scarce opportunities for educational investment are al- located on some basis other than talent – such as inherited wealth, for example – this will lead to an inefficient allocation of resources.18 The applied literature on the measurement of inequality of opportunity has focused primarily on opportunities for the acquisition of income, but there is ´ ndez and Galı 18. See, e.g. Ferna ´ (1999). 230 THE WORLD BANK ECONOMIC REVIEW no reason it cannot be adapted to the space of educational achievement.19 Two main approaches characterize that empirical literature. Both approaches begin by seeking agreement on a set of individual characteristics which are beyond the individual’s control, and for which he or she cannot be held responsible. These variables are known as ‘circumstances’. Once a vector C of circumstanc- es has been agreed upon, society can be partitioned into groups with identical circumstances. Formally, such a partition is given by a set of types: P ¼ fT1 ; T2 ; . . . ; TK g, such that T1 < T2 < Á Á Á < TK ¼ f1; . . . ; Ng, Tl > Tk ¼ ;; 8l; k, and the vectors Ci ¼ Cj ; 8i; jji [ Tk ; j [ Tk ; 8k: Given such a partition, the two approaches differ in how they define the benchmark of equality of opportunity. In the ex-ante approach, associated with van de Gaer (1993), the opportunity set faced by each type is evaluated, and equality of opportunity is attained when there is perfect equality in those values across all types. In practice, researchers have often used the mean income (or achievement) of the type as an estimate of the value of the opportu- nity set they face. Since equality of opportunity would imply equality in means across types, inequality of opportunity is then naturally seen as some measure of between-type inequality. In the ex-post approach, associated with Roemer (1998), equality of oppor- tunity obtains only when individuals exerting the same degree of effort, regard- less of their circumstances, receive the same reward. Under certain assumptions, this amounts to requiring equality in the full conditional outcome distributions across all types. Inequality of opportunity would, in this case, best be captured by the (appropriately weighted) sum of inequality within groups characterized by the same degree of effort.20 The two approaches are closely related but, for any society with a given joint distribution of achieve- ment and circumstance variables, they yield different answers to the question “How much inequality of opportunity is there?” See Fleurbaey and Peragine (2012) for a formal discussion of the relationship between the two approaches. In what follows, we adapt the ex-ante approach employed by Ferreira and Gignoux (2011a) to the distributions of test scores described earlier.21 These authors propose to measure inequality of opportunity (IOp) by between-type inequality. Specifically: ÀÈ ÉÁ I mki uIOp ¼ ð10Þ IðyÞ 19. Indeed Checchi and Peragine (2005), the working paper version of their 2010 paper, do apply the concept to educational achievement measures. See also Gamboa and Waltenberg (2012) for a more recent treatment. 20. Under the standard Roemerian assumptions, these groups are Checchi and Peragine’s (2010) ‘tranches’. 21. Ferreira and Gignoux (2011a), in turn, build on Bourguignon et al. (2007) and Checchi and Peragine (2010). Ferreira and Gignoux 231 È É where mk i is the smoothed distribution corresponding to the distribution y and the partition P.22 Naturally, uIOp can be computed non-parametrically by means of a standard between-group inequality decomposition ( provided the chosen inequality index I() is properly decomposable). However, this procedure is data-intensive when the vector C is large. As the partition becomes finer, cells become small and sparsely populated, and the precision of the estimates of cell means declines, giving rise to an upwards bias in the estimation of uIOp . Following Bourguignon et al. (2007), Ferreira and Gignoux (2011a) then propose a para- metric alternative for uIOp , based on an OLS regression of y on C: À Á ^ I C0i b ^ uIOp ¼ : ð11Þ IðyÞ ^ in (11) is the OLS estimate of the regression coefficients in a simple regres- b sion of y on C: yi ¼ C0i b þ hi : ð12Þ ^ denotes the vector of predicted test scores from regression (12). In (11), C0i b Under the maintained assumption of a linear relationship between achievement and circumstances, this vector is equivalent to the smoothed distribution, since all individuals with identical circumstances are assigned their conditional mean scores. Because of its unique path-independent decomposability properties, Checchi and Peragine (2010) and Ferreira and Gignoux (2011a) both use the mean log- arithmic deviation as the inequality index I(). However, as shown above, the mean log deviation is not ordinally invariant in the standardization to which test scores are submitted, and it is therefore unsuitable for use in the present context. Following the discussion in Section 3, we use the simple variance as our inequality index I(). This choice yields our proposed measure of inequality of educational opportunity, as a special case of (11): À Á ^ Var C0i b ^IOp ¼ u : ð13Þ Varðyi Þ This index has a number of attractive features. First, it is extremely simple to calculate: It is simply the R 2 of an OLS regression of the child’s test score on a vector C of individual circumstances. In our application to the PISA data sets, C includes the following ten variables: gender, father’s and mother’s education, father’s occupation, language spoken at home, migration status, access to 22. A smoothed distribution is obtained from a vector y and a partition P by replacing each element of y in a given cell Tk with the mean value of y in its cell, mk. See Foster and Shneyerov (2000). 