75432 State Planning Organization of the Republic of Turkey and World Bank Welfare and Social Policy Analytical Work Program Working Paper Number 3: Inequality of Economic Opportunity in Turkey: An assessment using asset indicators and women’s background variables Francisco H. G. Ferreira The World Bank Jérémie Gignoux The World Bank Meltem Aran Oxford University and 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 3: Inequality of Economic Opportunity in Turkey: An assessment using asset indicators and women’s background variables Francisco H. G. Ferreira The World Bank Jérémie Gignoux The World Bank Meltem Aran1 Oxford University and The World Bank Ankara, March 2010 1 We are grateful, without implication, to Deon Filmer, Jesko Hentschel, Peter Lanjouw, David McKenzie and participants at the SPO-World Bank Social Policy Workshop in Ankara, on 22 October 2008, for comments on an earlier version of this paper. 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. Inequality of Economic Opportunity in Turkey iii Inequality of Economic Opportunity in Turkey Table of Contents Abstract..................................................................................................................................................................... v 1. Introduction........................................................................................................................................................... 1 2. Data ...................................................................................................................................................................... 2 3. Perceptions of Inequality ..................................................................................................................................... 3 4. Inequality of Opportunity for Wealth .................................................................................................................. 4 5. Opportunity Profiles for Household Wealth ........................................................................................................ 9 6. Inequality of Opportunity for Consumption ........................................................................................................ 10 7. Conclusions . ........................................................................................................................................................ 13 Appendix . ................................................................................................................................................................ 15 References . .............................................................................................................................................................. 16 Inequality of Economic Opportunity in Turkey v Keywords Inequality of opportunity, asset indicators, family background, Turkey JEL Codes D31, D63, J62 Abstract Using information on asset ownership, housing quality, and access to services to construct an indicator of household wealth, we estimate the share of inequality among prime-age Turkish women that can be attributed to unequal opportunities. Both parametric and non-parametric estimation methods are used, and robustness to some sample re- definitions is verified. We find that at least one-third (one-fourth) of overall wealth (imputed consumption) inequality in Turkey is associated with morally irrelevant, pre-determined circumstances. The circumstances that account for the largest share of the variance are rural/urban birth area and father’s education. Controlling for rural birth, parents’ education, language spoken at home, and number of siblings, a three-way regional breakdown of birthplace is not an important predictor of wealth. An opportunity deprivation profile reveals that more than two thirds of the most deprived group in Turkey consists of women born in the rural areas of the Eastern region, from mothers with no formal education. Inequality of Economic Opportunity in Turkey 1 1. Introduction Arneson, 1989, Cohen, 1989), and which regards only the former as morally objectionable. An influential 1. At first glance, economic inequality in Turkey strand of this literature actually defines “equal does not appear to be particularly high. According opportunity� as a hypothetical situation in which to Aran et al. (2008), the Gini coefficient for the there is no inequality between groups that differ only distribution of consumption per equivalent adult in terms of pre-determined, exogenous and morally was 0.31 at the time of the latest household budget irrelevant circumstances, such as race, gender, family survey, in 2006. This is in the same broad range as background, place of birth, and so on (Roemer, 1998). Greece (0.36, consumption) and the United Kingdom Inequality in any particular outcome of interest (income, (0.34, income). It is considerably higher than in the education, wealth) might be ethically acceptable, but more egalitarian countries of northern Europe, such only insofar as the outcome is independent of those as Sweden (0.25, income), and much lower than in circumstances.4 neighboring Iran (0.43, consumption), or in the vast majority of countries in Africa and Latin America.2 5. It is quite possible that it is high levels of perceived inequality of opportunity – rather than other sources 2. Nevertheless, more than 85% of Turks either of outcome inequality, such as luck or effort – which agree or strongly agree with the statement that “The Turks find objectionable. There is certainly some gap between the rich and poor should be reduced� evidence that people are more resentful of inequality in their country, and 92% feel that the State ought when it is accompanied by less mobility. Alesina, to be “strongly involved� in reducing that gap.3 As a Di Tella and McCulloch (2004) compared the effect benchmark for comparison, only about 50% of World of local inequality on “happiness� in the United Value Survey respondents in 69 countries (in the year States and Europe, and found that the evidence was 2000) felt closer to the statement “Incomes should be supportive of what Bénabou and Ok (2001) call the made more equal in my country� than to “We need “Prospect of Upward Mobility� (POUM) hypothesis: larger income differences as incentives for individual the notion that inequality is less objectionable in efforts�. societies that are perceived to be more mobile. In societies with limited mobility, inequality is seen as 3. Are Turks particularly inequality averse? Why more permanent, and thus associated with greater does inequality, albeit moderate in level, appear to be unhappiness.5 Greater inequality of opportunity is so unpopular in Turkey? A possible clue is offered by closely related to an intergenerational version of the their answers to a third question in the same opinion POUM hypothesis: people object to inequality when survey: when asked what is the “main reason why there they feel that children’s prospects in their society are are some people in need in our country today?�, 63% heavily pre-determined by their background, so that choose “injustice in society� as their answer. Do Turks there is little intergenerational mobility. object more strenuously to inequality which they see as reflecting differences in circumstances over which 6. This paper investigates the degree and nature individuals have no control than to inequality which of inequality of economic opportunity in Turkey. they perceive as arising from an individual’s own Following Roemer (1998), Bourguignon et al. (2007) efforts? and Ferreira and Gignoux (2008), we associate inequality of opportunity with the between-group share 4. There is an established tradition in moral of inequality that is estimated when the population philosophy that distinguishes inequality of opportunity is partitioned exclusively on the basis of morally from other sources of inequality (Dworkin, 1981, irrelevant, pre-determined circumstances. In the case 2 Inequality measures for other countries reported in this paragraph are Gini coefficients for per capita distributions, taken from World Bank (2005). These comparisons are merely suggestive of broad ranges, since there are important methodological differences in how the figures are calculated across countries. In particular, the figure for Turkey incorporates equivalence scale adjustments that are absent from the other numbers. 3 These numbers summarize answers to opinion questions in the Life in Transition Survey (LITS-Turkey, 2006), which are discussed in more detailed in Section 3. 4 Slightly different versions of this idea have been applied in practice to the empirical measurement of inequality of opportunity in Europe (e.g. Checchi and Peragine, 2005; Lefranc, Pistolesi and Trannoy, 2008) and Latin America (e.g. Bourguignon, Ferreira and Menéndez, 2007; and Cogneau and Gignoux, 2009). 5 See also Hirschman and Rothschild (1973) for a pioneering discussion. Inequality of Economic Opportunity in Turkey 2 of Turkey, these circumstance variables include place women in Turkey is due to unequal opportunities. The of birth (both rural-urban status and region), as well as corresponding figure for household consumption is family background (mother’s and father’s education, just over a quarter. The opportunity profile suggests language spoken at home, and number of siblings). that opportunity deprivation is particularly pronounced in rural areas of the Eastern provinces, and among 7. As in many other countries, there is a non-trivial families headed by people with mothers with no data challenge associated with this approach to formal schooling. measuring inequality of opportunity in Turkey: There is no single survey that contains satisfactory information 10. The paper is organized as follows. Section 2 briefly both on household income or consumption, on the one describes the datasets used in the analysis. Section 3 hand, and on key pre-determined circumstances on the provides a brief assessment of public opinion about other. In particular, Turkey’s Household Budget Survey inequality in Turkey, largely as a motivation for what (HBS), which contains reliable data on household follows. This discussion is based on data from the consumption, does not report information on the family Life in Transition Survey (2006). Section 4 describes background of working-age individuals. Conversely, the method for and the results of the analysis of surveys such as Turkey’s Demographic and Health opportunity for wealth acquisition. Section 5 introduces Survey (TDHS), which do report family background the concept of opportunity profiles, and presents variables for a large subset of the population, do not the wealth opportunity profile for Turkey. Section 6 contain sufficiently detailed income or consumption discusses our assessment of inequality of opportunity information. for consumption, and Section 7 concludes. 