The impact of internal migration on educational outcomes: Evidence from Turkey

Economics of Education Review 28 (2009) 739–749

Contents lists available at ScienceDirect

Economics of Education Review journal homepage: www.elsevier.com/locate/econedurev

The impact of internal migration on educational outcomes: Evidence from Turkey Ali Berker ∗ Department of Economics, Abant I˙ zzet Baysal University, 14280 Gölköy-Bolu, Turkey

a r t i c l e

i n f o

Article history: Received 4 September 2007 Accepted 16 March 2009 JEL classification: J24 J61 I21 Keywords: Human capital Educational attainment Internal migration

a b s t r a c t Similar to the relation between the inflows of immigrants and educational outcomes that are found in immigration studies, the spatial distribution of internal migrants within a given country also may influence educational outcomes, at least in the short run. This could be particularly true in Turkey, where inter-provincial mobility is high and where striking differences in educational resources and therefore educational success across regions persist. Using the 1990 and 2000 Turkish Censuses, this study exploits variations over time in the inflow of internal migrants across provinces to identify the causal effect of internal migration on natives’ educational outcomes. The evidence suggests that the inflow of migrants lowers natives’ completion rates for middle school and high school. Evidence also indicates that while the negative effects appear to be greater among native children from low-SES households, native high-SES households are able to mitigate these adverse effects for their children. Furthermore, the estimated effects exhibit some differences by children’s gender and migrant status. © 2009 Elsevier Ltd. All rights reserved.

1. Introduction As induced by the migration of families and children, the school-age population’s spatial redistribution across regions in a given country might alter both educational opportunities and incentives for both those who previously resided (native population) and those who newly reside (migrant population) in local areas experiencing different densities of migrant inflows. Economic theory conjectures that for a given local area, the migrant inflow may generate changes in both the marginal benefit and marginal cost of education, which might result in two opposite effects for educational investments. As a result, migration has ambiguous effects on natives’ educational outcomes in a given local area (Betts, 1998; Betts & Fairlie, 2003; Gould, Lavy, & Paserman, 2004). For instance, the migrant inflow may increase the marginal cost of education for natives,

∗ Tel.: +90 374 254 1464; fax: +90 374 253 4521. E-mail address: [email protected]. 0272-7757/$ – see front matter © 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.econedurev.2009.03.003

decreasing their educational outcomes because an expansion in the number of migrant students lessens the efficient use of fixed-level school resources, at least in the short run. The migrant inflow’s negative effect may be reversed, however, when migrants have relatively lower skills than natives; the skill premium – the marginal benefit of education – is likely to increase for natives, thus improving their educational outcomes. Economic theory, therefore, calls for a well-designed empirical study to determine the direction and magnitude of the relation between the migrant inflow and natives’ educational outcomes. In addition to responding this call, increasing public concerns about immigrant flows from less-developed to developed countries such as the U.S. and European countries have led researchers to explore international migration’s causal effects on various outcomes, such as natives’ educational outcomes. For example, Betts (1998) used variation in the immigration ratio across states over time in the U.S. to estimate immigration’s effect on natives’ educational attainments. Gould et al. (2004) examined the long-term educational impact of immigra-