232 THE WORLD BANK ECONOMIC REVIEW books at home, durables owned by the households, cultural items owned, and the location of the school attended. Second, despite its simplicity, it is a very meaningful summary statistic. It is a parametric approximation to the lower bound on the share of overall in- equality in educational achievement that is explained by pre-determined cir- cumstances. A formal proof is provided by Ferreira and Gignoux (2011). But the basic intuition is to note that (12) can be seen as the reduced form of a (lin- earized version of a) model such as: y ¼ f ðC; E; uÞ ð14Þ E ¼ gðC; vÞ: ð15Þ In (14) and (15), y denotes achievement, and C denotes the vector of circum- stances, as before. E denotes a vector of efforts: all variables that affect achieve- ment and over which individuals do have some measure of control. u and v denote random shocks or innovations. Because 15 year-olds may conceivably affect the choice of school they attend, the class they are assigned to, and thus the teachers they interact with, all school characteristic variables, for example, are in- cluded in E. So are any direct measures of the student’s own efforts in preparing for exams, for instance. Of course, efforts E can be influenced by circumstances C, but the reverse cannot happen. Variables can only be treated as circumstances if they are pre-determined and entirely exogenous to the individual. Now return to (12) as a linearized reduced form of (14)-(15). We know that circumstances C are economically exogenous to y. We also know that all effort (E) variables (whether or not one could observe them in the data) are omitted deliberately: b is intended to capture the reduced-form effect of circumstances – both directly and through efforts. Since all relevant factors are classified into either circumstances or efforts, the only sources of bias to the estimates of b are omitted, unobserved circumstance variables. Although the observed vector C is economically exogenous, it may not be exogenous in the (econometric) sense that its components may be correlated with other (unobserved and thus omitted) circumstance variables. Individual elements of the vector b ^ suffer from these omitted variable biases, and cannot be interpreted as causal esti- mates of the individual impact of a particular circumstance on test scores. If one is interested, however, on the total joint effect of all circumstances on achievement and, more specifically, on the share of variation in y that is causally explained by the overall effect of circumstances (operating both directly and through efforts), then the R 2 of (12) - our u ^IOp - yields a valid lower bound for the object of interest. By construction, the only missing variables in (12) are other circumstances. If any were added, u ^IOp might rise, but it cannot fall. While individual coefficients in b ^ may be biased, u^IOp is a lower bound estimate of the joint causal effect of all circumstances on achievement, and thus an appropriate Ferreira and Gignoux 233 measure of inequality of opportunity. A formal proof is provided by Ferreira and Gignoux (2011a), for the perfectly analogous case of incomes.23 A third attractive feature of (13) is that it allows for the use of more infor- mation on circumstances than previous studies, which typically rely on a smaller set of background variables, and thus capture a more limited share of heterogeneity in family resources. Schultz, Ursprung and Wossmann (2008), for example, focus on the number of books at home. Macdonald et al. (2010) look at the effect of gender and an index of household wealth but ignore, for example, information on parental education and occupation. Gamboa and Waltenberg (2012) see inequality of opportunity as determined by gender, pa- rental education, and school type ( public or private), which they treat as a cir- cumstance. We consider the joint effect of all of these circumstances, and more. ^IOp as a measure of inequality of educational A fourth attractive feature of u opportunity is that, unlike any measure of the level of inequality (see Remark 1 above), it is a parametric estimator of a ratio (equation 10) that is cardinally in- variant in the standardization of test scores. To see this, note that any sub- group mean is affected by standardization in a manner analogous to equation (3), so that:  2 n o s ^ Varfmk i ðyÞg ¼ Var mk i ðxÞ : ð16Þ s n o Given (16) and equation (6), it follows that uIOp ¼ Var mki ðyÞ =VarðyÞ ¼ n o Var mk i ðxÞ =VarðxÞ. A fifth attractive feature of this IOp measure is that it is neatly decompos- able into components for each individual variable in the vector C. Equation (13) can be rewritten as: " # À1 X 1 XX À Á ^IOp ¼ ðvar y Þ u b2 j var Cj þ b b cov Ck ; Cj : ð17Þ j 2 k j k j This in turn can be written as the sum over all elements (denoted by j ) of the C vector: " # X X À1 1X À Á ^IOp ¼ ^j ¼ ðvar yÞ b2 u u j var Cj þ b b cov Ck ; Cj : ð18Þ j j 2 k k j 23. Note, however, the implication that the cross-country comparisons reported in this paper are comparisons of that lower-bound measure. If additional circumstance variables were observed across all of these countries, those rankings might change. 