8. We employ two alternative approaches to circumvent these data limitations. First we use household wealth (measured by a principal- 2. The Data components-based index of assets and access to amenities constructed from information available in 11. The paper uses data from three separate household the TDHS) as an alternative indicator of economic surveys recently conducted in Turkey. The study status. Second, we combine information from the of perceptions of inequality briefly summarized in two key datasets, the HBS and TDHS, by imputing Section 3 rests on data from the Life in Transition consumption from the former into the latter, on the Survey (LiTS), conducted in 2006. The measurement basis of the covariance between consumption and of inequality of opportunity for wealth in Section 4 is other variables which are observed in both surveys. based on data from Turkey’s Demographic and Health To avoid underestimating consumption variance in Survey (TDHS), fielded in 2003. The discussion the TDHS because of the imperfect nature of that of inequality of opportunity for consumption, in covariance, a bootstrap prediction method developed Section 5, is based on imputing consumption from by McKenzie (2005) is used. Parametric and non- the Household Budget Survey (HBS), fielded in 2006, parametric estimation methods are used for both into the TDHS. wealth and imputed consumption, to test the sensitivity of the estimates to small cell-sizes and functional form 12. The Life in Transition Survey was carried out by the assumptions. European Bank for Reconstruction and Development in 28 post-communist transition countries in Europe 9. The resulting measures of the opportunity share of economic inequality are lower-bound estimates, since and Central Asia and in Turkey. The data were the inclusion of other circumstance variables, which collected between August and October 2006. As in the are not observed in our combined datasets, would other countries surveyed, a nationally representative further refine the population partition (Ferreira and sample of 1,000 Turkish households was interviewed. Gignoux, 2008). We also use the population partition One adult was selected at random in each household to into types (groups of individuals with identical answer a set of detailed questions on living standards, observed circumstances) to construct opportunity poverty and inequality, trust in state institutions, profiles, which rank types by the first moment of the and attitudes to market economy and democracy. frequency distribution defined over their opportunity Information on socio-economic characteristics was sets. As lower bounds, we find that approximately also collected. We focus on a set of questions regarding one third of the observed wealth inequality among perceptions of inequality and economic mobility, Inequality of Economic Opportunity in Turkey 3 and on how they relate to individual and household Turkish households. Although the LiTS, for example, characteristics such as educational attainment, native also contains a consumption expenditure module, it is language, consumption expenditures, and residence much less detailed than that in the HBS, and covers a in an urban or rural area. Unfortunately, the region of much smaller sample. residence of the interviewees was not provided in the data. 13. The paper uses data from the Turkey Demographic 3. Perceptions of Inequality and Health Survey (TDHS) fielded between December 16. The Life in Transition Survey (LiTS), which was 2003 and March 2004 by the Hacettepe Institute. briefly described above, asked a number of subjective The data were collected from a sample of 10,836 questions to a representative sample of the Turkish households, representative at the national level but population in 2006. Among these questions, four are also at the level of the five major regions of the country particularly informative of Turkish attitudes towards (the West, South, Central, North and East regions). economic inequality. The first asks for people’s views Information on basic socio-economic characteristics on the statement “The gap between the rich and poor of the population was collected for all household today in this country should be reduced�. The second members, and all ever-married women between 15 asks “Should the state be involved in reducing the gap and 49 years old answered a detailed questionnaire between the poor and the rich?� The distribution of on demography and health. 8075 women provided responses is presented in Table 1, both for the whole information. sample and by four different characteristics of the respondent: type of area (rural, urban or metropolitan), 14. Although very limited information is provided on native language, level of education, and consumption earnings and consumption, the TDHS, as other DHS expenditures. surveys, collected reasonably detailed information on certain durables owned by households, on 17. The results suggest that Turks are highly averse to housing conditions, and on access to amenities. This the inequality they observe in their society: 85.4% of information is used to construct a measure of household the people either agree or strongly agree that the gap wealth, and to study its distribution. The DHS survey between rich and poor in Turkey should be reduced. also contains information on a set of circumstance This is the single highest proportion for all 29 countries variables for the sample of ever-married women, where LiTS surveys were conducted, and compares namely the birth region, the type of area of the place with a mean proportion of 47% in those countries. This of birth, the levels of education of both the mother is particularly remarkable since the LiTS countries and father, the respondent’s mother tongue, and the include some with much higher observed levels of number of siblings.6 In the remainder of the paper, the income inequality than Turkey (e.g. Russia, with an information on these pre-determined circumstances is income Gini of 49.6).7 used to measure inequality of opportunity for wealth acquisition, for the sample of ever-married women. 18. Not only does a large majority of Turks feel that inequality is too high, but there is almost a consensus 15. The 2006 Household Budget Survey (HBS) col- that there is a clear role for state-led redistribution: lected information among a nationally representative 92% of respondents argue that the State should be sample of about 8,500 households and their members, “strongly involved� in reducing the gap between rich including gender, area type, parental education and poor. This compares with a 68% average across and father’s occupation. It is the staple survey for the 29 LiTS countries. Like the proportion who agree assessing the distribution of household consumption that the gap is too great, the share favoring state-led expenditures, and thus contains a reasonably detailed redistribution is actually somewhat higher among the questionnaire on that topic, which provides the most rich and the more educated (although both shares are reliable estimates of current living conditions for also markedly higher among ethnic minorities).8 This 6 Region was classified into three broad regions: West, Center, and East; the type of area of birth place into rural or urban according to whether the respondent considered it as a village or sub-district or not; parental education into four categories: no education or unknown level, primary, secondary, and higher education; mother tongue into Turkish or another language; and number of siblings into: less than 3, 4 to 5, 6 to 8, 9 or more. 7 Source: World Bank www.worldbank.org/depweb/beyond/wren/wnrbw_05.pdf 8 Ethnic minorities are those who do not speak Turkish at home. Inequality of Economic Opportunity in Turkey 4 (slightly) larger degree of inequality aversion among determining economic success declines (slightly) richer households is reminiscent of the pattern found with actual economic and educational achievement: for the United States using happiness data: in the US whereas 50% of the poor (and 55% of those with no (unlike in Europe), the negative effect of inequality on educational degree) favor hard work, only 45% of subjective well-being is only statistically significant the rich (and 43% of those with completed higher among the better-off. (Alesina et al., 2004). education) agree. Conversely, whereas only 8% of the uneducated feel that political connections are key 19. The remaining two LiTS questions on inequality in determining economic success, 26% of college may help shed some light on what is behind Turkey’s graduates agree.9 marked aversion to (its apparently moderate levels of) income inequality. When asked what would be the 22. But what is the objective evidence on the relative “main reason why there are some people in need in our importance of “effort and hard work�, vis-à-vis pre- country today?