740

A. Berker / Economics of Education Review 28 (2009) 739–749

tion for Israeli high school students who were exposed to different densities of immigrant students during their elementary education. These two studies provide evidence that as native students, particularly those with disadvantaged family backgrounds, are exposed to a greater immigrant influx, immigration’s adverse effect becomes stronger. Immigrant inflows may influence not only individuals’ educational level, but also the types of school they attend. Betts and Fairlie (2003) provide evidence that for a given metropolitan area, a higher immigrant inflow may lead more native students, particularly white native students, to attend private schools. Building on the econometric methods of previous studies that examined the spatial correlation between the immigrant population’s density and natives’ educational success in the international migration context, this study is the first to examine the causal relation between the migrant inflow and natives’ educational outcomes in the internal migration context. It is also the first to evaluate to what extent a shock to the educational sector (in this case a change to the school-age population within the local area caused by the migrant inflow) can be handled by the Turkish educational sector, which has been characterized by excessive centralized governance, the absence of specific policy rules for determining the distribution of resources across schools, and sharp differences in educational outcomes among regions as well as well across students with varying backgrounds. Using the 1990 and 2000 Turkish Censuses, I exploit changes over time in the inflow of internal migrants across provinces to estimate the effects of the migrant–native ratio on the likelihood of completing middle school for the 16–19 age group and of completing high school for the 18–20 age group. To measure the density of migrants for a given province, I obtain the recent migrant–native ratio by dividing the number of recent migrants ages 6–20 by that of natives in the same age group. A simple analysis of the association between the migrant inflow and natives’ educational outcomes might capture a spurious relation between migration and natives’ educational outcomes. This is mainly because the province-level fixed effects that may influence natives’ outcomes also might be related to the migrant inflow. Applying a twostage estimation method on 2-year, province-level panel data, I first estimate the first-difference specification to remove the province-level fixed effects. However, the first-difference estimation method still might yield biased estimates of the migrant–native ratio’s effect, because temporary shocks at the province level may be correlated with both the migrant–native ratio and natives’ educational outcomes. To address this possible problem, I use the migrant–native ratio in 1990 as an instrument for the change in the migrant–native ratio between 1990 and 2000 in the first-difference equation. Econometric approaches applied in this study provide reduced-form estimates that net out the relative strengths of the effects of changes in both labor market returns and the cost of acquiring additional education, where both changes were presumably induced by a change in the migrant inflow at the province level. Overall, the estimation results provide evidence for a negative association between the migrant inflow and natives’ educational outcomes. I fur-

ther investigate how the estimated effects differ with the native households’ socio-economic status (SES). While children from low-SES households appear to be most harmed by an increase in the migrant inflow, a statistically significant negative effect is limited to their middle-school completion rate, and no significant impact was observed for children from high-SES households. The estimation results also indicate heterogeneity in the estimated effects with respect to children’s gender and migrant status. 2. Theoretical framework Similar to international migration, by changing the school-age population’s distribution across regions within a given country, internal migration could affect educational production for native students in migrant-receiving local areas in various ways (Betts, 1998; Betts & Fairlie, 2003; Betts & Lofstrom, 1998; Gould et al., 2004). First, a higher inflow of migrant students may adversely alter the efficient use of school inputs. Indeed, a given local area experiencing a higher inflow of migrant students may experience a dramatic increase in the average number of students per educational input, such as teachers and laboratories, and thus this local area may fail to meet its student population’s educational needs. In addition to changing the size of school-age population, the migrant inflow may also change the composition of peer inputs for native students at the neighborhood, school, and class levels (Gould et al., 2004). The effects of the change in the peer composition on natives’ educational outcome may hinge on differences between native and migrant parents in their educational attainment and their preferences regarding their children’s education. Thus, by changing the efficient use of school inputs and native students’ peer composition, which are the most important determinants of the supply of education for a given local area, the migrant inflow may alter the marginal cost of education and, consequently, influence natives’ educational outcomes. For example, coupled with an increase in the school-age population, the inflow of migrant students with a lower parental background may adversely affect the education production in migrant-receiving areas, increasing the time period in which native students complete any given level of education.1 Consequently, this type of migrant inflow may increase the marginal cost of education, thus reducing natives’ educational outcomes. In addition to influencing the supply of education in a local area, the migrant inflow may also alter economic returns to education, leading to changes in natives’ demand for education (Betts, 1998). The direction and magnitude of this effect are closely related to the skill composition of both migrants and natives in a local area. For instance, when the migrant inflow results in a disproportional increase in the number of middle-school graduates, an increase in the wages of high-school graduates relative to those of middle-school graduates may produce incentives for addi-

1 In contrast, for example, migrant students with a higher parental education background may serve as a complementary input in educational production, increasing native students’ educational success.

