75447 State Planning Organization of the Republic of Turkey and World Bank Welfare and Social Policy Analytical Work Program Working Paper Number 4: Inequality of Opportunity for Education: The case of Turkey Francisco H. G. Ferreira The World Bank Jérémie Gignoux The World Bank Ankara, March 2010 State Planning Organization of the Republic of Turkey and World Bank Welfare and Social Policy Analytical Work Program Working Paper Number 4: Inequality of Opportunity for Education: The case of Turkey Francisco H. G. Ferreira The World Bank Jérémie Gignoux The World Bank Ankara, March 2010 State Planning Organization World Bank Copyright @ 2010 The International Bank for Reconstruction and Development The World Bank 1818 H Street, NW Washington, DC 20433, USA All rights reserved The World Bank enjoys copyright under protocol 2 of the Universal Copyright Convention. This material may nonetheless be copied for research, educational or scholarly purposes only in the member countries of The World Bank. Material in this report is subject to revision. Inequality of Opportunity for Education: The case of Turkey iii Inequality of Opportunity for Education: The case of Turkey Table of Contents Abstract ................................................................................................................................................................... v 1. Introduction ......................................................................................................................................................... 1 2. The Data .............................................................................................................................................................. 2 3. Circumstances at Birth and School Enrollment................................................................................................... 4 4. Inequality of Opportunity for Educational Achievement .................................................................................... 6 5. Conclusions ......................................................................................................................................................... 9 References ............................................................................................................................................................... 11 Inequality of Opportunity for Education: The case of Turkey v Keywords Inequality of Opportunity; Education; Enrollment, Achievement; Turkey JEL Classification D39, D63, I21 Abstract This paper seeks to measure inequality of opportunity for education in Turkey, taking into account both the quantity (attainment) and the quality of schooling (achievement). Using DHS data, large gaps in age-enrollment profiles are documented across genders, regions and family backgrounds. The gender gap is particularly pronounced in the Eastern provinces, in rural areas, and for poorer and larger households. PISA data show that morally irrelevant circumstances also affect achievement. The lower bound for the share of the variance in test scores that is accounted for by such circumstances is between a quarter and a third, depending on the subject and on the procedure adopted to correct for sample selection. Among those circumstances, family background variables such as parental education, father’s occupation and the ownership of books, cultural possessions and electronics, seem to account for the largest inequality shares. Once the composition of families is controlled for, spatial location is considerably less important. Inequality of Opportunity for Education: The case of Turkey 1 1. Introduction – including incomes and well-being more generally – are at least in part the result of individual decisions, 1. Questions about the relationship between equity and if there is an ethical role for individual responsibility, and growth are at least as old as economics itself. In then perhaps equity – understood as the form of equality the preface to his Principles of Political Economy, that is socially just – requires equality in another space, David Ricardo (1817) wrote: in some sense logically prior to final outcomes. At the “The produce of earth – all that is derived from its risk of greatly over-simplifying, the search for this prior surface by the united application of labor, machinery space has taken us to concepts such as primary goods and capital – is divided among three classes of the (Rawls), capabilities (Sen), equality of resources community, namely the proprietor of the land, the (Dworkin) and equality of opportunity (Arneson, and Roemer). owner of the stock or capital necessary for its cultivation, and the laborers by whose industry it 4. While the conceptual literature on these different is cultivated. distributional domains is now both rich and well- established, its influence on applied economics has “But in different stages of society, the proportions remained marginal. Despite Amartya Sen’s Herculean of the whole produce of the earth which will be efforts to move us ‘from commodities to capabilities’, allotted to each of these classes […] will be the temptation to look for lost keys where the light essentially different… To determine the laws which shines has remained exceedingly powerful. Though regulate this distribution is the principal problem 1 incomes and consumption expenditures may be hard to in Political Economy” measure accurately, they are immensely easier to observe and measure than concepts such as capabilities or 2. The quest to understand the links between opportunities. development and distribution remained central to modern economics as well. Lewis (1954), Kuznets (1955) and 5. We think that this has recently begun to change, a long line of followers explored causation running and that it has begun to change in large part thanks to from economic growth and the patterns of structural a particular formalization of the concept of equality of change which are associated with it, to the distribution opportunities, due to John Roemer (1998), which lends of income. More recently, a thriving literature has itself reasonably well to observation in the kind of explored the reverse direction of causality, operating household data which -- even if somewhat more from different degrees of inequality to the nature and demanding than data on consumption or income alone 2 rate of economic growth. -- do exist in many countries. 