�, 63% of respondents choose “injustice determined circumstances, in accounting for economic in society�. If one considers that “luck� and “inevitable status in Turkey? What share of the inequality observed part of modern life� also refer to factors beyond the in Turkey is due to unequal opportunities, and what control of the individual, then a full three-quarters of share to personal responsibility and effort? This is the the Turkish population feel that the poor should not question to which we turn in the next section. be held responsible for their condition. Only 24.4% of respondents attribute poverty to “laziness and lack of will power� of the poor themselves. 4. Inequality of Opportunity for 20. Interestingly, however, perceptions about the re- Wealth lative roles of circumstances and efforts in determining 23. “Consequential� approaches to inequality of economic outcomes appear to be asymmetric in opportunity measure it in terms of its contribution to, Turkey. The fourth LiTS question we examine asks or share of, inequality in some observed outcome of what “factors are most important to succeed in life in interest, y.10 Suppose all determinants of this outcome this country�. Whereas three quarters of the sample variable can be divided into three groups: those over attributed poverty to factors other than a person’s which the individual has no control (represented by “laziness� and “will power�, only 22.2% attribute a vector of circumstances, C); those which can be economic success to “political connections� or affected by individual decisions (denoted efforts, E); “criminal or corrupt ties�. Just over 75% feel, instead, and purely idiosyncratic factors (such as luck) grouped that success is due to effort and hard work (48.4%) under a zero-mean random variable, u. Then, at a very or intelligence and skills (27.2%). It would appear general level, one can write: that Turks see economic failure and deprivation as the result of an unjust system, or bad luck, but are y = f (C , E , u ) (1) prepared to view the rich predominantly as deserving of their position. In terms of the Alesina et al. (2004) 24. Because they are, by definition, variables over contrast between American and European views – with which individuals have no control, circumstances can the former attributing economic status predominantly be treated as economically exogenous. Efforts, on the to personal desert and the latter largely to social other hand, may clearly be affected by circumstances, circumstances – Turkish attitudes would appear to be as well as by other factors grouped under v. A person’s “American upwards�, but “European downwards�. own educational attainment, for instance, is an outcome over which individuals can exercise a measure of 21. Table 1 also reveals that “American-style� op- control. It is therefore an “effort� variable. But it is timism about the role of effort and hard work in affected both by unobservable factors 9 This pattern would appear to be consistent with Bénabou and Tirole’s (2006) argument that people may be prepared to process information selectively, so as to reinforce the weight of evidence consistent with the belief that the world they live in is fair, and that they have a chance to succeed. This may be particularly relevant if they need the encouragement in order to overcome current predicaments. 10 Most empirical measures of inequality of opportunity in the literature are “consequential� in this sense. See. e.g. Checchi and Peragine (2005), Lefranc et al. (2008), etc. See, however, Barros et al. (2008) for an attempt to measure inequality in access to individual opportunities, without recourse to a relevant concept of (consequential) “advantage�. Inequality of Economic Opportunity in Turkey 5 (such as the individual’s ability) and by observable predetermined circumstances then, conversely, the circumstances, such as the educational level of one’s degree of inequality of opportunity empirically parents. Following Bourguignon et al. (2007), we can observed in a given society must be related to the then rewrite (1) as: degree to which, in practice, y is correlated with C. y = f[ C , E (C , v ), u ] (2) 28. In most empirical applications, each element of C is a discrete variable, such as race (black or white), 25. This formulation captures the idea that gender (male or female), region of birth, and so on. circumstances potentially affect final outcomes For a given vector of (such discrete) circumstances C, through two different pathways: a direct effect of define { y ik } as the partition of the population the circumstance on the outcome (controlling for such that C ik = C k ⇔ i ∈ k , k = 1,..., K. Ferreira efforts), and an indirect effect, through efforts. In this and Gignoux (2008) propose a relative measure of framework, Roemer’s (1998) definition of equality inequality of opportunity given by the simple mapping, of opportunity can be stated in very simple terms: , given by12: opportunities are equally distributed in a society if (and only if) the outcome of interest is distributed (3) independently of pre-determined, morally irrelevant circumstances: F (y C )= F (y ) , ∀C . This, in turn, implies three conditions: (i) that circumstances must have 29. Equation (3) defines a measure of inequality of no direct impact on outcome y; (ii) that efforts must opportunity as the between-group share of overall also be distributed independently of circumstances; inequality in y, where the groups are given by a full and (iii) that the random term u be orthogonal to partition of the population such that members of each circumstances.11 group have identical circumstances for all elements of C. Ferreira and Gignoux (2008) also note that, for a 26. To fix ideas, think of wages as the outcome (y); {} given partition y ik , depends on the ethnicity as a circumstance (C); and of education as specific (decomposable) inequality index I() and, an effort (E). Ethnicity might affect wages directly, for some measures, it may also depend on the path controlling for education, because, for instance, of of decomposition. Additionally, given sample size discrimination in the labor market. It might also limitations, they note that parametric estimates may affect wages indirectly, if different ethnic groups have certain advantages over the fully non-parametric have differential access to education, for any reason. decomposition.13 Roemer’s definition of equal opportunities, F (y C )= F (y ) requires that neither of these two channels operate: 30. Given that not all relevant circumstances C are there should be no differences in the conditional wage {} observed, the partition y ik is an incomplete partition distributions across ethnic groups at all, whether driven by the full set of circumstances. Our data does not by ethnicity directly (controlling for education), or contain information on how good the schools to indirectly, driven by differences in education levels which one’s parents went were, for instance. Or on the across ethnicities. quality of care women received as infants. These are relevant circumstances that lie beyond an individual’s 27. As noted by Ferreira and Gignoux (2008), to own control, but may affect their lifetime wealth or measure inequality of opportunity is then to try to well-being. If we did observe them, and were able to quantify the extent to which F (y C )= F (y ). If the ideal further partition the population into groups defined by of equal opportunities attains when the distribution of those variables, the between-group share of inequality advantage y is independent of all morally-irrelevant might rise, and could certainly not fall. is 11 See Bourguignon et al. (2007) and Ferreira and Gignoux (2008) for more detailed discussions of this Roemer-based approach to the measurement of inequality of opportunity. 12 The same authors also define an absolute measure of inequality of opportunity, : 13 The between-group share defined by (3) corresponds to a standard decomposition of inequality by population subgroups, which uses overall inequality among individuals as the denominator. An alternative decomposition, proposed by Elbers et al. (2008), adjusts the reference inequality (the denominator) to take into account the number and relative sizes of groups in the partition. This alternative approach is specially well-suited to identifying the most salient cleavages in a particular society. While we find it less satisfactory as a lower-bound measure of inequality of opportunity – precisely because both the numerator and the denominator are sensitive to the design of the partition – future research should investigate its uses in describing the profile of opportunity. Inequality of Economic Opportunity in Turkey 6 therefore a lower-bound on the actual share of 33. Table 2 describes the elements underpinning between-group across all possible circumstances. This Turkey’s household wealth index, by listing each is also true of the parametric estimation approach, element of the vector x, as well as its mean and standard which we discuss in more detail below: if additional deviation. The last column presents the scoring factor circumstance variables were to be added to a regression for each element of x in the TDHS sample (the vector specification, the R2 might rise, but would not fall. a), divided by the standard deviation. The standard interpretation is that a yields the set of weights 31. In this paper, we apply this decomposition to a providing the maximum discrimination between household wealth index, constructed on the basis of households in the sample, in terms of their ownership information contained in the TDHS 2006. This choice of these particular assets (x).16 of outcome variable y takes the following factors into consideration. First, Turkey’s household budget 34. McKenzie (2005) lists a number of reasons why survey (HBS), from which a reliable consumption an asset index such as this might in fact be preferable aggregate can be constructed, does