A. Berker / Economics of Education Review 28 (2009) 739–749

741

Table 1 The distribution of migrant population who changed the province of residence across destination regions for each origin of region and regional socioeconomic development index scores. Destination regions Marmara

Aegean

Central Anatolia Mediterranean

Black Sea

Southeastern Anatolia

Eastern Anatolia

Census Census Census Census Census Census Census Census Census Census Census Census Census Census 1990 2000 1990 2000 1990 2000 1990 2000 1990 2000 1990 2000 1990 2000 Panel A: regional migration patterns Origin regions Marmara 40.66 36.47 Aegean 21.15 22.92 Central Anatolia 30.28 28.87 Mediterranean 19.60 21.61 Black Sea 60.05 54.43 Southeastern Anatolia 23.63 26.92 Eastern Anatolia 48.19 44.25

11.67 38.88 12.49 11.12 5.97 13.49 15.16

12.68 36.61 12.75 11.77 7.03 16.36 13.86

15.81 17.26 29.06 17.46 13.58 11.09 11.09

14.94 15.14 29.43 19.79 16.10 12.58 14.02

6.93 8.55 12.12 31.12 3.31 28.98 9.93

7.84 10.39 12.06 25.79 3.96 19.93 9.58

14.95 5.84 8.30 4.81 12.95 4.27 4.27

16.87 5.28 8.00 4.59 14.01 3.21 5.18

3.61 3.64 3.24 10.48 1.48 15.57 3.15

4.07 3.86 3.40 9.90 1.23 14.29 3.55

6.36 4.67 4.50 5.42 2.65 5.68 8.22

7.13 5.79 5.48 6.55 3.24 6.71 9.57

Regions Marmara

Aegean

Central Anatolia Mediterranean

Panel B: regional socio-economic development index scores Years in which index 1996 2003 1996 2003 1996 scores calculated Index scores 1.694 1.702 0.500 0.481 0.460

Black Sea

Southeastern Anatolia

Eastern Anatolia

2003

1996

2003

1996

2003

1996

2003

1996

2003

0.481

0.061

0.020

−0.543

−0.513

−1.036

−1.011

−1.137

−1.162

Source: Regional socio-economic development index scores are taken from the report prepared for Turkish Republic State Planning Organization (Dincer, Özaslan, & Kavasoglu, 2003).

tional educational investments, thereby increasing natives’ educational outcomes. 3. Background information on Turkey Compared to European and Organisation for Economic Co-operation and Development (OECD) countries, Turkey has devoted a smaller share of its gross domestic product (GDP) to the educational sector, and it has a stronger central authority that determines how educational resources are allocated (World Bank, 2005). In fact, Turkey’s Ministry of National Education (MONE) carries out of 94% of all educational decisions by itself. Despite its overwhelming centralized structure, there is no well-defined procedure to determine the distribution of resources across schools. In addition, because there is no efficient monitoring system that holds them accountable for their performance, local educational administrators do not have any incentive to improve their local areas’ educational outcomes. In such a system, where the central authority determines the level of resources and their use for each school, one would expect to observe an equal distribution of educational resources and outcomes across both schools and local areas. Research on education in Turkey, however, establishes that both educational resources and outcomes exhibit significant differences with respect to students’ characteristics, such as their gender and their families’ socio-economic status, as well as between regions (Mete, 2004; Tansel, 2002). Within such framework, assessing the educational consequences of population movements induced by migrant inflows for local areas may provide some information on the Turkish educational sector’s performance. But before performing this assessment, it is necessary to understand the magnitude and direction of internal migration in Turkey.