3. But just as interest in the role of income inequality 6. Reduced to its essential core, Roemer’s definition experienced something of a resurgence in mainstream of equality of opportunity relies on a distinction between economics in the 1990s (on which, see Atkinson, 1997), two normatively different kinds of determinants of a many social scientists, philosophers and (even) particular outcome of interest (which he calls advantage). economists appeared to become less certain that the He calls those determinants over which individuals can inequality they should be fundamentally concerned with exercise some discretion (i.e. which are subject to some was the one they observed in the income space – or degree of individual choice or responsibility) “efforts”. indeed the one they might imagine in the space of Other determinants, over which individuals have no utilities. Building on Rawls (1971), influential authors control, are called “circumstances”. Equal opportunities like Sen (1980) and Dworkin (1981) challenged us to are said to attain in a society if circumstances are ask ourselves what equality – or the equality of what immaterial to the attainment of advantage. In such a –societies should really aim for? If individual outcomes situation, there will in general exist some inequality in 1 David Ricardo, Preface to Principles of Political Economy, 1817 (1911 edition, p.1), as cited by Atkinson and Bourguignon (2000), in their Introduction to the Handbook of Income Distribution. Atkinson and Bourguignon already sound somewhat apologetic for failing to resist the temptation to begin with this well-known quotation – an apology which is therefore even more clearly warranted here… 2 This literature has spanned economics, including the pioneering work of Loury (1981) and Galor and Zeira (1993), and economic history, including work by Engermann and Sokoloff (1997). The literature has now been reviewed so often that space here is too short even for a survey of surveys. One good recent survey is Voitchovsky (2005). Inequality of Opportunity for Education: The case of Turkey 2 advantages. But such differences will be attributable dimensions: quantity, or attainment (which we capture only to differences in “efforts”, and not in circumstances. through enrollment-age profiles), and quality, or achievement (which we measure through standardized 7. Roemer’s recent (and growing) influence over test scores. Few countries are better suited for such an applied economists arises because such a definition has endeavor than Turkey – both for data availability and an immediate statistical implication: given the law of for intrinsic interest reasons. Turkey has good data both large numbers, equality of opportunity would imply on enrollment, from the Turkish Demographic and that advantage should be distributed independently of Health Survey (TDHS) of 2003/4, and on achievement, circumstances. To the extent that some advantages from the Program for International Student Assessment (incomes, educational attainment, health status) and (PISA) 2006 data set – both of which are described in some circumstances (race, gender, family background, more detail below. birthplace) can be observed in large enough samples, the hypothesis of stochastic independence can be tested 11. What the analysis of these data reveals is a complex statistically (see, e.g. Lefranc, Pistolesi and Trannoy, pattern of inequality of opportunity, with profound 2008). In addition, to the extent to which inequalities differences in enrollment rates across genders, regions between circumstance-homogeneous groups (which and family backgrounds, which are generally Roemer calls types) can be associated with inequality compounded by additional differences in student of opportunity, the latter concept can be measured achievement. But not all circumstances matter in the cardinally – albeit in consequential terms. In other same ways, and exclusion patterns are not always as words, a certain amount (or share) of inequality in a they at first appear. Gender is a dominant factor in particular advantage can be related to inequality of explaining differences in enrollment, but not in opportunity for the attainment of that particular outcome achievement: once girls get to school, they tend to do (See, e.g. Bourguignon, Ferreira and Menendez, 2007; no worse than boys (better, if one does not control for and Ferreira and Gignoux, 2008). selection). Regional differences in enrollment, which are large in absolute terms, are not statistically significant 8. By empirically identifying, describing and once one controls for other circumstances. Differences quantifying inequality in a normatively more appropriate in family background, whether measured in terms of space for assessing social justice, we believe this incipient parental education, father’s occupation or asset literature contributes towards an economic understanding ownership, matter for all children, but much more so of equity. In due course, such empirical measures may for girls. There is much one can learn about Turkish even be related to the broader processes of growth and society, and the nature of its inequalities of opportunity, 3 development, in ways to analogous to those in which from applying these concepts to these rich data sources. income and wealth inequality have often been. 12. The remainder of the paper is organized as follows. 9. In this paper, we extend the cardinal approach to Section 2 briefly describes the three data sets used in the measurement of inequality of opportunity to an the analysis. Section 3 reports the pattern of correlations advantage other than income, namely education. Because between individual circumstances and school enrollment. education has intrinsic value to individuals, it can Section 4 briefly discusses the measurement of inequality certainly be considered an advantage, in Roemer’s of opportunity, and applies it to educational achievement. terms. Because it has such well-documented instrumental Section 5 concludes. value for the achievement of other valued outcomes, such as health and incomes, it also reinforces an opportunity loop. 2. The Data 10. Education is itself difficult to measure, and the 13. Data from three surveys are used in the analysis. paper investigates its distribution along two key Section 3 constructs profiles of school enrollment rates 3 There is a recent literature on the determinants of school enrollment in Turkey, including Tansel (2002) and Kirdar (2007). We take a slightly different approach here, focusing on the description and measurement of inequality of opportunity for education, rather than seeking to estimate the causal effects of specific circumstances on enrollment. Nevertheless, some of our descriptive results are close to findings in that literature. Inequality of Opportunity for Education: The case of Turkey 3 by age using Turkey’s latest Demographic and Health in the 2006 Program for International Student Survey (TDHS), which was fielded between December Assessment (PISA) data set for Turkey. The PISA survey 2003 and March 2004 by the Hacettepe Institute. The was fielded at Turkish schools by the Organization for data were collected from a sample of 10,836 households, Economic Cooperation and Development between representative at the national level but also at the level March and November 2006, at the same time as in 56 of the five major regions of the country (the West, other countries. It collected information for a sample South, Central, North and East regions). Information of 4,942 15 year-old pupils enrolled in school, at grade on basic socio-economic characteristics of the population 7 or above. All children surveyed took tests in reading, was collected for all household members, and all ever- mathematics and science. In addition to the test scores married women between 15 and 49 years old further in these subjects, the PISA data set also reports the answered a detailed questionnaire on demography and student’s gender and some information on family health. 