not contain to consumption or income as a basis for inequality information on some of the most important candidate measurement, including the likelihood that recall bias circumstance variables, such as the education of the might be smaller for asset ownership questions than for father and mother of present-day workers. Second, some income or expenditure questions. But McKenzie (2005) also highlights two potential pitfalls in using the consumption aggregate available from the LiTS asset indices, namely the possibilities of truncation and suffers from two serious shortcomings: it is compiled clumping. Whereas truncation would most likely arise from a highly aggregated consumption questionnaire, from not observing assets capable of distinguishing and is available for a relatively small sample. Third, either the very poor from those just above them, or the wealth indices constructed from DHS information on very rich from those just below them, clumping might the ownership of durable goods (such as fridges, TV be caused by using too few assets, leading to “false sets, cars, computers, etc.), on housing characteristics modes� in the distribution, arising from insufficient (such as the type of roof materials and floor cover), discriminating power in the index. Figure 1 plots the and on access to utilities (such as water and sanitation) superimposed histogram and kernel density estimate have been widely used in estimating household welfare of our asset index for Turkey, revealing the absence of and in ranking households for targeting purposes.14 either clumping or truncation. 32. Following Filmer and Pritchett (2001), we define 35. Once we are satisfied that the distribution of y our wealth index as the first principal component constitutes an appropriate basis from which to estimate of a vector of assets x (including durables, housing inequality of opportunity, the next problem is to characteristics and utility access indicators) owned by choose a suitable inequality index, I(), for computing households in the TDHS sample.15 For each household through equation (3). By construction, y is i, the index is therefore given by: distributed with mean zero and a variance equal to the largest eigenvalue in the correlation matrix of x. (4) As noted by McKenzie (2005), these properties mean that most standard inequality measures routinely where the p-dimensional vector a is chosen so as used for income or consumption are unsuitable to maximize the sample variance of y, subject to for the wealth index y. A zero mean impedes .s denotes a standard deviation, and the computation of most relative inequality measures overbar denotes a mean. 14 See Filmer and Scott (2008) for a recent (and sanguine) assessment of the robustness of household rankings based on asset indices originating from DHS information, when compared, inter alia, to detailed consumption expenditure data. 15 Like Filmer and Pritchett (2001), we have treated each and every category of our housing characteristic and utility access variables as an independent dummy variable. In some cases, such as access to sanitation or water sources, there is arguably an ordinal nature to the alternative categories, and it may be statistically preferable to treat those variables explicitly as ordinal in the analysis. See Kolenikov and Angeles (2009). This alternative treatment is left for future work. 16 The TDHS data files contain a pre-constructed asset index, supposedly also given by (4). As the survey documentation does not describe the details of how that index is constructed, best research practice generally involves computing the index from the underlying data, as we have done here. The correlation coefficient between our index and the TDHS index is 0.94, and the kernel density functions for both indices are very similar, although the kernel for our index is considerably smoother. Inequality of Economic Opportunity in Turkey 7 (which generally divide by the mean), including the while to estimate (5) parametrically. This is done by Gini coefficient and all members of the Generalized imposing a functional form assumption on equation Entropy class. Negative values are problematic for (2), such as: logarithm-based measures (such as the Theil indices, (6) the variance of logarithms, and many others.) The reduced form of (6) is 36. For our purposes, the simplest solution is to which can be estimated by OLS as revert to the variance, which is straightforwardly (7) decomposable and is also translation invariant.17 Our proposed measure of the opportunity share of inequality in wealth is thus given by: 40. Under the maintained functional form assumptions in (6), a parametric estimate of the opportunity share of inequality is given simply by the R2 of (7). Once (5) again, is the lower-bound estimate on the set of possible estimates of the share of circumstances. If which is a specific version of (3), when an additional element of C, which is presently omitted, were to become observable, the R2 of (7) might rise, 37. Since , but it would not fall. That is all that is needed to define this measure as a lower-bound, since (7) is a reduced- it is clear that (5) corresponds to the between- form regression that is intended to capture the effect group share in a standard variance decomposition. of circumstances both directly and through any effort Furthermore, since the weights in both the within-group variable whatsoever. and the between-group terms are simple population shares, and do not include income levels or shares, 41. To qualify as a “circumstance� in Roemer’s sense, (5) describes a path-independent decomposition in the variables must (i) be likely determinants of advantage Foster-Shneyerov (2000) sense.18 y, either directly or through their effect on efforts; and (ii) be impossible for the individual himself to 38. Equation (5) can be computed non-parametrically affect by choice. Given the information available in from partition . All that is required is the population the TDHS, our vector of circumstances consists of share and mean wealth index for each cell of the information on the type of area in which the woman partition, as well as the overall mean and variance for was born, the region where she was born, her mother’s the complete sample. However, as the dimension of and father’s levels of education, the reported mother the circumstance vector C, and the number of discrete tongue, and the number of siblings the individual had values that each element Cj can take (#Cj), rise, the at birth. The discrete categories for each variable, as number of cells in the partition increases geometrically. well as the distribution of the population across them, If the dimension of C is J, the number of cells in are reported in Table 3. is given by Naturally, for a given sample size, the precision of the estimates of group means will 42. Table 4 reports the results of regression (7) of the fall as J and #Cj rise. wealth index on circumstances. Since this is a reduced- form regression, coefficients should not be interpreted 39. If the number of cells with fewer than 10 causally. They reflect partial correlations between observations or so is non-trivial, it becomes worth individual circumstance variables and the household’s McKenzie (2005) also effectively reverts to the variance, although he is interested in the inequality of different sub-regions relative to total inequality, 17 and therefore defines a ratio of standard deviations to the overall standard deviation. Since our measure of inequality of opportunity (in equation 3) is by construction a ratio of inequality measures, we define the inequality index in outcome space as the unadjusted variance. The problem of scale dependence will vanish for the opportunity index, and the (related) issue of mean dependence would seem to be of no import for a variable that has mean zero by construction. In other words, if one were to define a “smoothed� distribution 18 , corresponding to a particular partition as the distribution that arises from replacing with the group-specific mean , and a “standardized� distribution as the distribution that would arise from replacing with (where μ is the overall mean), then Inequality of Economic Opportunity in Turkey 8 wealth index, conflating both direct and indirect effects. 46. The bottom panel of Table 5 reports the partial Nevertheless, the regression is informative. The share shares of each individual circumstance (J) included in of explained variance, , is 31%. Being the partition, defined as: born in an urban area, and having Turkish as mother tongue, are significantly associated with subsequent (9) wealth. So is having educated parents and, in contrast 47. Inspection of (9) immediately reveals that, for any to results for Latin America, father’s education appears given partition, these partial shares sum up to the overall to be a stronger predictor of wealth than the mother’s.19 parametric estimate of between-group inequality, given Being one of many children in the family is associated by (8). Besides this attractive additive decomposability with lower wealth. Perhaps most interestingly, once property, this definition of circumstance-specific shares these other circumstances are controlled for, there is also satisfies the Foster-Shneyerov path-independence no significant association between birth region (at the property. Although we have already noted that the three-region level) and future wealth. overall non-parametric decomposition (5) is path- 43. Our estimates of the opportunity share of inequality independent by construction, parametric estimation of in household wealth among women in Turkey are the partial shares – based respectively on the smoothed presented in Table 5. The first column reports results and standardized distributions – are not the same. for the full TDHS sample of ever-married women aged However, as we show in the Appendix, equation (9) 15-49. The second column restricts the sample to ever- is the simple average between the direct and residual married women aged 30-49, so as to eliminate some estimates of the partial shares, which correspond to those smoothed and standardized distributions, of the life-cycle inequality, which is arguably purely respectively. (9) is therefore a simple example of a transitory in nature. The third column repeats the Shapley-value based decomposition, where averaging analysis for the pre-constructed TDHS wealth index, across alternative paths eliminates path-dependence. discussed in footnote 15. See Shorrocks (1999). 