234 THE WORLD BANK ECONOMIC REVIEW This decomposition is an example of a Shapley-Shorrocks decomposition: it corresponds to the average between two alternative paths for estimating the contribution of a particular circumstance CJ to the overall variance. In the first (direct) path, all Cj, j = J are held constant. In the second (residual) path, CJ is itself held constant, and its contribution is taken as the difference between the total variance and the ensuing variance. Either path is conceptually valid, and the Shapley-Shorrocks averaging procedure yields (18) as the path-independent additive decomposition.24 ^IOp can be seen as isomorphic to a measure of intergenerational per- Finally, u sistence of inequality, itself the converse of a measure of educational mobility.25 In the canonical Galton regression of a child’s outcome ( yit) on the parent’s outcome ( yi,t-1): yit ¼ byi;tÀ1 þ 1it ð19Þ the coefficient b is sometimes used as measure of persistence, and 1-b as a measure of mobility. An alternative that gives equal weight to the variance in both father’s and son’s distributions is the R 2 of (19) which is, of course, also the square of the correlation coefficient between the two outcomes in the popu- lation. If one were to replace the parent’s outcome yi,t-1 with a vector of parental or family background variables, (19) would transform into something very close to (12), and the R 2 measure of immobility into our measure of inequality of op- portunity, u^IOp . Indeed, the only pre-determined circumstance among the ten variables previously listed which is not a family background variable is the child’s own gender. Apart from the child’s own gender, one could see u ^IOp as a measure of intergenerational persistence, or immobility, in which the missing value for the parent’s own test scores, yi,t-1, is replaced with a proxy vector of family background circumstances, Ci. Having separately regressed test scores for each subject (in each country) on the vector C (equation 12), and computed the R 2 of each regression to obtain ^IOp , we report them on Table 4. These are our estimates of the inequality of u educational opportunity (IOp) given by equation (13). They range between 0 and 1, and can be interpreted straight-forwardly as a lower-bound on the share of the total variance in educational achievement that is accounted for by 24. See Shorrocks (1999) for the original application of the Shapley value to distributional decompositions. Ferreira et al. (2011) provide a formal proof that (18) is the Shapley-Shorrocks decomposition of the variance into the effects of individual circumstances. 25. Mobility is a multifaceted concept, and there are many distinct measures of it, often attempting to capture different aspects of “movement” across distributions. See Fields and Ok (1996) for a discussion. In the present context, we adopt a view of mobility as time- or origin-independence. See also Shorrocks (1978). Persistence would therefore correspond to the concept of origin-dependence, which is closely related to the notions of inequality of opportunity in both van de Gaer (1993) and Roemer (1998). Ferreira and Gignoux 235 T A B L E 4 . Inequality of Educational Opportunity for Three PISA Subjects IOp Standard Error IOp Standard Error IOp Standard Error Reading (Reading IOp) Mathematics (Math IOp) Science (Science IOp) Asia & North Africa Azerbaijan 0.173 0.028 0.044 0.012 0.112 0.024 Hong Kong 0.177 0.016 0.154 0.016 0.166 0.018 SAR, China Indonesia 0.250 0.038 0.237 0.042 0.220 0.045 Israel 0.197 0.018 0.206 0.019 0.195 0.016 Japan 0.206 0.017 0.203 0.020 0.189 0.016 Jordan 0.346 0.024 0.272 0.024 0.271 0.019 Korea 0.214 0.022 0.209 0.021 0.173 0.019 Kyrgyzstan 0.314 0.023 0.306 0.027 0.269 0.023 Macao-China 0.127 0.012 0.102 0.009 0.111 0.008 Qatar 0.309 0.010 0.254 0.009 0.264 0.009 Russian 0.238 0.021 0.165 0.020 0.183 0.020 Federation Chinese Taipei 0.300 0.017 0.275 0.022 0.281 0.019 Thailand 0.325 0.023 0.230 0.021 0.265 0.022 Tunisia 0.215 0.024 0.273 0.031 0.191 0.026 Turkey 0.251 0.026 0.241 0.033 0.249 0.032 Latin America Argentina 0.289 0.024 0.315 0.007 0.312 0.026 Brazil 0.268 0.020 0.318 0.005 0.286 0.021 Chile 0.248 0.022 0.330 0.001 0.299 0.021 Colombia 0.181 0.018 0.216 0.007 0.193 0.018 Mexico 0.278 0.024 0.261 0.002 0.271 0.024 Uruguay 0.221 0.015 0.245 0.004 0.248 0.012 Australia 0.199 0.010 0.153 0.009 0.164 0.009 Canada 0.242 0.011 0.211 0.011 0.207 0.010 New Zealand 0.276 0.013 0.241 0.012 0.269 0.013 United States 0.279 0.020 0.282 0.019 Eastern Europe Bulgaria 0.377 0.028 0.331 0.030 0.364 0.030 Czech Republic 0.296 0.021 0.268 0.019 0.279 0.020 Estonia 0.271 0.013 0.206 0.013 0.208 0.012 Croatia 0.297 0.017 0.222 0.015 0.239 0.014 Hungary 0.345 0.023 0.326 0.022 0.326 0.019 Lithuania 0.318 0.017 0.279 0.017 0.262 0.016 Latvia 0.254 0.017 0.201 0.020 0.187 0.016 Montenegro 0.252 0.013 0.223 0.012 0.197 0.011 Poland 0.275 0.014 0.241 0.013 0.241 0.014 Romania 0.301 0.026 0.313 0.028 0.310 0.027 Serbia 0.311 0.018 0.276 0.017 0.255 0.016 Slovak Republic 0.292 0.026 0.317 0.030 0.297 0.024 Slovenia 0.336 0.018 0.263 0.016 0.268 0.014 Western Europe Austria 0.296 0.019 0.300 0.020 0.324 0.022 Belgium 0.335 0.015 0.329 0.018 0.338 0.015 Switzerland 0.313 0.013 0.282 0.013 0.322 0.012 Germany 0.368 0.021 0.351 0.018 0.352 0.019 (continued ) 236 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Continued IOp Standard Error IOp Standard Error IOp Standard Error Reading (Reading IOp) Mathematics (Math IOp) Science (Science IOp) Denmark 0.229 0.015 0.219 0.014 0.249 0.017 Spain 0.243 0.013 0.239 0.012 0.258 0.013 Finland 0.247 0.014 0.179 0.010 0.167 0.011 France 0.305 0.019 0.335 0.019 0.345 0.018 United Kingdom 0.274 0.014 0.258 0.012 0.275 0.012 Greece 0.261 0.023 0.228 0.022 0.245 0.019 Ireland 0.259 0.018 0.235 0.017 0.240 0.016 Iceland 0.234 0.009 0.167 0.009 0.184 0.009 Italy 0.207 0.015 0.178 0.014 0.206 0.014 Liechtenstein 0.388 0.031 0.323 0.034 0.379 0.030 Luxembourg 0.344 0.008 0.291 0.008 0.328 0.009 Netherlands 0.247 0.022 0.271 0.023 0.283 0.023 Norway 0.271 0.016 0.195 0.014 0.220 0.018 Portugal 0.303 0.021 0.274 0.019 0.267 0.020 Sweden 0.265 0.014 0.233 0.012 0.250 0.013 Notes: IOp denotes the measure of inequality of educational opportunity, defined in equation (13). It is the share of the total variance in test scores which is accounted for by the student’s pre- determined circumstance