8075 women provided such information. background, including the mother’s and father’s education; the father’s occupation; number of books 14. Information on enrollment was collected for all owned by the household; durable ownership; and 7 18,376 household members aged 6 to 24. For all these “cultural possessions”. School location variables also children and young people, information was also allow us to allocate the student to a geographic region collected on the following circumstance variables: of the country, as well as to rural or urban areas 8 gender, region of residence, the type of area of residence, (including a disaggregation into large or small cities). the levels of education of the mother and the father, the number of children in the household, the mother’s 16. The PISA sample is representative for the national mother tongue, and household wealth. We classified population of 15 year-olds enrolled in grades 7 and region of residence into five broad categories (South, above. However, because of the incomplete enrollment North, West, Center, and East) and the type of area into at age 15, repetition, and sample design issues, the 4 rural, small urban areas, and large cities. Parental sample coverage rate (the ratio of the population education was coded into three categories: no formal represented by the survey to the total population of 15 education or unknown level; primary education; and year-olds) is only 47% in Turkey. Particularly worrisome, secondary or higher education. The mother’s mother of course, is the fact that selection into enrollment and tongue was coded into Turkish or another language; repetition is clearly non-random. The analysis in Section and the number of children in the household into: 1 or 4 below presents results mostly for the universe for 5 2; 3 to 5; and 6 or more. Household wealth was which the sample is representative, in line with the measured by means of a Filmer-Pritchett (2001) asset literature based on PISA surveys. However, we also index, constructed by principal components analysis report two alternative attempts to correct for selection from information on household ownership of various biases in estimating inequality of opportunity for durable goods, housing quality and access to amenities. educational achievement. For this analysis, households were simply divided into 6 quartiles of the distribution of this index. 17. In order to correct for selection into the PISA sample, we use two-sample reweighting techniques and 15. Section 4 analyses information on scholastic the data from the 2006 Household Budget Survey (HBS). achievement based on standardized test scores reported This survey has a nationally representative sample of 4 In the TDHS, urban areas are defined as settlements with populations larger than 10,000, and large cities include Istanbul, Izmir, Bursa and Adana, in addition to the capital Ankara. Although information on area or region of residence is less consistent with an interpretation as an exogenous circumstance than area of birth, this is the only information available for this age group in the TDHS. We assume that the two are very closely correlated for children. 5 Children within households are identified as the children of the head or spouse, aged less than 18. 6 Details of the construction of this asset index, and of its distribution, can be found in Ferreira and Gignoux (2009). 7 Parental education was classified as in the TDHS, but secondary and higher education were coded separately. Father’s occupation was coded into three groups, following the ISCO 88 classification: a) legislators, senior officials, professionals, technicians and clerks; b) service workers, craft and related trade workers, plant and machine operators or assemblers, or unoccupied; and c) skilled agricultural or fishery workers, or workers with an elementary occupation. Durables include: dishwasher, VCR/ DVD, cell phones, car, computer, and TV. Cultural possessions refer to the household’s ownership of works of literature, art and poetry. 8 Because of differences in the sample frame, the regional breakdown for the PISA survey is coarser than for the TDHS, and includes only three main categories: West, Center, and East. Inequality of Opportunity for Education: The case of Turkey 4 about 8,500 households, which includes 683 individuals boys and girls in every other region of the country. aged 15. A set of circumstance variables comparable to Their enrollment rates peaks at just over 85%, at age some of those from the PISA survey are available, 9, and is below 40% by age 15. Although there are other 9 including gender, area type, parental education and cross-regional differences among girls (but almost none father’s occupation. From the HBS sample weights, the among boys), they pale into insignificance when total national population of the groups of 15 year-olds compared to the gap between the Eastern region and with each specific set of characteristics can be estimated. the rest of the country. These estimates of the populations of the different “types” are used in section 4 to provided two alternative 20. Figure 2 investigates this further, by disaggregating estimates of the effects of selection on the measures of profiles for urban (further classified into large cities inequality of opportunity for educational achievement. and other towns) and rural areas. The top panel presents results for the country as a whole: while there are small differences between large and smaller cities, the rural- 3. Circumstances at Birth and School urban enrollment deficit is more pronounced throughout, Enrollment and becomes particularly substantial in the transition to secondary schooling. The bottom panels focus on 18. A natural way to begin an investigation of the females, and explore the differences between the Eastern distribution of opportunities for schooling is to consider region (panel b) and other regions (panel c). Being born how the age-enrollment profile varies by population in a rural area is a disadvantage for girls across the sub-group, where the subgroups are defined by country, but only from age 13 upwards outside the characteristics over which the students have no individual Eastern region. It is only in the East that (some 20% control – i.e. by circumstance variables. Figure 1 presents of) rural-area girls are excluded from schooling this profile for the overall population of 6-24 year-olds throughout the early primary years as well. By age 15, in Turkey in 2006, as taken from the TDHS. The top when approximately 80% of girls in urban areas in the panel presents the overall, as well as the gender-specific, rest of country are still enrolled, just over 50% of urban profiles. Just over 50% of children are enrolled at age girls and fewer than 20% of rural girls are enrolled in 6, while the other half are enrolled between the ages of the East. 7 and 8. There is almost universal enrollment between the ages of 8 and 12, although 7% of girls and 2% of 21. Gender, region and area of residence are not the boys never make it to school even at those peak ages. only morally irrelevant, pre-determined circumstances A substantial drop in enrollment occurs from age 13 correlated with educational attainment in Turkey. The (roughly sixth grade), and it accelerates at age 16 (when educational background of one’s parents is strongly secondary school begins). Only about a quarter of associated with enrollment, as shown by Figure 3, for students are enrolled at age 18, when secondary school mother’s education. For both boys and girls, the profiles should be completed. Average enrollment in tertiary of those whose mothers have no formal education lie education is approximately 16% between ages 18 and 23. Throughout the enrollment decline range (ages 13- everywhere below the profiles of those with more 18), girls’ enrollment falls earlier and faster than boys’. educated mothers. Once again, however, the gaps are At age 15, for instance, female enrollment is almost 20 considerably larger for girls than for boys: At age 16, percentage points below male enrollment. 