44. The first line simply reports the total variance in 48. Unsurprisingly, the partial shares echo some of the the wealth index, . The second line reports the preliminary findings from Table 4. Whether a Turkish non-parametric estimate of between-group inequality, woman is born in an urban or rural area appears to be a given by equation (5), while the third line gives its powerful predictor of her likely household wealth as an parametric analogue, the R2 of (7), or: adult. More than a third of the overall (lower- bound) (8) opportunity share of inequality is accounted for by this circumstance alone. Father’s education and mother’s 45. The non-parametric estimates are consistently education follow, in that order. Taken together, they are higher than the parametric ones. For our own asset roughly of the same magnitude as rurality in accounting index, the non-parametric estimate of the opportunity for the overall share: just over a third. Mother tongue share of inequality is 35% for the full sample, and 37% and number of siblings follow. The number of siblings for the sample of ever-married women aged 30-49. results, with roughly 10% of the share of overall The parametric estimates are roughly four percentage variance accounted for by circumstances, is not points lower in each case. As discussed earlier, these trivial, particularly when considering that this is after differences are consistent with the expectation of large controlling for the education of both parents, as well sampling variances around the estimated cell means in as the geography of birth. (5), owing to the fine partition of a finite sample. Since the exercise aims to estimate lower bound measures of 49. As before, and despite the salience of regional inequality of opportunity as a share of observed wealth differences in the literature on Turkey, the three-way inequality, we choose the parametric estimates in line (East, Center, West) partition of the country has no 3 as our benchmark result. This yields a tight range importance in accounting for differences in opportunity of 31% - 32%, robust to considering only the adult for wealth, once the other determinants have been women. controlled for. 19 Although this may be because the Latin American regression contained information on the father’s occupation as well. See Ferreira and Gignoux (2008). Inequality of Economic Opportunity in Turkey 9 5. Opportunity Profiles for Household than in that for native Turkish-speakers. There are Wealth also substantial differences in skewness and kurtosis, particularly in the urban/ rural, regional, and native 50. The partition of the population into circumstance- language partitions.21 homogeneous groups (called types by Roemer, 1998), which was used above to compute a lower-bound 53. Looking at the overall shape and position of these measure of inequality of opportunity for wealth, conditional distributions, some other differences can also be used to shed light on the distribution of across social groups are also clearly apparent. There opportunities among Turkish women in a more direct are large wealth gaps between women with uneducated and disaggregated manner. As noted by Ferreira mothers and those whose mothers have completed and Gignoux (2008), each cell in the partition of either primary or higher levels of schooling, although the population implemented in Section 4 (such that there is little positional difference across the latter two ), corresponds to a distributions. In contrast, there are clear positional Roemerian type . We have seen that equal differences across all three distributions conditional on opportunities attain when , which father’s education. The distribution for native Turkish is a different way of writing for a discrete speakers lies well to the right of that for non-Turkish partition. Differences in the outcome distributions speaking minorities. As expected from the results among types, therefore, are taken to reveal (or arise in Table 5, those who were born in urban areas are from) inequality of opportunity. considerably wealthier than those born in rural areas. Women who grew up in large families (with six or more 51. It follows that to plot the conditional wealth children) tend to fare less well than those who grew up distributions, , for each type should be an in smaller households. By each of these circumstance- informative way to graphically depict inequality of specific cuts, it is clear that the conditional distributions opportunity. The cardinal measures presented in are not identical: social background – as measured Section 4 rely fundamentally on differences across by predetermined circumstances that individuals conditional means. A complementary exercise is themselves can not be held responsible for – does to visually inspect differences across the entire powerfully affect the distribution of opportunities distributions. Because of sample size restrictions, faced by individual Turkish women. it is impossible to estimate density or distribution functions for all 768 types used in our decomposition. 54. At least conceptually, it is not unreasonable to see But it is still informative to look at more aggregated the support of such a conditional distribution as an conditional distributions, where the population is individual i’s ( ) opportunity set for outcome y, decomposed into groups by one specific circumstance and as the probability distribution associ- at a time. The results, in the shape of kernel estimates ated with the opportunity set. After all, given i’s of the conditional density functions, are presented in circumstances , only i’s own choices, effort and Figure 2.20 luck will determine his final position, . If it were possible, therefore, to rank across 52. Figure 2 reveals that the conditional wealth k in a meaningful way, we would obtain a ranking distributions differ across social groups not only in of opportunity sets across types, which Ferreira and means, but in other moments and in general shape Gignoux (2008) call an opportunity profile. as well. The distributions for women born in rural areas (and in the East) have visibly larger dispersion 55. At the level of disaggregation implicit in Figure than the distributions of those born in urban areas 2, one can look for robust rankings across conditional (and in the Central and Western regions). Similarly, distributions by means of stochastic dominance there is greater dispersion in the distribution of those relationships (see Lefranc et al., 2008). However, who grew up in non Turkish-speaking households such broad groupings may be less useful for policy 20 Diagnoses of inequality of opportunity by means of stochastic dominance test across such conditional distributions have been used in the literature. See Lefranc et al. (2008). 21 Throughout this subsection, the sample is restricted to households containing ever-married women between ages 30-49. This is so as to minimize possible selections biases arising from more frequent early marriages among rural women in the Eastern part of the country. Inequality of Economic Opportunity in Turkey 10 makers interested in identifying pockets of exclusion education level is fairly low even among the most than a distributions cannot be plotted and stochastic advantaged groups, as a result of the overall low dominance relationships cannot be established, because educational attainment of women: more than 70% of of data insufficiency, the types can still be ranked by the top decile in terms of opportunities had mothers a particular moment of their conditional distributions. who had no schooling or had only primary schooling. While this is certainly less robust than a dominance- based ranking, there are offsetting gains in terms of 59. The contrast with those in the bottom opportunity the ability to generate a complete ranking of types by decile could not be starker: 97% of the women in this their opportunity sets, and in terms of a much sharper group were born in rural areas, and 89% of them were description of the disadvantaged groups. born in the East of the country. 