variables. Source: Authors’ analysis based on data from PISA 2006. pre-determined circumstances (gender and family background) in each country. Bootstrapped standard errors are reported next to each IOp measure. The IOp estimates range between 12.7% and 38.8% of the total variance of test scores in reading; between 4.4% (10.2% excluding the outlier Azerbaijan) and 35.1% of the variance of test scores in math; and between 11.1% and 37.9% in Science.26 No clear regional pattern emerges from the estimates presented in Table 4. Among the countries with the highest levels of inequality of opportunity, with shares above 30%, are Western European countries (such as Belgium, France, and Germany) but also Eastern European countries (such as Bulgaria and Hungary), and Latin American countries (such as Argentina, Brazil and Chile). Among the countries with the lowest IOp, with shares below 20%, are Asian countries (such as Azerbaijan, Macao (China), and Hong Kong SAR, China), Nordic countries (such as Finland, Iceland, and Norway), Russia, Australia and Italy. The United States, the UK, and Spain lie in an intermediate range, with shares close to 25%. One can use these results to make specific comparisons. For example, the degree of inequality of educational opportunity seems to be significantly higher in a few large European countries, such as France and Germany, than in the 26. If one were interpreting these shares as proxies for the persistence measure given by the R2 of (19), one should note that the numbers correspond to squares of the correlation coefficient. The square root of IOp for mathematics scores, for example, ranges from 0.21 to 0.59. Ferreira and Gignoux 237 United States. However these inequalities are significantly lower in Nordic countries, such as Finland and Norway, or in Japan and Korea. Regarding de- veloping economies, countries in Latin America tend to rank in the upper half of the distribution, while Asian countries, such as Indonesia and Thailand, rank in the lower half. Although the estimates are very imprecise for Indonesia, Thailand exhibits significantly lower inequalities than Latin American countries such as Brazil. The results for reading and science are not discussed in detail here, but IOp measures for the three subjects are highly correlated: The Spearman rank correlation coefficients for shares in Reading, Math and Science range from 0.75 to 0.92. The absence of a clear geographical pattern in the cross-country distribution of inequality of educational opportunity is mirrored in the absence of a correla- tion between IOp and either the level of educational achievement, as measured by mean test scores, or the level of economic development, as measured by GDP per capita.27 Simple regressions of IOp in mathematics against both mean achievement in mathematics and GDP per capita yield insignificant coefficients. Scatter plots and some additional robustness analysis are presented in the working paper version of this article, Ferreira and Gignoux (2011b). Ferreira and Gignoux (2011b) also report the exact decomposition of inequal- ity of opportunity into partial shares by individual circumstance, described in equation (18). These partial shares are functions of individual regression coeffi- cients from (12) which, as noted earlier, are likely to be biased. The partial shares reported in our working paper version should therefore not be interpreted causally in any way. They are useful only as a description of the variables under- pinning the overall (lower-bound) measure of inequality of opportunity. I V. A D E S C R I P T I V E A P P L I C A T I O N : C O R R E L A T I O N S BETWEEN IOP AND EDU CATION POLICIES As an illustration of potential applications, we now briefly investigate the cross- country correlation between the measure of inequality of educational opportuni- ty presented in the previous section and two specific educational policy variables: the distribution of public spending across different levels of the education system, and the extent of early tracking of pupils between general and vocational schools or classes. The incidence of public spending in education and the allocation of financial resources among the different segments of the education system have been ex- amined by various studies (e.g. Birdsall 1996; Castro-Leal et al. 1999; and Van de Walle and Nead 1995). Given that children with disadvantaged back- grounds tend to drop out from school earlier than others, the allocation of 27. GDP per capita is measured at purchasing power parity exchange rates, in 2006 US prices; the data are from the World Development Indicators (WDI) database. 238 THE WORLD BANK ECONOMIC REVIEW resources to the primary level of schooling is generally thought more likely to be progressive. The impacts of tracking policies on the efficiency and equity of educational systems are another example of education policies that have received considerable attention in recent studies (e.g. Ariga et al. 2006; Bertocchi and Spagat 2004; Brunello and Checchi 2007; Brunello et al. 2006; Hanushek and Woessman 2006; Jabukowski et al. 2010; Manning and Pisckhe 2006; Pekkarinen et al. 2009). Theory does not provide clear-cut predictions for the effect of early track- ing on educational achievements. On the one hand homogenous classrooms, and the associated specialization of teaching and curricula to the needs and abilities of specific students, could lead to efficiency gains. But on the other hand, disad- vantaged groups might be harmed by unfavorable allocations of resources, in- cluding less well endowed schools, teacher sorting, peer effects, or differences in curricula28. Moreover, since much of the early inequality in achievement—and thus the track placements themselves—are driven by differences in