90% (60%) of boys with highly educated (uneducated) mothers are enrolled. For girls, the corresponding rates 19. The two bottom panels of Figure 1 disaggregate are around 90% and 30%. At that age, the parental boys’ and girls’ enrollment profiles by the five regions. education gap is twice as large for girls as for boys, The broad pattern of the profiles is similar across regions, suggesting that the intergenerational persistence of with one striking exception: the profile for girls in the educational inequality will be more pronounced for Eastern region lies everywhere below the profiles for women than for men. 9 A few differences remain in the definition of the circumstances in the two surveys. In particular, in the HBS, urban areas are identified as settlements with more than 20,000 inhabitants, whereas in the PISA, the urban threshold is at 15,000 inhabitants. Inequality of Opportunity for Education: The case of Turkey 5 22. pattern of gender inequality in educational attainment correlated. Table 1 presents the results of an attempt to is consistent across all other circumstances. Growing disentangle their partial effects on enrollment, by means up in a household with many other children (also a of a simple probit regression of enrollment at age 15 circumstance beyond the control of the individual child), on all of the previously discussed circumstances, as or a in a poorer household, is associated with lower well as father’s education and the mother tongue of the 10 enrollment across the age range, and for both boys child’s mother. These marginal effects are clearly not and girls (Figures 4 and 5, respectively). However, the interpretable as causal, since there are many potentially strength of the negative correlation, say, between the relevant determinants which are omitted from the number of children in the household and enrollment, is specification. They do, however, provide partial markedly greater for girls than for boys. correlations that complement the description so far. 23. When the population of children is disaggregated 27. The probit regressions are estimated on the sample by quartiles of the distribution of the household wealth of the total population of 15 year olds, as well as on index, as in Figure 5, the powerful effect of samples of girls and boys separately. Gender remains socioeconomic background on education is evident. a powerful correlate of enrollment even after controlling This is true for all age groups for girls (since first- for the other observed circumstances, with girls quartile girls never reach the 90% enrollment mark), appearing to be 50% less likely to be enrolled than boys and becomes pronounced for boys after ages 12-13. By at the sample mean. Family wealth also has a powerful age 20, more than half of the young men (and women!) partial correlation with enrollment, with coefficients hailing from the top quartile of the wealth distribution on all three quartiles being significantly higher than the are attending college, while the same is true for fewer first, in the joint and in each gender-specific sample. than 10% of men and women from the bottom quartile. Parental education is important, although father’s education is only significant in the combined sample 24. These various profiles document that school and in the girls sample – suggesting that both less and enrollment in Turkey is evidently not independent from more educated fathers try to send their sons to school. circumstances at birth. Family background (in terms of Children in larger households (or, more specifically, wealth, parental education, and family size), gender those with more children) are less likely to be enrolled and place of residence (both the region and whether in at age 15 although this effect, too, is driven by girls. a city or the countryside) are all statistically significantly Similarly, girls in rural areas are significantly less likely associated with how long a Turkish child is likely to to be enrolled in schools, whatever their family stay in school. And this amount of schooling, as we background. Interestingly, however, residence in the know from a copious international literature, is in turn East region and mother’s mother tongue become causally related to future earnings and standard of living insignificant when controlling for the entire set of more broadly. circumstances. 25. In the specific case of Turkey, no circumstance 28. Another way to illustrate the inter-relationship appears to be more important in influencing school- between the different circumstance variables and leaving than gender. But the disadvantage that girls enrollment is to consider the “cumulative” effect of experience vis-à-vis boys is by no means uniform across belonging to many disadvantaged sub-groups, when the country. Spatially, it is clearly more pronounced in compared to those who belong to the most advantaged the East, and in rural areas. It is also more marked for cells in the partition. Figure 6 illustrates one possible girls born in poorer and larger families, or to less such comparison, by plotting the enrollment-age profile educated mothers, than it is for girls from a higher for girls born in the rural areas of the Eastern Region, socio-economic background, or from smaller families. in households with six or more children, to uneducated, non-Turkish speaking mothers (a group that accounts 26. While Figures 1-5 are descriptively powerful, these for roughly 1% of the population in the 6-24 age range). various circumstances are obviously mutually inter- It also plots the profile for boys in urban areas of Central 10 Though the differences become less pronounced (and often statistically insignificant) above age 20. Inequality of Opportunity for Education: The case of Turkey 6 Turkey, in households with two children or fewer, and (non-parametric) inequality decomposition, or using a native Turkish-speaking mothers with some education parametric alternative, which relies on regression 11 (a group which accounts from some 2.5% of the analysis. population). The difference in enrollment rates is striking at every age. It is lowest