97% had uneducated mothers (as compared to 7% among their advantaged 56. To explore this option, we follow Ferreira and counterparts), and 81% had illiterate fathers as well. Gignoux (2008) and rank each type in our fine partition 91% hailed from non-Turkish speaking households, by the mean of its conditional wealth distribution. Once and only 4% came from relatively small families (with types are so ordered, the circumstances which define fewer than three siblings). Clearly, when Turkish them constitute an opportunity profile. Table 6 lists the households are ranked by the mean wealth indices of circumstances that define those types that make up the the types they belong to, there are very stark differences bottom tenth of the opportunity distribution. In other in the degree of their access to opportunity for wealth words, it lists the individual types with the lowest mean acquisition. levels of wealth (by asset index) in the population, until the cumulative population share reaches 10%. A full 66% of the people in this group live in households 6. Inequality of Opportunity for whose ever-married woman was born in the rural areas Consumption of the eastern provinces, to an uneducated mother, and in a non-Turkish speaking household. These women 60. Although the distribution of wealth, as measured almost always had more than three siblings, and a by an asset index, is of intrinsic interest, it is unlikely father with either none or primary schooling only. to be the best description of the distribution of current economic well-being. Among other concerns, the 57. This is a heavy concentration of deprivation among asset index contains no information on liabilities, and a fairly specific social group, which is defined by a few is thus a better indicator of gross than of net wealth. observable characteristics over which they have no The distribution of consumption expenditures is control. It may present analysts and policymakers in likely to be a better guide to the distribution of well- Turkey with a reasonably clear picture of which groups being, and it thus provides an alternative angle on enjoy the least opportunity for acquiring wealth, based economic opportunity in Turkey. However, as noted on exogenous circumstances of their birth. in the introduction, the Household Budget Survey, which contains a good measure of consumption, does 58. Table 7 compares characteristics for the bottom not report information on a number of important and top tenths of the opportunity profile defined over circumstance variables. In this final section, we the population in our TDHS sample (of households follow a simple statistical procedure to combine containing 30-49 year old ever-married women). information on circumstances from the TDHS survey In other words, the table reports the circumstance with information on consumption from the HBS. composition of the top and bottom deciles of the Ultimately, since the link between the two surveys distribution when households are ranked by the mean is provided largely by components of the asset index wealth level of their types (as in Table 6). According to (and a few additional covariates), the exercise can also Table 7, 99% of those women in the most advantaged be seen as an alternative way of using information group were born in urban areas, and 62% were born on assets to measure inequality of opportunity in in the western provinces. 85% of them had fewer Turkey. Our approach here closely follows McKenzie than 3 siblings, and 98% had Turkish as their mother (2005) in imputing consumption from the HBS into tongue at home. It is interesting to note that mother’s the TDHS, using a bootstrap prediction method.22 This approach is a simplified version of the consumption imputation procedures proposed by Elbers, Lanjouw and Lanjouw (2003). 22 Inequality of Economic Opportunity in Turkey 11 61. We are interested in examining the relationship G = 10 groups, defined according to the deciles of the between the distribution of consumption c and a vector distribution of the first principal component (the wealth of circumstances C. The TDHS contains information index) y for the set of wealth indicators common to the on circumstances, but no comprehensive information two surveys.24 Separate distributions of the predicted on consumption. However, the HBS contains detailed residuals are identified for each of the 10 groups. information on consumption, and both surveys collected a common set of information on ownership (3) The sample of the DHS survey is then divided of durable goods, housing characteristics, and access into the same 10 groups, using the same cut-off values to utilities. The HBS may then be used as an auxiliary for y as in the auxiliary sample. survey to impute consumption into the main survey, (4) For each household i in group g in , a residual the TDHS, from the common information on wealth is drawn from the empirical distribution of residuals indicators and a set of demographic and other controls. for households in group g in , The predicted value The imputation may be implemented in different ways, of per capita consumption is given by: but the bootstrap prediction method appears to be most reliable for studies of inequality (McKenzie, 2005). This procedure consists in combining a direct prediction based on a regression model, with a repeated draw of (11) residuals comparable to a bootstrap. The relationship (5) Measures of inequality of opportunity are between wealth indicators X and per capita consump- computed using the predicted distribution of per capita tion c is estimated, on sample (from the auxiliary consumption. HBS survey), using a log-linear regression model: (6) Following the bootstrap principle, steps (4) and (10) (5) are repeated for a number R of drawn replicate distributions of the residuals, and the measures of where are demographic controls. The estimation inequality of opportunity are computed as the mean of (10) provides the fitted coefficients and as over the measures obtained for each replication. In our well as estimated residuals . In order to reproduce analysis, we use R=20 replications. This replication the observed levels of inequality, the prediction of per process allows averaging out the bootstrap sampling capita consumption in sample (from the “main� error. DHS survey) is constructed by adding the linear prediction of per capita consumption, and 62. The opportunity share of consumption inequality a prediction of the residual . (A traditional direct may be obtained through the same equation (3), as for prediction, consisting exclusively of the linear wealth inequality. However, the choice of a suitable prediction, would underestimate actual inequality.) inequality index, I(), for computing might The predicted residual is drawn, for the sample differ. Imputed consumption ci takes only positive of the main survey, from the empirical distribution values, so that members of the generalized entropy class of residuals obtained in fitting (10) to the auxiliary (GE) can be computed. Their main advantage over the sample . Following McKenzie (2005), the procedure variance, in this case, is that the distributions of imputed allows for heteroskedasticity by drawing from the consumption do not have mean zero by construction, distribution of residuals for households with similar so that mean- or scale-independence becomes, once assets.23 This is done through six steps: again, a desirable property for I(). All GE indices also satisfy the additive decomposability property required (1) The regression in (10) is carried out using the to identify the share of between-circumstance groups common set of wealth indicators, and the parameters inequality. However, in this class, only the mean log and residuals are obtained. deviation allows for a path-independent decomposition in the Foster-Shneyerov sense because of the income (2) The sample of the HBS survey is divided into weights in the decomposition.25 Using this index, 23 Heteroskedasticity might stem from a non-linear relationship between log consumption and wealth assets X and also from the higher noise in this relationship for richer households than for poorer and middle-class ones. 24 We partition the sample into 10 groups in order to allow a sufficiently high degree of heteroskedasticity and keep group sizes of the order of a few hundreds observations. 25 See Foster and Shneyerov (2000) and Ferreira and Gignoux (2008) for discussions of this point. Inequality of Economic Opportunity in Turkey 12 our measure of the opportunity share of inequality in other countries, the coefficient on the gender of the consumption is given by: head is insignificant (which is generally thought to reflect the endogeneity of headship status). Levels of consumption are also higher in urban areas. (12) 65. In order to identify the G = 10 groups with comparable levels of wealth, the first principal As in Section 4, we compute this share non- component for the set of wealth indicators is computed parametrically (using equation 12), as well as using both samples for the HBS and DHS data. The parametrically. The parametric estimate uses a log deciles of this wealth indicator in the HBS auxiliary linear specification of the relationship between sample are used to identify the groups in the DHS circumstances and per capita consumption: main sample. (13) 66. Per capita consumption is then imputed using the 63. Under these functional form assumptions, a fitted coefficients and presented in Table 8 and parametrically-standardized distribution is estimated the draws of the residuals. The descriptive statistics by , and a parametric alternative in Table 8 suggest that the set of regressors used for the imputation have similar distributions in the two to is given by:26 samples.27 Figure 3 depicts kernel density estimates of the distributions of total household consumption (14) observed in the auxiliary HBS sample and imputed in the main TDHS sample.28 The two distributions have 64. The set of wealth indicators common to the DHS reasonably similar shapes, and the levels of inequality and HBS surveys contains 14 variables for ownership in actual consumption in the HBS and in imputed of durable goods, and four variables for housing consumption in the TDHS are also close: for the characteristics and access to utilities. We also use a sample of 30-49 year-old women, the E(0)s are 0.337 variable indicating the ownership of agricultural land, and 0.360 respectively. and nine variables for demographic controls. Table 8 presents descriptive statistics for those variables in the 67. For each of the 20 draws, a set of parametric and two samples. The results for the regression of total nonparametric indexes of inequality of opportunity consumption on assets using the HBS data are then are computed. These measures are computed for both presented in Table 9. We use a log linear specification samples of 15 to 49 and 30 to 49 ever-married women. because of the likely nonlinear relationship between Table 10 presents the regression estimates of imputed the ownership of assets and consumption. The per capita consumption on the same set of circumstance R-squared for the regression is 0.53. The coefficients