parental re- sources, a frequent concern has been that tracking might reinforce the effects of family background on educational achievements. I.e. that it might reduce inter- generational mobility, and exacerbate inequality of educational opportunity. We briefly examine the correlation between our measure of IOp and these two policies, using data on the policy indicators from the UNESCO Institute for Statistics (UIS).29 Our indicator of the distribution of educational expendi- tures is the share of spending in primary schools—defined as the first ISCED level, corresponding to grades 1 to 6—in total public educational expenditure. The indicator of tracking is the share of technical or vocational enrollment at the secondary level (including lower and upper secondary or the second and third ISCED levels, usually corresponding to grades 7 to 12) in total enrollment at that level.30 The information on the distribution of education expenditure across levels is missing for six countries (Canada, Montenegro, Qatar, Russia, Serbia and Taiwan, China) and the information on the share of technical and vocational enrollment at the secondary level is missing in five countries (Latvia; Montenegro; Serbia; Taiwan, China; and the United States). Two other coun- tries are excluded from the analysis: Liechtenstein and Luxembourg. The number of observations for Liechtenstein (339 examinees) makes the estimates of learning inequalities unreliable and Luxemburg is too much of an outlier in 28. Early tracking may also be costly in terms of the misallocation of students to tracks, and in terms of forgone versatility in the production of skills (Brunello and Checchi, 2007). 29. The data for 2006 correspond to the school year 2005-06 for countries where the school year laps over two calendar years. 30. As rightly noted by an anonymous referee, a preferable measure of tracking would focus exclusively on the proportion of students in vocational or technical tracks in lower (as opposed to both lower and upper) secondary education. Unfortunately, information on tracking at that more specific level is considerably scarcer: Of the 57 countries in PISA 2006 with information on enrollment at the lower secondary level in the UIS policy indicators data, 36 do not report enrollment in vocational or technical streams at that level. Ferreira and Gignoux 239 terms of GDP per capita in 2006 (at about 69.000 US dollars, with the US in second place at 44.000 US dollars). There is considerable variation in the share of expenditures allocated to the primary level of education in the remaining country sample. While the mean share is 27.0%, the lowest share is observed in Romania at 13.8% and the highest in Jordan at 41.7% (the first quartile is at 20.2% and the third quartile at 34.0%). Figure 1 provides an illustration of the relationship between the primary share of expenditures and IOp. The regression line and a 95% confi- dence interval for the mean are shown. Table 5 gives the tests of significance of F I G U R E 1. Inequality of Educational Opportunity and Public Expenditure at the Primary Level Source: Authors’ analysis based on data from PISA 2006. T A B L E 5 . Coefficients on the Primary Share of Public Education Expenditure in Regressions of IOp on that Variable; with and without Controls Reading Math Science No controls All countries 2 0.00217*** (0.00092) 2 0.00077 (0.00112) 2 0.00152 (0.00105) Excluding outliers 2 0.00300*** (0.00078) 2 0.00113 (0.00101) 2 0.00172* (0.00101) Controlling for GDP and public expenditure in education per pupil All countries 2 0.00197** (0.00087) 2 0.00013 (0.00120) 2 0.00103 (0.00113) Excluding outliers 2 0.00184*** (0.00072) 2 0.00181* (0.00102) 2 0.00185* (0.00108) Notes: Regression coefficients of the share of public expenditure in education allocated to the primary level. Dependent variable: IOp in the subject at column header. Standard errors in parenthe- ses. Where indicated, outliers are identified using the method proposed by Besley, Kuh and Welsch (1980). Data source: UNESCO Institute for Statistics database; ***/**/*: significant at 1/5/10%. 240 THE WORLD BANK ECONOMIC REVIEW this relationship both without any controls (first panel) and controlling for per capita GDP and public education expenditure per pupil (second panel). Once outliers are excluded, significant negative correlations exist both for reading and science, with or without controls. For math, the negative correlation is only significant with controls. The coefficients lie between -0.001 and -0.003, indicating that an increase of 10 points in the share of resources allocated to primary schooling is associated with decreases of 1 to 3 points in inequality of educational opportunity. There is also considerable heterogeneity in tracking in our country sample. The mean share is 20.8 percent and values range from 0.9% in Qatar to 51.4% percent in the Netherlands (the first quartile is at 12.9 and the third at 31.2). Figure 2 provides a scatter plot of the relationship between tracking and IOp in this sample, while Table 6 lists coefficients and standard errors, both without any controls (upper panel) and controlling for per capita GDP and public education expenditure per pupil (bottom panel). There is a clear pattern of significant positive relationships across all three subjects and both regression specifications, with the statistical significance being stronger in the specification with controls. Higher inequality of opportunity tends to be associated with higher shares of technical and vocational enrollment. The regression coeffi- cients lie between 0.001 and 0.002, indicating that an increase of 10 points of the share of technical or vocational enrollments is associated with an increase of 1 to 2 points in inequality of opportunity. F