According to the latest Census in 2000, one in 10 changed their province within the last 5-year interval, and three in 10 reside in a province that is different from their province of birth.2 Shifting the focus to determine the direction of internal migration flows for the period analyzed in this study, Table 1 provides information for population movements between regions focusing on individuals who changed the province of residence (Panel A) and for regions’ socio-economic development index scores, which enable us to differentiate regions in terms of their social and economic development level (Panel B). In particular, by focusing on internal migration that occurred during the 1985–1990 period for the 1990 Census and that occurred during the 1995–2000 period for the 2000 Census, each row in Panel A in Table 1 shows the distribution of migrant population across destination regions for each region of origin. Together with information in Panel B in Table 1, the examination of interregional patterns suggest that individuals moved from less-developed regions, such as the Black Sea and Eastern and Southeastern Anatolia, to the most developed regions, such as Marmara and Aegean, confirming the fundamental prediction of migration theories that individuals move for better economics and social opportunities. For example, more than half of the migrant population from the Black Sea Region, located in Northern Turkey with the third-lowest scores on the socio-economic development index, chose the Marmara Region, located in Northwest Turkey with highest score on this development index, as a destination region. Likewise, the Marmara Region is also most attractive destination region for those

2 Note that because the Census data provide information on the change of residence only one time in a fixed time interval, such as 5 years prior to the Census, individuals’ mobility is likely to be underreported.

742

A. Berker / Economics of Education Review 28 (2009) 739–749

from the Eastern Anatolia Region, which has the lowest score on the socio-economic development index. Thus, it appears that population movements for 1985–2000 in Turkey have been mainly characterized by east-west and north-west migration flows, which also reflected the main direction of population movements during the 1960–1980 period (Gedik, 1996). 4. Data The data used in this paper come from a randomly drawn 5% sample from both 1990 and 2000 Censuses. To distinguish individuals’ migration behaviors at the provincial level, the critical information that the Census data provide is the province of residence for all individuals who are older than five at three different points in time: at birth, 5 years prior to the Census, and at the time of the Census. By restricting the analysis sample to the non-institutional population, I first utilize information about individuals’ mobility across provinces within the last 5 years prior to the Census. Individuals who were residing in the same province 5 years prior to the Census and at the time of the Census are defined as natives. If individuals were residing in different provinces 5 years prior to the Census and at the time of the Census, they are defined as recent migrants. Furthermore, I split the natives into groups by matching information regarding their province of residence 5 years prior to the Census with information on their birthplaces. I define individuals whose information on provinces matched as permanent natives, and all remaining natives are grouped as old migrants. To determine whether a child belongs to any of these four groups, I use information on the household head’s migrant status. To construct the province-level, 2-year panel data, I recode provinces and districts in the 2000 Census based on the structure of the administrative divisions in the 1990 Census year, resulting in a total of 73 provinces for both Censuses. Furthermore, the analysis sample is restricted to individuals who were residing in the province or district centers, excluding rural places of residence. This restriction enables me to control differences between urban and rural residences in the demand and supply of education and migration behavior, isolating the contribution of these urban–rural differences in the estimated effects of migrant inflows. For example, while all levels of pre-tertiary education are provided in almost all urban areas, the provision of educational services in rural areas is limited to primary education—there are even some rural settlements with no primary school educational services. In addition, the quality of educational services differs significantly between urban and rural residences, in that compared to their peers in urban schools, students in rural schools are more likely to receive their education in integrated classes, where students from different grades receive their education at the same time and in the same class; are more likely to have poor school resources; and are more likely to suffer a high rate of teacher absenteeism (World Bank, 2005). Furthermore, compared to individuals in urban areas, those in rural areas may have a lower demand for education because of a higher cost of receiving education beyond primary school (for example, because of commuting costs), and the