variables used for the analysis of opportunities in for the asset variables are all highly significant and wealth. The R-squared for this regression is 0.26. have the expected signs: ownership of any durable is The coefficients on the circumstance variables are positively correlated with consumption, as are access all significant and have the expected signs: per capita to piped water, the number of rooms, and ownership consumption is higher for individuals born in a urban of land. As for the demographic controls, household area, those living in the West and Center regions, size is positively and the number of children Turkish native-speakers; it increases with parental negatively correlated with consumption. An inverted education and decreases with number of siblings. U-shaped relationship is observed between the age of household head and consumption, and a convex 68. Table 11 gives the opportunity shares of positive relationship attains for education. As in many consumption inequality for the two samples: ever- 26 A parametrically smoothed distribution can also be computed. Significant differences are found only for the share of urban residence because of the difference in the definitions of urban areas in the two surveys 27 (agglomerations with 20,000 inhabitants for the HBS survey and 15,000 for the TDHS one), and access to piped water (the definition is more restrictive in the DHS). The distribution of imputed consumption in the TDHS that is shown corresponds to the first one of the R=20 draws. 28 Inequality of Economic Opportunity in Turkey 13 married women aged 15-49; and ever-married women circumstances that the inequality shares we estimate aged 30-49. Reassuringly, the overall levels of must be interpreted as a lower bound. Even with inequality in per capita consumption (measured by such a limited set of circumstances, however, there the mean log deviation), are 0.35 and 0.36 in the two is no single household survey dataset in Turkey that samples. Parametric and non-parametric estimates for contains reliable information on both these variables the total share of inequality of opportunity, presented and on consumption expenditures. Consequently, we in the second and third lines, indicate that lower have used information on household asset ownership, bounds for the opportunity shares of inequality in housing characteristics, and access to amenities to consumption are comprised between 25% and 29% construct a composite asset indicator that has frequently been used as a measure of household wealth. for the broader sample (with life-cycle variations potentially introducing a downward bias for this 72. Given the statistical properties of the composite sample) and between 27% and 32% for the sample of asset index, we use a path-independent variance women aged 30 to 49. These results suggest that the decomposition to calculate the opportunity share of opportunity shares of consumption inequality are about inequality. The standard non-parametric estimates of five percentage points lower than the corresponding this share for Turkey are 35% for the full sample, and shares for wealth inequality. 37% when we focus on a more restricted age range, to eliminate some of the variation due only to the 69. The next lines give estimates for the partial shares life-cycle. The parametric estimates, based on the of inequality associated with each circumstance R2 of a reduced-form regression of the wealth index variable. As in the case of wealth, birth in an urban or on the observed circumstances, are 31% and 32% rural area and parental education are the circumstances respectively. Although the non-parametric estimates associated with the largest shares, between 9 and 11%, have the advantage that they impose no arbitrary of inequality of opportunity in consumption. Mother functional form, they suffer from a potential upward tongue and number of siblings also capture significant bias arising from imprecision in the estimation of shares, at about 5%, of consumption inequality. These conditional means (as cell sizes fall). We treat our results are very close to those obtained for wealth. parametric estimates as conservative estimates of the lower-bound share of inequality of opportunity in Turkey. 7. Conclusions 73. The parametric estimates have the additional 70. Inequality may be found more objectionable when advantage that they are additively decomposable it originates from exogenous differences in people’s into partial shares corresponding to each individual initial circumstances, rather than from differences in circumstance variable. These shares indicate that the their relative levels of effort or responsibility, or in the rural or urban status of a woman’s birthplace accounts wisdom of their choices. Using both non-parametric for the largest component of inequality of opportunity and regression-based techniques, this paper has sought in Turkey – a third of the overall opportunity share. to estimate a lower bound for the share of overall Rural status is followed by father’s education; mother’s economic inequality among women in Turkey that education; mother tongue and number of siblings, in is due to those exogenous circumstances. Following that order of importance. Roemer (1998) and Bourguignon et al. (2007), this share is interpreted as a measure of inequality of 74. Interestingly, once these various characteristics economic opportunity in Turkey. are controlled for, the broad geographical region in which a woman was born (Eastern, Central or 71. Our lower bound estimate relies on a small set Western) accounts for almost no variance. Since wealth of observed personal characteristics which can be distributions do differ substantially across these regions confidently interpreted as completely independent (as do consumption and education levels), this finding of individual choices: region and area of birth, the suggests that such differences are due to heterogeneity educational attainment of both parents, mother tongue, in the composition of the population across regions, and the number of siblings a person grew up with. in terms of the other circumstances, rather than to any It is precisely because this is an incomplete set of intrinsic regional effects. Inequality of Economic Opportunity in Turkey 14 75. All of these findings, which are based on an as opportunity-unequal as the most unequal countries analysis of variance of the composite asset index, are in Latin America, such as Guatemala and Panama. robust to an alternative empirical strategy, in which It ranks alongside the lower end of the inequality of we imputed household consumption levels from the opportunity spectrum in Latin America, with Colombia Household Budget Survey into the Demographic being a good comparator. and Health Survey. Although the overall opportunity shares of inequality were somewhat lower for imputed 77. The paper also explored the opportunity profile consumption, at 25% - 26% of the mean log deviation, for Turkey, constructed by ranking circumstance- the partial ranking of circumstances was identical: homogeneous types of households by their mean rural or urban birth status; father’s education, mother’s wealth levels. Once households are so ranked, the education, mother tongue, number of siblings and birth bottom 10% of the distribution is 97% rural and 88% region. Eastern (by birth). 91% of them hail from non-Turkish speaking households, and 97% had mothers with no 76. In addition to providing some sense of qualitative formal education. A full 66% of this opportunity- robustness of the results, the consumption-based deprived decile (of woman aged 30-49) has the analysis also permits a limited degree of international following combination of characteristics: she was born comparability. Using the lower-bound parametric in a rural area of an Eastern province, to an illiterate estimates of the opportunity share of consumption mother, and in a non-Turkish speaking household. inequality calculated by Ferreira and Gignoux (2008) The contrast with the top decile in the opportunity for five Latin American countries, we can place the distribution was striking along every dimension. Turkish results into some context. The overall shares for the mean log deviation were 24% for Colombia, 78. Such marked differences in economic opportunity 32% for Ecuador, 34% for Peru, 39% for Panama, across groups defined by morally irrelevant and pre- and 50% for Guatemala. Given methodological determined characteristics might explain, at least in differences, comparisons between Turkey and these part, why Turks appear relatively inequality averse, other countries should not be over-interpreted or despite a middling position in the world’s ranking of emphasized. In particular, the Turkish results are consumption inequality. Perhaps more importantly, based on imputed, rather than observed consumption. the opportunity profile of social groups, constructed Also, the Latin American study did not include on the basis of these pre-determined circumstances, number of siblings as a circumstance, but did use may be useful to Turkish policymakers as they seek father’s occupation. Nevertheless, as an indication of to target scarce resources and policy attention with the rough relative position, it is clear that Turkey is not aim of fostering a more inclusive growth process. Inequality of Economic Opportunity in Turkey 15 Appendix 1. Table 5 reports partial shares of inequality of opportunity, associated with each individual element Cj of the vector of circumstances C. These partial shares, which are computed through equation (9), using the regression coefficients from (7), have the attractive property that they sum up to the total share of inequality of opportunity computed through equation (8), using the same regression coefficients. 