I G U R E 2. Inequality of Educational Opportunity and Tracking Note: Tracking is measured as the share of enrollment in technical or vocational curricula at the secondary level. Source: Authors’ analysis based on data from PISA 2006. Ferreira and Gignoux 241 T A B L E 6 . Coefficients on Tracking in Regressions of IOp on that Variable; with and without Controls Reading Math Science No controls All countries 0.00106* (0.00059) 0.00130* (0.00070) 0.00179*** (0.00063) Excluding outliers 0.00158** (0.00060) 0.00109* (0.00062) 0.00160*** (0.00059) Controlling for GDP and public expenditure in education per pupil All countries 0.00148*** (0.00057) 0.00173*** (0.00074) 0.00214*** (0.00068) Excluding outliers 0.00090* (0.00047) 0.00175*** (0.00065) 0.00205*** (0.00067) Notes: Regression coefficients of tracking (measured as the share of technical and vocational enrollment at the secondary level). Dependent variable: IOp in the subject at column header. Standard errors in parentheses. Where indicated, outliers are identified using the method proposed by Besley, Kuh and Welsch (1980). Source: UNESCO Institute for Statistics database; ***/**/*: significant at 1/5/10%. These correlations suggest that our measure of inequality of opportunity is negatively associated with the share of public spending on primary education, and positively associated with tracking into general or technical/vocational schooling at the secondary level. These associations allow for absolutely no in- ference of causality, of course, but the results seem in line with and extend those of studies devoted to these relationships. For instance, while Hanushek and Woessman (2006) find tracking to be associated with higher levels of overall inequality in test scores, our results suggest it also tends to come with higher levels of inequality of learning opportunities. The analysis here is de- scriptive and is only meant to illustrate the potential use of indicators of in- equality of opportunity for future studies of the distributive impacts of education policies. Future extensions—notably involving the use of panel data— might allow for causal analysis of these relationships. V. C O N C L U S I O N S Internationally comparable information on learning outcomes, such as the stan- dardized test scores collected by PISA surveys, represents a revolution in the quality of data available for research on education. It allows for potentially much greater insight into the determinants of educational achievement, and might therefore contribute to the design of policies that raise average learning levels, or that reduce educational disparities. The measurement of educational disparities using this kind of data is not, however, a trivial extension of inequality measurement in years of schooling, or in other variables like income. This paper has highlighted two issues that require special attention in the measurement of inequality in educational achievement, and which appear to have been overlooked so far. The first is the standardization of test scores, to which all meaningful measures of inequality are cardinally sen- sitive. More importantly, many common measures of inequality, including the 242 THE WORLD BANK ECONOMIC REVIEW Gini coefficient and the Theil indices, are not even ordinally invariant to stand- ardization, invalidating country rankings that are based on them. We show that the simple variance (or the standard deviation) of test scores is ordinally invariant to standardization, and present estimates for all 57 coun- tries that took part in the 2006 round of PISA surveys, in all three subjects for which tests are carried out: reading, mathematics and science. There is consid- erable international variation in educational inequality thus measured. The standard deviation in Math scores ranges from around 80 in Indonesia, Estonia and Finland, to nearly 110 in Belgium and Israel. The second measurement issue that may compromise international inequali- ty comparisons based on PISA test scores is the possibility of sample selection. The surveys are designed to be representative of the population of 15 year-olds enrolled in school, and attending grades 7 or above. While this stipulation covers most of the population of that age group in OECD countries, it purpo- sively excludes substantial numbers in poorer countries. Selection into the sample is clearly correlated with determinants of test scores, leading to a classic problem of sample selection bias. Using information on characteristics of fifteen year-olds included in other, ancillary household surveys, we use sample re-weighting methods to assess the implications of the selection bias for our measures of educational inequality in achievement and opportunity. Results for Brazil, Indonesia, Mexico and Turkey suggest that the inequality measures are relatively robust to selection on the basis of three observed vari- ables (gender, mother’s education and father’s occupation). Under a more strin- gent scenario of strong selection on unobservables with no common support, however, the current measures of educational inequality in these countries would appear to be substantially underestimated. Finally, we also propose and compute a measure of inequality in educational opportunity. The measure is simply the share of the total variance in achieve- ment that can be accounted for by pre-determined circumstance variables in a linear regression. The index is simple and intuitive, and provides a lower-bound estimate of the joint causal effect of all pre-determined circumstances on educa- tional inequality. It is cardinally invariant to the standardization of test scores, and closely related to the origin-independence concept of inter-generational ed- ucational mobility. Thus measured, inequality of opportunity