at age 9, when enrollment for 31. Ferreira and Gignoux (2008) describe each of these the disadvantaged group reaches a peak at around 70%. methods in some detail (in the context of earnings, In absolute terms, it is highest at the crucial 14-15 age income and consumption inequality) and note the range, when children are transitioning from primary to potential trade-off between parametric methods (which secondary schooling. At this age, children in our impose a functional form assumption on the relationship advantaged group begin to come away from 100% between advantage and circumstances) and the non- enrollment, while only some 10% of children in the parametric decompositions (where conditional mean disadvantaged group are enrolled. estimates become imprecise and small sample biases can be considerable for fine partitions of the sample). In the present context, the wealth of circumstances 4. Inequality of Opportunity for available in the data would be consistent with a partition Educational Achievement of the sample into as many as 589,824 cells! Non- parametric inequality decompositions are therefore not 29. The extent to which children accumulate human an option, and we rely below on a regression-based capital at school depends not only on how many years decomposition. they attend classes for, but also on how good those classes are. While information on attainment is essential 32. Furthermore, another feature of the data makes the for an understanding of the distribution of educational use of the regression-based decomposition the most opportunities in Turkey, it is not sufficient. It must be natural. Because different items (or questions) in any complemented by information on actual educational test have different degrees of difficulty, a simple achievement. In this section, we use data from the PISA proportion of right answers is a poor measure of the 2006 survey for Turkey, which contains standardized latent variable of interest, which is the student’s test scores for a sample of nearly 5,000 15-year old knowledge, or achievement. PISA surveys around the students across the country, in three subjects: Turkish world (as well as many other applications) therefore (reading), Mathematics and Science. It also contains a use Item Response Theory methods to adjust for these rich information set on those children’s circumstances differences in item difficulty, under some assumption which, as described in Section 2, includes gender, about the underlying distribution of ability in the father’s and mother’s educational attainment, father’s population. The process, which is described in detail in occupation, region and area of residence, language Mislevy (1991) and Mislevy et al. (1992), generates a spoken at home, durable ownership, book ownership, set of scores with no inherent metric, which are then and cultural possessions. standardized around an arbitrary mean (typically 500), and with an arbitrary variance. 30. To the extent that we are prepared to treat each of these variables as representing true Roemerian 33. The arbitrary nature of the mean (which precludes circumstances – i.e. characteristics that lie beyond the the need for scale invariance as a property of the influence of the children themselves – then we can inequality index) and the normal distribution of the estimate a lower-bound measure of inequality of scores that results from standardization suggest the opportunity for educational achievement by calculating variance as the natural inequality indicator of choice. the share of the overall inequality in achievements When the variance is used, the parametric estimate of which is accounted for by these circumstances. In the (lower bound) share of inequality due to principle, this can be done either by means of a standard circumstances is given simply by the R2 of a linear 11 These are lower-bound estimates of the effect of circumstances because not all circumstances are observed. If additional circumstances were to become observable, this might raise (but could not lower) the between-group component of the inequality decomposition, or the explanatory power of a regression. For a more detailed discussion, which incorporates “effort” variables explicitly, see Ferreira and Gignoux (2008). Inequality of Opportunity for Education: The case of Turkey 7 regression of the test score on circumstances. And from achievement, a word is needed on how we have sought such a regression, estimates of the additively to address the important selection problem present in decomposable partial effects of each circumstance can the Turkish (and in a number of other) PISA data, which be calculated straightforwardly, as shown in Ferreira was briefly mentioned in Section 2. Enrollment at age and Gignoux (2009). 15 is 67% in Turkey. Repetition and non-response among those enrolled lowers the representativeness of 34. Three separate regressions (for reading, the PISA data set to 47% of the 15 year-old population mathematics, and science scores, respectively) are in the country (OECD 2007). To restrict an assessment reported in Table 2. As in many other countries, girls of educational inequality only to PISA respondents, perform significantly better than boys in reading, but therefore, would ignore a potentially important share worse in mathematics. There is no significant gender of the overall inequality, by excluding nearly half of difference in science scores. Children whose parents the relevant population, which is clearly not randomly have secondary or college education score more highly selected. in all three subjects, but the effect of primary education alone is not significant (as compared to the reference 37. The difficulty with addressing this problem, as with category of children whose parents have no formal any correction for selection, is that the counterfactual 12 education). test scores which non-participating 15 year-olds would have obtained if they had taken the test are not observed. 35. Children whose fathers are occupied as service While standard Heckman correction procedures would workers, craft and related trade workers, plant and help, by controlling for selection on observables, many machine operators or assemblers, or who are unoccupied, important likely determinants of participation in the have significantly lower scores than children whose exams are not observed. We therefore propose a non- fathers work as legislators, senior officials, professionals, parametric two-sample procedure that generates plausible technicians or clerks, even after controlling for parental higher and lower alternative estimates for selection education. A father’s occupation in agriculture, fishery, correction. or other elementary occupations, however, is not significant once urban-rural differences are controlled 38. We exploit the fact that Turkey’s Household Budget for. These differences are both large and significant, Survey, which is nationally representative, is also with rural areas at a substantial achievement available for 2006. We partition the population of 15 disadvantage. Scores from schools in the Central and year-olds in both the HBS and PISA into groups with West regions