sharp wage differential between the urban and rural labor markets in disfavor of those in the latter market. Finally, migrant flows to rural areas are likely to last for a short time and are observed with higher frequencies, making it difficult to detect the possible educational consequences of migrant inflows on migrant-receiving rural areas. This problem might become more severe when considering the fact that the Census data are not appropriate to measure and thus analyze relatively more volatile, transitory, and short-term population movements that may reflect the characteristics of migration to rural areas. In the empirical analysis, because the Census’s information on individuals’ education success is limited to their final degrees received, I focus on two educational outcomes at the provincial level: middle-school and high-school completion rates. To uncover the exact nature of the relation between migrant inflows and these educational outcomes, a researcher should have individuals’ information on where they resided during the time when they received their education. Given that the Census data provide information on individuals’ one-time (im)mobility within the last 5 years, to increase the likelihood of the matching information on where native children resided with where they received their education, I focus on 16–19 year olds who were 11–14 years old 5 years prior to the Census for the analysis of middle-school completion rate. Similarly, for the analysis of high-school completion rate, I focus on 18–20 year olds who were 13–15 years old 5 years prior to the Census.3 When the independent variable of interest is constructed, because the change in the schoolage population (6–20) that was induced by the inflow of recent migrants may matter most for changes in educational resources at the provincial level, the inflow rate of recent migrants is measured as the ratio of the number of recent migrants ages 6–20 to that of natives ages 6–20.4 5. Empirical strategy To estimate the causal effect of the recent migrant inflow on natives’ educational outcomes, I implement a two-stage estimation method that has been applied when examining the effects of immigration on labor market and educational outcomes (Altonji & Card, 1989; Betts & Fairlie, 2003). Given that household characteristics may be important determinants of both children’s educational outcomes and their migrant–native status, in addition to the vector of province dummies, the first-stage estimation uses individual-level observations to control for differences in the native population’s observable characteristics across provinces.5 For each

3 In the Turkish educational system, primary education (grades 1–5) is compulsory, and the primary school-age population normally consists of children ages 6–11. Since the compulsory educational period was extended from 5 to 8 years in 1997, children ages 11–14 also have been required to attend and complete middle school (grades 6–8). Finally, high-school education (grades 9–11) is given to children ages 15 and older, and its educational period was extended from 3 to 4 years in 2006. 4 In the article, when referring the recent internal migrant–native ratio, I use two terms – the recent migrant ratio and the migrant ratio – interchangeably. 5 For the first-stage estimation, the analysis sample includes all children of interest in a given household. This is mainly because examining

A. Berker / Economics of Education Review 28 (2009) 739–749

year of Census data, the estimated coefficients on province dummies capture the regression-adjusted, province-level averages of educational outcomes and serve as the dependent variable in the second-stage estimation, in which the following regression model was constructed by taking the first-difference of the specified variables defined below between 1990 and 2000: ˆ j2000 − ˆ j1990 = (RMj2000 − RMj1990 ) + (Sj2000 − Sj1990 ) + (nj − nj ) + (vj2000 − vj1990 ) where for the denoted year of the Census, t,  jt measures the adjusted provincial average of educational outcomes in province j at time t; RMjt denotes the province-level recent migrant ratio; and Sjt represents the vector of the provincelevel characteristics that may be related to province-level educational outcomes.6 Because j denotes the provincelevel characteristics that are assumed to be unobserved and time-invariant for any given province, the first-differencing procedure eliminates them. Finally, jt denotes a provincespecific error term.7 The coefficient of interest is , which measures the extent to which natives’ educational outcomes between 1990 and 2000 differed in provinces with a greater migrant inflow increase relative to provinces that did not experience such a change. Using the equation stated above, the migrant ratio’s effects on natives’ educational outcomes are obtained by implementing the first-difference-instrumental variable method (the first-difference-IV), in which the change in the migrant ratio between 1990 and 2000 is instrumented with the recent migrant ratio in 1990. The first-difference component of this method controls for time-invariant, province-specific effects and time-specific effects that uniformly influenced all provinces during the 1990–2000 period.8 The IV component of this method accounts

only children with family background information may bias the estimated effects because information on children’s family background is available only if they reside with their parents. Nevertheless, the estimation results reported in the paper are robust to restricting the analysis sample to those children who were residing with their parents; the estimation results for this analysis are available from the author upon request. When the firststage equation is estimated by a linear probability model with a robust option, it controls for the child’s age and gender, a set of dummy variables indicating the household head’s educational attainment, and the number of individuals aged 0–5, 6–20, 21–64, and 65 or older in the household. In the appendix, Tables A.1 and A.2 provide the descriptive statistics of these control variables used in the first-stage estimation for educational outcomes. 6 Specifically, these province-level variables include the logarithm of the number of primary or middle-school graduates, high-school graduates, and university graduates ages 25–65, the logarithm of the number of employed individuals ages 25–64, the provincial average of household size, and the logarithm of the province population. The descriptive statistics of these control variables used in the second-stage estimation are given in Table A.3 in the appendix. 7