2. This appendix shows that (9) is a simple average of the two alternative paths of the variance decomposition. It therefore corresponds to the Shapley value decomposition proposed by Shorrocks (1999). This explains its additive decomposability. Recall that (7) Therefore (A1) 3. The partial contribution of a particular circumstance CJ to var (y) can be calculated in two alternative ways. Both focus on the first two terms in (A1), i.e. set 4. var (e) = 0. The direct estimate holds all constant in (A1), and computes the remaining variance as a share of the total: (A2) 5. The indirect, or residual, estimate takes holds CJ itself constant, and takes the difference between var (y) and the ensuing variance: (A3) Taking the average between (A2) and (A3) yields (9): Inequality of Economic Opportunity in Turkey 16 References Elbers, Chris, Jean O. Lanjouw and Peter Lanjouw (2003): “Micro-level Estimation of Poverty and Inequality�, Econometrica, 71 (1): 355-364. Alesina, Alberto, Rafael Di Tella, and Robert Elbers, Chris, Peter Lanjouw, Johan Mistiaen and Berk MacCulloch (2004): “Inequality and happiness: Özler (2008): “Reinterpreting between-group are Europeans and Americans different?�, Journal inequality�, Journal of Economic Inequality, 6 (3): of Public Economics, Elsevier, 88(9-10): 2009- 231-245. 2042. Ferreira, Francisco H. G. and Jeremie Gignoux (2008): Aran, Meltem, Sirma Demir, Ozlem Sarica and Hakan “The measurement of inequality of opportunity: Yazici (2010): “Poverty and Welfare Changes in theory and an application to Latin America�, Turkey 2003-2006�, State Planning Organization Policy Research Working Paper Series, 4659, The of the Republic of Turkey and the World Bank World Bank. Welfare and Social Policy Analytical Work Program Working Paper Number 1. Filmer, Deon and Lant Pritchett (2001): “Estimating Wealth Effects Without Expenditure Data – or Arneson, Richard (1989): “Equality of Opportunity Tears: An application to educational enrollments for Welfare�, Philosophical Studies, 56: 77-93. in states of India�, Demography, 38 (1): 115-132. Barros, Ricardo P., Jose Molinas Vega and Jaime Filmer, Deon and Kinnon Scott (2008): “Assessing Saavedra (2008): “Measuring Inequality of asset indices�, Policy Research Working Paper Opportunities for Children�, World Bank Series, 4605, The World Bank. (Washington, DC) mimeo. Foster, James and Artyom Shneyerov (2000): “Path Bénabou, Roland and Efe A. Ok (2001): “Social Independent Inequality Measures�, Journal of Mobility And The Demand For Redistribution: Economic Theory, 91: 199-222. The Poum Hypothesis�, The Quarterly Journal of Hirschman Albert O. and Michael Rothschild (1973): Economics, MIT Press, 116 (2): 447-487. “The Changing Tolerance for Income Inequality in the Course of Economic Development�, The Bénabou, Roland and Jean Tirole (2006): “Belief in a Quarterly Journal of Economics, 87(4): 544-566. Just World and Redistributive Politics�, Quarterly Journal of Economics, 121 (2): 699-746. Kolenikov, Stanislav and Gustavo Angeles (2009): “Status Measurement with Discrete Proxy Bourguignon, François, Francisco H.G. Ferreira Variables: Is principal component analysis a and Marta Menéndez (2007): “Inequality of reliable answer?�, Review of Income and Wealth, Opportunity in Brazil�, Review of Income Wealth, 55 (1): 128-165. 53 (4): 585-618. Lefranc, Arnaud, Nicolas Pistolesi and Alain Trannoy Checchi, Daniele and Vitoroco Peragine (2005): (2008): “Inequality of Opportunities vs. Inequality “Regional Disparities and Inequality of of Outcomes: Are Western Societies All Alike?�, Opportunity: the Case of Italy�, IZA Discussion Review of Income and Wealth, 54 (4): 513-546. Paper, 1874/2005. McKenzie, David (2005): “Measuring Inequality Cogneau, Denis and Jérémie Gignoux (2009): with Asset Indicators�, Journal of Population “Earnings Inequality and Educational Mobility in Economics, 18: 229-260. Brazil over Two Decades�, Chapter 2 in S. Klasen Roemer, John E. (1998): Equality of Opportunity. and F. Nowak-Lehmann (eds.) Poverty, Inequality, Cambridge, MA: Harvard University Press. and Policy in Latin America (Cambridge, MA: MIT Press). Shorrocks, Anthony (1999): “Decomposition Procedures for Distributional Analysis: A Cohen, Gerry A. (1989): “On the Currency of unified framework based on the Shapley Value�, Egalitarian Justice�, Ethics, 99: 906-944. University of Essex, mimeo. Dworkin, Ronald (1981): “What is Equality? Part 2: World Bank, (2005): World Development Report Equality of Resources�, Philosophy and Public 2006: Equity and Development. Washington, DC: Affairs, 10 (4): 283-345. The World Bank and Oxford University Press. Inequality of Economic Opportunity in Turkey 17 Figure 1: The Household Asset Index for Turkey: density Distribution of the asset index (with weights) .15 .1 Density .05 0 -10 -5 0 5 10 Asset index Figure 2: Household Wealth Distributions for Different Circumstance Groups in Turkey: Kernel Density Estimates Kernel density estimates for the conditional distributions of wealth. Source data: Turkey TDHS 2003 ever-married women 30 to 49 years old. Inequality of Economic Opportunity in Turkey 18 Figure 3: Distribution of household annual expenditure observed in HBS 2003 and imputed in TDHS 2003 Table 1: Perceptions of the Magnitude and Nature of Inequality in Turkey Overall By type of area By native language By level of education By level of expenditures metropolitan urban rural Turkish Other no degree primary secondary professional or higher Poor Intermediate Rich "The gap between the rich and the poor today in strongly disagree 2.5 3.5 3.1 0.8 3.0 0.6 1.0 2.2 3.6 5.5 2.3 2.9 2.4 this country should be reduced." disagree 3.4 3.7 2.4 4.0 3.8 1.7 2.8 4.4 1.2 5.2 2.6 3.0 4.4 neither disagree not agree 6.3 5.6 4.3 9.3 6.7 4.8 8.7 5.8 5.6 4.9 8.7 3.7 6.7 agree 18.2 16.3 19.8 18.5 18.1 18.2 24.0 16.2 20.2 7.0 23.9 16.7 14.7 strongly agree 67.2 69.0 67.8 64.7 66.0 72.8 59.6 69.0 68.3 76.1 58.9 71.5 70.4 "Should the state be involved in reducing the gap not involved 1.9 3.5 1.5 0.6 2.2 0.6 1.9 2.2 1.4 1.5 1.2 2.2 2.3 between the rich and the poor?" moderately involved 6.2 4.8 5.3 8.5 7.5 0.3 6.6 7.2 3.2 6.9 8.2 7.0 3.7 strongly involved 91.9 91.7 93.2 90.9 90.3 99.1 91.5 90.5 95.4 91.6 90.6 90.9 94.1 "In your opinion, what is the main reason why there unlucky 7.5 9.4 5.3 7.6 6.7 10.9 12.7 7.3 3.5 4.3 11.2 7.2 4.6 are some people in need in our country today?" laziness and lack of willpower 24.4 21.8 29.1 22.5 26.3 15.9 21.2 28.3 17.8 29.3 22.6 27.2 23.4 injustice in society 62.9 64.1 60.1 64.5 61.4 69.7 63.9 58.9 70.5 61.9 61.4 60.9 66.0 inevitable part of modern life 2.6 2.6 3.5 1.5 3.0 0.6 0.3 2.5 4.6 3.7 2.1 0.8 4.5 "Which of the factors in this list is the most effort and hard work 48.4 46.7 52.1 46.6 50.2 40.9 54.5 49.5 43.2 39.4 50.1 51.2 44.5 important to succeed in life in this country?" intelligence and skills 27.2 23.2 22.3 36.8 25.7 33.8 30.9 30.2 20.0 20.1 31.9 23.0 27.1 Inequality of Economic Opportunity in Turkey political connections 11.4 17.0 8.6 8.1 10.4 15.9 8.0 9.4 13.1 27.3 8.2 12.4 13.2 criminal/corrupt ties 10.8 11.8 12.9 7.7 11.4 8.4 5.4 9.4 19.5 11.3 8.9 10.6 12.7 Distribution of the population 35.2 32.8 32.0 81.4 18.7 23.5 45.9 22.2 8.4 30.5 33.1 36.4 Source: Tabulations from the Life in Transition Survey for Turkey, 2006. 19 Inequality of Economic Opportunity in Turkey 20 Table 2: The Household wealth index Principal components and summary statistics for asset indicators Inequality of Economic Opportunity in Turkey 21 Table 3: Partition of the population by circumstances; wealth analysis Table 4: Reduced-form regression of household asset index on circumstances Robust standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 Sample of 15-49 year-old ever-married women. Inequality of Economic Opportunity in Turkey 22 Table 5: The Opportunity Shares of Wealth Inequality for Women Note: partial shares are given by the means of smoothed and standardized estimates; sample of ever married women aged 15 to 49, second column restricts the sample to women aged 30 to 49. Table 6: The (Wealth) Opportunity-Deprivation Profile for Turkey Inequality of Economic Opportunity in Turkey Source: TDHS 2003. Sample includes Ever Married Women between Ages 30-49. The index has an overall mean of 0.283 and a standard deviation of 2.637. 23 Inequality of Economic Opportunity in Turkey 24 Table 7: The opportunity-Deprived and the Opportunity-Hoarders: characteristics of the bottom and top tenths of the opportunity profile Note: Wealth index in this analysis is recalculated from the assets in the DHS data. Sample includes only ever-married women ages 30-49. Table 8: Descriptive statistics for the asset indicators and demographic variables common to the HBS and DHS samples Notes: Statistics given for the full samples of each survey. Inequality of Economic Opportunity in Turkey 25 Table 9: Regression of household annual consumption on assets in HBS Log household annual expenditure 0.05 [0.03] 0.08*** 0.04** [0.02] [0.02] 0.21*** -0.05*** [0.01] [0.01] 0.09*** 0.39*** [0.01] [0.02] 0.23*** -0.03*** [0.01] [0.01] 0.29*** -0.05*** [0.03] [0.00] 0.22*** 0.02 [0.03] [0.01] 0.16*** 0.02*** [0.03] [0.00] 0.05*** -0.01*** [0.02] [0.00] 0.32*** 0.02*** [0.02] [0.00] 0.21*** 0.01** [0.01] [0.00] 0.12*** 0.10*** [0.02] [0.01] 0.16*** 18.23*** [0.03] [0.06] 0.20*** 25764 [0.01] 0.525 0.06*** [0.02] 0.05*** [0.01] 0.08*** [0.02] 0.11*** [0.01] 0.05*** [0.01] Ref. Robust standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 Table 10: Reduced-form regression of imputed per capita consumption on circumstances Robust standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 Source: TDHS 2003 with consumption imputated from HBS 2003; sample of ever-married women 30-49 Inequality of Economic Opportunity in Turkey 26 Table 11: The Opportunity Shares of Consumption Inequality for Women Note: Parametric standardized simulations. Source: TDHS 2003 with consumption imputed from HBS 2003; sample of ever-married women 30-49 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