are significantly higher than those from identical observable circumstances, using region, urban- the East region. This is in contrast with the result for rural status and mother’s education as the defining 13 the quantity of schooling reported in Table 1, where characteristics. If the expanded population of 15 year- regional dummies were not significantly partially olds in cell k of this partition in the PISA (HBS) correlated with enrollment, once other circumstances were controlled for. In Table 2, however, the coefficients survey is ( ), then our “low alternative” for the area and regions variables are not only significant, estimate of the selection correction consists of but quantitatively substantial, with absolute values of reweighing each score by . about half a standard deviation of the overall distribution of test scores, whereas the coefficients estimated for 39. This “corrects” for selection on observables, because the father’s occupation and parental education are it “reintroduces” the 15 year-olds who dropped out (or comprised between 20% and half of a standard deviation. otherwise did not participate in the PISA exam), under the assumption that the distribution of test scores for 36. Before presenting our estimates of the lower-bound non–participants would have been identical to the opportunity shares of inequality in educational distribution for participants, within each cell k. This 12 For reading, secondary education is not significant for fathers, and tertiary education is not significant for mothers. 13 These three variables are defined identically in the two surveys, so that the partitions should be strictly comparable. A finer partition would have been possible, but would have generated statistically imprecise estimates of population weights in the HBS, given the sample size (683 individuals aged 15). Inequality of Opportunity for Education: The case of Turkey 8 assumption is, of course, the familiar assumption of no the variances between two- and threefold. The selection on unobservables. opportunity shares of inequality, on the other hand, do not vary much across subjects and turn out to be 40. But it is quite likely that selection did not depend relatively insensitive to the alternative corrections for only on the variables used to partition the population selection. With no selection correction, a minimum of into . In particular, it is plausible that, within each 26% of the variance in reading and Math scores, and 27% of the variance in science scores is due to cell k, non-participants would on average have had circumstances. These rise only very slightly to 27-28% worse scores than participants. Under that assumption, under the more conservative selection correction a likely “higher alternative” effect of selection would procedure. Even under the higher alternative correction be obtained by giving a proportion the procedure, the lower-bound estimate of the opportunity share of inequality in educational achievement rises to lowest score in cell k, , after the reweighting process some 32-33%. 14 described above. 44. The bottom panel of the Table 3 reports the partial 41. Intuitively, this corresponds to counterfactually shares of inequality of opportunity associated with attributing to each and every non-participant, the worst individual circumstances, namely gender, father’s test score actually observed among participants, within education, mother’s education, father’s occupation, type each cell k. Figure 7 shows the kernel density functions of area, and region. These shares are calculated so as for the standardized PISA test scores in reading in 2006, to add up to the overall effect, in the manner described under three different scenarios. The top panel depicts in Ferreira and Gignoux (2009). The partial shares are the observed sample distribution, with no correction only reported for the regression without selection for selection. The middle panel depicts the counterfactual correction, because the partition used for that distribution with the “lower alternative” correction for selection. The bottom panel depicts the counterfactual correction is based on some but not all of the independent distribution with the “higher alternative” correction for variables in the regression. This would make analysis selection. of the partial effects in the corrected regression difficult to interpret. 42. Table 3 summarizes our results on inequality of opportunity for educational achievement. For each of 45. When all circumstances are considered together the three distributions of test scores (no correction, and controlled for, family background seems to be the lower alternative and higher alternative correction for dominant source of inequality of opportunity for selection) and for each subject (reading, mathematics achievement in Turkey. Taking Mathematics scores as and science), the top panel of the Table reports both an example, mother’s and father’s education combined total variance and the lower-bound on the share of this account for 7.5 of the 26 percentage points in the overall inequality which corresponds to inequality of share. Add father’s occupation and the three “asset” opportunity. This latter estimate is simply the R2 of the indicators (numbers of books, durables and cultural regressions reported in Table 2 (for the distribution with possessions), and these family background variables no correction for selection), and the R2 of analogous add up to 21 of the 26 percentage points. Interestingly, regressions for the adjusted distributions. the largest part of this ‘family effect’ shows up through material possessions – books, durables and ‘cultural 43. While the lower selection correction increases the possessions’ represent 12.8 percentage points. When variance for the distribution of reading scores, it has these three variables are omitted from the regression, almost no effect on the other two variances. The higher some (though by no means all) of their effect is picked alternative correction, on the other hand (and as could up by mother’s and father’s education, with which they 15 be expected from an inspection of Figure 7), increases are collinear. Although there are some small 14 In this procedure, the specific observations whose scores are modified are chosen randomly within each cell. 15 In that specification, father’s education and mother’s education shares (for reading scores) are 0.041and 0.024 respectively. They are higher in the Math and Science decompositions. Inequality of Opportunity for Education: The case of Turkey 9 differences, the dominance of the family background across the whole relevant age range. In particular, we variables is consistent across all three subjects. found that girls residing in Eastern provinces – and particularly (but not exclusively) in their rural areas – 46. Although the area type and the region where schools are much less likely to attend school than their are located were highly significant in the regression in counterparts in other parts of the country, and than boys. Table 2, their partial shares are relatively small in Other circumstances associated with a lower probability magnitude, generally