−1 −1 The first-difference speciations use (Nj1990 + Nj2000 )

−1/2

743

for temporary, province-specific shocks that occurred between 1990 and 2000. One example of such shortlived province-specific change is the 1999 earthquake that caused widespread destruction in the Northwestern part of Turkey, which had achieved greater economic and social development than other regions of the country. When comparing scores from the socio-economic development index, which was measured in provinces that had been hit by the earthquake in 1996 – about 3 years before the earthquake – and in 2003—about 4 years after the earthquake, these provinces improved or at least protected their relative positions. This evidence points out the persistence of the province-level fixed effects even after such a shock. At the same time, some permanent changes may have taken place within provinces during the 1990–2000 period, when important market-oriented reforms were introduced and Turkey was integrated into the world economy. However, to the extent to which permanent changes occurred within the provinces does not alter the rank ordering of province in terms of their economic and social development whose relative measures are also key determinants of migrant flows, the identification strategy of the IV method may remain valid. To assess the extent to which the ranking of provinces changed between 1990 and 2000, I calculate correlations between the province-level economic and social indicators measured in 1990 and those measured in 2000, when they are available. The highly positive correlations between these indicators, ranging from 0.902 to 0.987, suggest that the economic and social development rankings across provinces remained unchanged, providing support for the IV’s identification strategy.9 Nevertheless, there are some other threats to the validity of the instrument used in the first-difference-IV specifications, which are difficult to address due to fact that the Census data is the only data set with information on the province-level migration and educational outcomes and that it contains limited information on these outcomes. When the change in the recent migrant ratio is instrumented by the recent migrant ratio in 1990, the maintained assumption is that recent migrants are more likely to migrate to provinces where previous migrants are densely populated, and thus the recent migrant ratio in 1990 can only influence changes in educational outcomes by exclusively affecting the change in the recent migrant ratio. However, for example, if this positive correlation stems from the fact that recent and previous migrants may share some common unobservable characteristics that are also related to their migration and educational outcomes, the IV estimation might yield biased estimates. Because research on migrant networks provide causal evidence that the stock of previous migrants in terms of both their size and quality perform important functions in migrants’ destination choices and their economic and social success as well, this

as a weight,

where Nj1990 and Nj2000 are the number of observations for the groups of interest in province j in the 1990 amd 2000 Censuses. The square root of the number of observations for the groups of interest in the province is used as a weight to estimate the cross-sectional specfications. 8 Thus, in addition to, for example, provincial-level preferences or tastes for education that are common to all of Turkey’s residents and do not change over time in a given province, the first differencing also controls for shifts in natives’ educational outcomes during the time period analyzed

here because of policy changes at the national level, such as extending compulsory education from 5 to 8 years in 1997. 9 In particular, I calculate correlations (shown in parentheses) for the following province-level variables between the specified years: GDP per capita, 1990–2001 (0.902); socio-economic development index, 1996–2003 (0.986); share of value added in the manufacturing industry, 1990–2002 (0.987); and bank deposits per capita (1990–2000) (0.979).

744

A. Berker / Economics of Education Review 28 (2009) 739–749

Table 2 Educational attainments of children by a household head’s migrant status: 1990 and 2000 Census. Middle school

Natives Permanent natives Old migrants Recent migrants

High school

1990

2000

Change (%)

1990

2000

Change (%)

0.442 (0.496) 0.432 (0.495) 0.458 (0.498) 0.430 (0.491)

0.585 (0.492) 0.595 (0.500) 0.571 (0.494) 0.612 (0.487)

32.3 37.7 24.6 42.3

0.275 (0.446) 0.262 (0.439) 0.295 (0.456) 0.317 (0.465)

0.454 (0.497) 0.462 (0.498) 0.442 (0.496) 0.594 (0.491)