in our sample of countries ranges from approximately 0.10 – 0.16 in Macao (China), Australia, and Hong Kong SAR, China, to 0.33 – 0.35 in Bulgaria, France and Germany. Although the measure is uncorrelated with average educational achievement and with GDP per capita, it appears to be higher in Latin America and parts of continental Europe (including France, Germany and Belgium). It is lower in Asia, the Nordic countries, and Australia. It is negatively correlated with the share of public educational spending allocated to primary schooling, and positively cor- related with the extent of educational tracking, defined as the share of technical and/or vocational enrollment in secondary schools. Ferreira and Gignoux 243 This paper sought to place the measurement of educational inequality on a sounder footing, given the specific characteristics of data on educational achievement. Some of our findings, however, raise new questions that may mo- tivate future work. First, although we confine our work to PISA surveys, a number of the issues we addressed are also relevant for the TIMSS, PIRLS and IALS surveys, all of which provide standardized achievement measures, and many of which also provide family background variables.31 Since those surveys have different country coverage, applying these measures to those surveys may shed light on educational inequality in other parts of the world.32 A second issue where additional research is clearly necessary is that of sample selection. The results presented in our Table 3 suggest that while inequality mea- sures appear to be robust to selection on a small set of observables, there is scope for very large selection biases on other variables (which we are treating as unobservable). This serves as a cautionary remark, but also as a call for addi- tional work on selection, either by closer examination of the sub-sample of fifteen year-olds who are not enrolled in school, or by combining and comparing different achievement surveys with distinct sampling frames, such as IALS (which samples households) and PISA (which samples schools). This might be done for countries which participate in both surveys, such as Chile.33 Finally, the measures of inequality of achievement and opportunity proposed here could be used as dependent variables in the evaluation of different educa- tional policy interventions, such as teacher training or school governance reforms. This would require achievement surveys that contain the right vari- ables, and are representative at much more finely disaggregated spatial units, likely in combination with interventions that were randomized at the level of those spatial units. But this is not as difficult to achieve as it may appear: It is close in spirit, in fact, to the approach applied by van de Gaer et al. (2012) to the impact of Mexico’s Oportunidades program on the distribution of health opportunities. To the extent that the effect of policy interventions on inequality of educational opportunity (or achievement) is a question of policy interest, the measurement tools developed in this paper may be useful. REFERENCES Ariga, K., G. Brunello, R. Iwahashi, and L. Rocco. 2006. “On the Efficiency Costs of De-Tracking Secondary Schools.” IZA Discussion Paper No. 2534. Institute for the Study of Labor, Bonn, Germany. Baker, F.B. 2001. The Basics of Item Response Theory. College Park, MD: ERIC Clearinghouse on Assessment and Evaluation, University of Maryland. 31. Although it is important to recognize that the sample selection issues in the TIMSS and PIRLS surveys, which define their samples by grade, rather than age, are somewhat different. See Jakubowski and Pokropek (2011) for an attempt to address those issues. 32. Indeed, see Salehi-Isfahani and Belhaj Hassine (2012) for an application of our approach to TIMSS data for countries in the Middle East and North Africa. 33. We are grateful to an anonymous referee for this suggestion. 244 THE WORLD BANK ECONOMIC REVIEW Bedard, K., and C. Ferrall. 2003. “Wage and Test Score Dispersion: Some International Evidence.” Economics of Education Review 22: 31– 43. Bertocchi, G., and M. Spagat. 2004. “The Evolution of Modern Educational Systems: Technical vs. General Education, Distributional Conflict, and Growth.” Journal of Development Economics 73 (2):559–582. Besley, D.A., E. Kuh, and R.E. Welsch. 1980. Regression Diagnostics: Identifying Influential Data and Sources of Colinearity. New York: Wiley. Birdsall, N. 1996. “Public Spending on Higher Education in Developing Countries: Too Much or Too Little?” Economics of Education Review 15 (4): 407–19. Blau, F., and L. Kahn. 2005. “Do Cognitive Test Scores Explain Higher US Wage Inequality?” Review of Economics and Statistics 87: 184– 93. Bourguignon, F., F.H.G. Ferreira, and M. Mene´ ndez. 2007. “Inequality of Opportunity in Brazil.” Review of Income Wealth 53 (4): 585– 618. Brown, G., J. Micklewright, S.V. Schnepf, and R. Waldmann. 2007. “International Surveys of Educational Achievement: How Robust are the Findings?” Journal of the Royal Statistical Society 170 (3): 623– 46. Brunello, G., M. Giannini, and K. Ariga. 2006. “The Optimal Timing of School Tracking.” In L. Wo ¨ ßmann, and P.E. Peterson, eds., Schools and the Equal Opportunity Problem. Cambridge, MA: MIT Press. Brunello, G., and D. Checchi. 2007. “Does School Tracking Affect Equality of Opportunity? New International Evidence.” Economic Policy 22: 781 –861. ´ , A., and R. Dome Castello ´ nech. 2002. “Human Capital Inequality and Economic Growth: Some New Evidence.” Economic Journal 112: C187– 200. Castro-Leal, F., J. Dayton, L. Demery, and K. Mehra. 1999. “Public Social Spending in Africa: Do the Poor Benefit?” World Bank Research Observer 14 (1): 49 –72. Checchi, D., and V. Peragine. 