accounting for between five and of enrollment – or a higher probability of early drop- six percentage points of the 26 - 27 percentage points out – such as a lower household wealth index; a larger of the overall lower-bound circumstance share. Except number of children; or lower levels of parental education, in the case of mathematics, the rural-urban divide is appeared to be systematically more important for girls more important than the broad regional location. As in than for boys. In other words, disadvantageous the case of enrollment, while spatial variables remain circumstances such as a poorer family background are significant once other population characteristics are more likely to lead girls than boys to drop out of school controlled for, they account for much smaller variance early – thus potentially generating a more resilient shares than one might expect from the raw, absolute inequality trap for Turkish women than for their menfolk. regional differences. These appear to be explained to a large extent by differences in the family background 50. Once the pattern of co-variances between compositions across the residents of different regions circumstances is taken into account, by means of a and areas. simple multivariate discrete choice model, a more nuanced picture emerges. Gender remains a key 47. The student’s gender, which was so important in cleavage, with girls only half as likely to be enrolled at explaining enrollment, is much less important in age 15 as boys, at sample mean values of other correlates. accounting for differences in achievement. It is largest But while family background variables – such as in the case of reading, where it accounts for 4.1% of household wealth; secondary or higher levels of parental total variance; and this is a subject where girls do education; and family size – retain importance in the significantly better than boys. While this may to some multivariate analysis, spatial variables become much extent reflect differences in selection across genders – less significant (or even acquire counter-intuitive signs). fewer girls are enrolled, so perhaps average ability is We interpret these results as suggesting that there is higher among enrolled girls than among enrolled boys nothing inherent about the East as a region, or about – there is clearly no evidence whatever to suggest that smaller towns, or about being a non-Turkish ethnic girls do worse at school than boys in Turkey. minority, that prevents children with these characteristics from going to school. They do have lower enrollment rates, but this reflects lower levels of advantage in 5. Conclusions family background: their households are larger and poorer, and their parents have less formal schooling. 48. Although the relationship between equity and growth Once those factors are accounted for, the only spatial has long been of concern to economists, recent circumstance that retains its original sign and significance developments in the conceptualization of inequality of from the univariate analysis is rural residence for girls. opportunity have arguably made it easier for applied economists to measure and decompose the kinds of 51. Broadly similar conclusions apply with regard to inequality that matter most, rather than simply those on educational achievement, as measured by PISA test which data is more readily available. Following John scores, at least in terms of the relative importance of Roemer’s (1998) definition of inequality of opportunity school location once family background is controlled as that kind of inequality which is driven by morally for. In this case, schools located in the East or in rural irrelevant, pre-determined circumstances, this paper areas are statistically significantly associated with lower has investigated the nature and magnitude of unequal test scores, even controlling for all other circumstances. opportunities for education in Turkey. But quantitatively, spatial variables account for no more than a fifth of the overall opportunity share of inequality 49. Using DHS data, we documented large differences in achievement in Turkey. This overall share, estimated in enrollment by student’s gender and spatial location parametrically as a lower-bound, is not trivial. Morally Inequality of Opportunity for Education: The case of Turkey 10 irrelevant circumstances account for over a quarter of additional purchasing power it generates for various total inequality in achievement, even when no correction goods that function as inputs into a broadly defined for sample selection bias is attempted. When (a two- ‘production function’ of human capital. sample reweighting) correction for selection is implemented, the lower bound share of the variance 53. Gender, which was of paramount importance as a attributable to opportunities rises to between 27 and 33%. circumstance determining access to education (via differences in enrollment and retention) is not an 52. Three-quarters to four-fifths of these shares are important determinant of achievement, conditional on accounted for by family background variables, being in school. In fact, possibly as a result of selection prominently including indicators for ownership of into enrollment, the one subject where gender accounts durable electronics, books and other cultural possessions. for a sizable share of inequality (reading), is one where Father’s and mother’s education remain important even it is the boys who are disadvantaged. The policy lesson controlling for those variables, but father’s occupation for those concerned with girls’ education in Turkey is much less important. A prestigious occupation for seems to be that the big priority is really to get - and one’s father does not seem to contribute much to a keep – them in school. Once there, they seem to do child’s achievement directly – but only through the well enough. Inequality of Opportunity for Education: The case of Turkey 11 References Arneson, Richard (1989): “Equality of Opportunity for 2649, University Library of Munich, Germany. Welfare”, Philosophical Studies, 56: 77-93. Kuznets, Simon P. (1955): “Economic Growth and Atkinson, Anthony B. (1997): “Bringing Income Income Inequality.” American Economic Review, Distribution in from the Cold”, Economic Journal, 45(1): 1-28. 