2005. “Regional Disparities and Inequality of Opportunity: The Case of Italy.” IZA Discussion Paper No. 1874. Institute for the Study of Labor, Bonn, Germany. ———. 2010. “Inequality of Opportunity in Italy.” Journal of Economic Inequality 8 (4): 429–50. Cohen, G.A. 1989. “On the Currency of Egalitarian Justice.” Ethics 99: 906–44. DiNardo, J., N. Fortin, and T. Lemieux. 1996. “Labor Market Institutions and the Distribution of Wages, 1973-1992: A Semi-Parametric Approach.” Econometrica 64 (5): 1001–44. Dworkin, R. 1981. “What is Equality? Part 2: Equality of Resources.” Philosophy and Public Affairs 10 (4): 283 –345. ´ ndez, R., and J. Galı Ferna ´. 1999. “To Each According to . . . ? Markets, Tournaments, and the Matching Problem with Borrowing Constraints.” Review of Economic Studies 66: 799– 824. Ferreira, F.H.G., and J. Gignoux. 2011a. “The Measurement of Inequality of Opportunity: Theory and an Application to Latin America.” Review of Income Wealth 57 (4):622–57. ———. 2011b. “The Measurement of Educational Inequality: Achievement and Opportunity.” World Bank Policy Research Working Paper No. 5873. Washington, DC. Ferreira, F.H.G., J. Gignoux, and M. Aran. 2011. “Measuring Inequality of Opportunity with Imperfect Data: The Case of Turkey.” Journal of Economic Inequality 9 (4): 651–80. Fields, G.S., and E.A. Ok. 1996. “The Meaning and Measurement of Income Mobility.” Journal of Economic Theory 71 (2): 349– 77. Fleurbaey, M. 2008. Fairness, Responsibility, and Welfare. Oxford: Oxford University Press. Fleurbaey, M., and V. Peragine. 2012. “Ex ante versus ex post equality of opportunity.” Economica ( published online 12 July 2012). Foster, J., and A. Shneyerov. 2000. “Path Independent Inequality Measures.” Journal of Economic Theory 91: 199–222. Ferreira and Gignoux 245 Gamboa, L.F., and F.D. Waltenberg. 2012. “Inequality of Opportunity for Educational Achievement in Latin America: Evidence from PISA 2006–2009.” Economics of Education Review 31 (5): 694– 708. Hanushek, E., and L. Woessmann. 2006. “Does Educational Tracking Affect Performance and Inequality? Differences-In-Differences Evidence across Countries.” Economic Journal 116: C63– C76. Jakubowski, M., and A. Pokropek. 2011. “Measuring Progress in Reading Achievement between Primary and Secondary School across Countries.” Working Paper n820. University of Warsaw, Faculty of Economic Sciences, Warsaw, Poland. Jakubowski, M., H.A. Patrinos, E.E. Porta, and J. Wisniewski. 2010. “The Impact of the 1999 Education Reform in Poland.” World Bank Policy Research Working Paper 5263. Washington, DC. Manning, A., and J.-S. Pisckhe. 2006. “Comprehensive versus Selective Schooling in England in Wales: What Do We Know?” IZA Discussion Paper 2072. Institute for the Study of Labor, Bonn, Germany. Macdonald, K., F. Barrera, J. Guaqueta, H. Patrinos, and E. Porta. 2010. “The Determinants of Wealth and Gender Inequity in Cognitive Skills in Latin America.” World Bank Policy Research Working Paper 5189. Washington, DC. Marks, G.N. 2005. “Cross-National Differences in Accounting for Social Class Inequalities in Education.” International Sociology 20 (4): 483–505. Micklewright, J., and S. Schnepf. 2007. “Inequality of Learning in Industrialized Countries.” In S. Jenkins and J. Micklewright, eds., Inequality and Poverty Re-examined. Oxford: Oxford University Press. Mislevy, R.J. 1991. “Randomization Based Inference about Examinees in the Estimation of Item Parameters.” Psychometrika 56: 177 –96. Mislevy, R.J., A.E. Beaton, B. Kaplan, and K. Sheehan. 1992. “Estimating Population Characteristics from Sparse Matrix Samples of Item Responses.” Journal of Educational Measurement 29 (2): 133–61. Morrisson, C., and F. Murtin. 2007. “Education Inequalities and the Kuznets Curves: A Slobal Perspective Since 1870.” Paris School of Economics Working Paper 2007-12. Paris, France. OECD 2006. PISA 2006 technical report. Paris: OECD. ———. 2009. PISA Data Analysis Manual. Paris: OECD. Psacharopoulos, G. 1994. “Returns to Investment in Education: A Global Update.” World Development 22:1325– 43. Pekkarinen, T., R. Uusitalo, and S. Kerr. 2009. “School Tracking and Development of Cognitive Skills.” IZA Discussion Paper No. 4058. Institute for the Study of Labor, Bonn, Germany. Roemer, J.E. 1998. Equality of Opportunity. Cambridge, MA: Harvard University Press. Salehi-Isfahani, D., and N.B. Hassine. 2012. “Equality of Opportunity in Education in the Middle East and North Africa.” Working Papers e07-33. Virginia Polytechnic Institute and State University, Department of Economics, Blacksburg, VA. Schultz, G., H.W. Ursprung, and L. Wossmann. 2008. “Education Policy and Equality of Opportunity.” Kyklos 61 (2): 279–308. Sen, A. 1985. Commodities and Capabilities. Amsterdam, The Netherlands: North-Holland. Shorrocks, A. 1978. “The Measurement of Mobility.” Econometrica 46: 1013–24. ———. 1999. “Decomposition Procedures for Distributional Analysis: A Unified Framework Based on the Shapley Value.” unpublished manuscript, University of Essex, Colchester, UK. Thomas, V., Y. Wang, and X. Fan. 2001. “Measuring Education Inequality: Gini Coefficients of Education.” World Bank Policy Research Working Paper 2525. Washington, DC. van de Gaer, D. 1993. “Equality of Opportunity and Investment in Human Capital.” PhD dissertation, Catholic University of Leuven, Leuven, Belgium. van de Gaer, D., J. Vandenbossche, and J.L. Figueroa. 2012. “Children’s health opportunities and project evaluation: Mexico’s Oportunidades program.” World Bank Economic Review 28 (2): 282–310. 246 THE WORLD BANK ECONOMIC REVIEW van de Walle, D., and K. Nead. 1995. Public Spending and the Poor: Theory and Evidence. Washington, DC: Johns Hopkins and World Bank. Zheng, B. 1994. “Can a Poverty Index be Both Relative and Absolute?” Econometrica 62 (6): 1453–58.