107 (441): 297-321. Lefranc, Arnaud, Nicolas Pistolesi and Alain Trannoy Atkinson, Anthony and François Bourguignon (2000): (2006): “Inequality of Opportunities vs. Inequality “Income Distribution and Economics”, of Outcomes: Are Western Societies All Alike?”, Introduction to Atkinson and Bourguignon (eds) ECINEQ Discussion Paper, 2006-54. Handbook of Income Distribution (Amsterdam: Lefranc, Arnaud, Nicolas Pistolesi and Alain Trannoy, Elsevier). (2008): "Inequality of Opportunities Vs. Inequality Bourguignon, François, Francisco H.G. Ferreira and Of Outcomes: Are Western Societies All Alike?" Marta Menéndez (2007): “Inequality of Review of Income and Wealth, Blackwell Opportunity in Brazil”, Review of Income Wealth, Publishing, vol. 54(4): 513-546 53 (4): 585-618. Lewis, Arthur W. (1954): “Economic Development with Dworkin, Ronald (1981): “What is Equality? Part 2: Unlimited Supply of Labour”, The Manchester Equality of Resources What is Equality? Part 2: School, 22(2): 139-91. Equality of Resources”, Philosophy and Public Loury, Glenn C. (1981): “Intergenerational Transfers Affairs, 10(4): 283-345. and the Distribution of Earnings”, Econometrica, Engerman, Stanley, and Kenneth Sokoloff (1997): 49: 843-867. "Factor Endowments, Institutions, and Differential Mislevy, R. (1991): "Randomization Based Inference Paths of Growth Among New World Economies: About Examinees in the Estimation of Item A View from Economic Historians of the United Parameters", Psychometrika, 56, 177-196. States." In Stephen Haber, eds., How Latin America Fell Behind. Stanford, C.A.: Stanford Mislevy, R., A. Beaton, B. Kaplan and K. Sheehan University Press. (1992): "Estimating Population Characteristics from Sparse Matrix Samples of Item Responses", Ferreira, Francisco H. G. and Jeremie Gignoux (2008): Journal of Educational Measurement, 29(2): 133- "The measurement of inequality of opportunity: 161. theory and an application to Latin America", Policy Research Working Paper # 4659, The Rawls, John (1971): A Theory of Justice. Cambridge, World Bank. MA: Harvard University Press. Ferreira, Francisco H. G. and Jeremie Gignoux (2009): Ricardo, David (1817): On The Principles of Political “Inequality of Economic Opportunity in Turkey: Economy and Taxation. London: Dent, 1911 As assessment using asset indicators”, World edition. Bank Research Department, unpublished Roemer, John E. (1998): Equality of Opportunity. manuscript. Cambridge, MA: Harvard University Press. Filmer, Deon and Lant H. Pritchett (2001): “Estimating Sen, Amartya (1980): “Equality of What?”, in S. Wealth Effects without Expenditure Data—or McMurrin (ed.) Tanner Lectures on Human Tears: An Application to Educational Enrolments Values, Cambridge University Press. in States of India”, Demography, 38: 115 - 32 Tansel, Aysit (2002): "Determinants of school attainment Galor, Oded and Joseph Zeira (1993): “Income of boys and girls in Turkey: individual, household Distribution and Macroeconomics Income and community factors," Economics of Education Distribution and Macroeconomics”, The Review Review, Elsevier, 21(5): 455-470. of Economic Studies, 60(1): 35-52. Voitchovsky, Sarah (2005): “Does the Profile of Income Kirdar, Murat (2007): "Explaining Ethnic Disparities Inequality Matter for Economic Growth”, Journal in School Enrollment in Turkey", MPRA Paper of Economic Growth, 10(3): 273-296. Inequality of Opportunity for Education: The case of Turkey 12 Figure 1: Enrollment-age profiles by gender and region Note: Enrollment-age profiles for total population, girls and boys (Figure 2a), and enrollment-age profiles by region of residence for girls (Figure 2b) and boys (Figure 2c). The distribution of the population of 6-24 years-old by region is the following: West 34.3%, South 13.7%, Center 21.1%, North 7.3%, and East 23.7%. Inequality of Opportunity for Education: The case of Turkey 13 Figure 2: Enrollment-age profiles by type of area. Note: Enrollment-age profiles for total population by area type (Figure 3a), and enrollment-age profiles by area type for girls in the East region (Figure 3b) and in the other regions (Figure 3c). Urban areas are defined as settlements with populations larger than 10,000 and large cities include Istanbul, Izmir, Bursa and Adana, beyond the capital Ankara. The distribution of the population of 6-24 years-old by area type is the following: capital or large city 26.5%, city or town 40.3%, and rural areas 33.2%. Inequality of Opportunity for Education: The case of Turkey 14 Figure 3: Enrollment-age profiles by mother’s education Note: Enrollment-age profiles for total population by mother’s education for girls (Figure 4a) and girls (Figure 4b). The distribution of the population of 6-24 years-old by mother’s education is the following: no education 34.4%, primary 50.5%, and secondary or higher 15.0%. Inequality of Opportunity for Education: The case of Turkey 15 Figure 4: Enrollment-age profiles by number of children Note: Enrollment-age profiles for total population by number of children aged less than 18 living in the household for girls (Figure 5a) and girls (Figure 5b). The distribution of the population of 6-24 years-old by number of children is the following: one or two 55.4%, three to five 36.2%, and six or more 8.5%. Inequality of Opportunity for Education: The case of Turkey 16 Figure 5: Enrollment-age profiles by quartiles of the asset index Note: Enrollment-age profiles for total population by quartiles of asset index, defined at the household level, for girls (Figure 6a) and girls (Figure 6b). The distribution of the population of 6-24 years-old by quartiles of household level asset index is the following: first quartile 31.0%, second quartile 25.6%, third quartile 24.6, and fourth quartile 18.9%. Figure 6: Enrollment-age profiles for one highly disadvantaged and one highly advantaged group. Note: Enrollment-age profiles for one highly disadvantaged and one highly advantaged group. The highly disadvantaged group consists of girls in rural areas of the East region whose mother has no education and is a non-Turkish native speaker living in a household with six children or more; it encompasses 1.0% of the population of 6-24 years-old. The highly advantaged group encompasses boys in urban areas of the Center region whose mother has some education and is a Turkish native speaker living in a household with one or two children; it encompasses 2.5% of the population of 6-24 years-old Inequality of Opportunity for Education: The case of Turkey 17 Figure 7: Distribution of standardized Turkish reading test scores under three alternative assumptions about selection into PISA participation Inequality of Opportunity for Education: The case of Turkey 18 Table 1: Circumstances at Birth and Enrollment at Age 15: A Partial Correlation Analysis Note: Probit estimates of enrollment at age 15. Marginal effects at sample mean reported. Source: DHS 2003, sample of 15 year-olds. The excluded categories are West region, Capital or large city, mother with no education or missing information, father with no education or missing information, one or two children, first quartile of the asset index, and mother’s mother tongue Turkish. * significant at 10%; ** significant at 5%; *** significant at 1%. Inequality of Opportunity for Education: The case of Turkey 19 Table 2: Reduced-form regression of standardized test scores on circumstances Note: Regression estimates of test scores in reading, mathematics and science. Source: PISA, Turkey, 2006, only the first plausible values are used. The standard deviations for the test scores are 87.3 in reading, 88.9 in math and 80.0 in science. * significant at 10%; ** significant at 5%; *** significant at 1%. Inequality of Opportunity for Education: The case of Turkey 20 Table 3: The Opportunity Share of Inequality in Educational Achievement with correction for selection into PISA participation at age 15 World Bank Copyright @ 2010 The International Bank for Reconstruction and Development The World Bank 1818 H Street, NW Washington, DC 20433, USA All rights reserved