65.1 76.3 49.8 87.3

In each native and migrant group, the sample of analysis is children ages 16–19 for middle-school completion rate, and those ages 18–20 for high-school completion rate. See text for detailed information on the definition of native and migrant groups. Standard deviations are reported in parentheses.

assumption seems to maintain through the analysis (Bartel, 1989; Munshi, 2003). The other issue that may pose a threat to the validity of the IV approach is the extent to which the recent migrant ratio in 1990 is correlated with the change in the migrant ratio. If this correlation is weak, there might be the problem of weak identification, leading to have biased IV estimates in the direction of OLS estimates. When the Staiger-Stock (1997) rule of thumb suggests that for a single endogenous regressor the first-stage F-statistics must be larger than 10 to avoid the problem of weak instruments are applied, all IV specifications with one endogenous regressor and one instrument satisfy this requirement, suggesting that my results are not driven by the weak identification problem. In the empirical analysis, I finally investigate the possibility that individuals may move out the province of residence as a means to mitigate the adverse effect of a shock induced by the inflow of migrants, thereby underestimating the impact of the migrant inflows on natives’ educational outcomes. As shown below, the main estimation results are robust to the presence of such a channel, which may lessen the adverse impacts of migrant inflows on natives’ educational outcomes.

6. Results 6.1. Main results The Census data provide limited information on individuals’ education outcomes, including only information regarding the last educational level that individuals had attained at the time of the Census. Using this limited information, I estimated the effects of the migrant ratio on the middle-school completion rates for 16–19 year olds children and the high-school completion rate for 18–20 year olds. As reported in Table 2, regardless of their migrant status, children’s educational attainment improved significantly between 1990 and 2000. In particular, the largest increases are observed for recent migrants whose highschool completion rate significantly exceeds that of their peers in 1990. To assess the presence and direction of bias in the estimated effect of the recent migrant ratio, the empirical analysis begins by providing of the first-difference-IV estimates, along with the cross-sectional and the firstdifference estimates. As presented in Table 3, for each specification, the estimation result suggests a negative association between the migrant ratio and both the middle-school and high-school completion rates, while the

magnitude and significance of the estimated effect differs across specifications.10 The first-difference-IV estimates are larger in absolute value. One possible explanation for this finding is that migrants may choose to not move to provinces where educational resources, for example, have been wiped out by natural disasters, such as the 1999 Marmara Earthquake in Turkey and, consequently, where their native population’s educational success might have been declining for the specified time period, thus reducing the negative estimated effects of both the first-difference and cross-sectional specifications. That is, the estimated effects of the latter specifications might be underestimated because of a positive selection mechanism through which families prefer to migrate to local areas where they perceive the native population’s higher educational success as an indicator for better local educational resources. Another possible explanation is that the migrant ratio may be measured with error for a given province. More importantly, the first-differencing transformation of the migrant ratio may exacerbate this measurement error problem, biasing the recent migrant ratio’s negative estimated effects toward zero. Therefore, the relatively larger IV estimates suggest that the IV specification may account for problems associated with the positive selection of migrants across provinces and with measurement error in the migrant ratio. Furthermore, the first-difference-IV estimates suggest that among native children, the migrant ratio’s impact is higher for middle-school than high-school completion rates. It may be that the sequential nature of the educational process may selectively eliminate students with lower ability and aspirations, and those with a disadvantaged background, from the educational ladder, thus reducing the

10 As suggested by an anonymous reviewer, using estimated parameters of the first-stage equation (or their first-differencing transformation) as a dependent variable may bias statistical inferences when their standard errors in the second-stage equation are not properly adjusted. For this reason, each cross-sectional speciation is reestimated by weighting with the inverse of the standard error of the estimated dependent variable in the second stage. In addition, for first-difference and first-difference-IV specifications, I use a modified version of the method used by Betts and Fairlie (2003), in which I pool 2 years of Census data and use a linear probability model to obtain the estimated coefficients on the interaction terms between a dummy variable for year 2000 and province dummies and, finally, use these estimated coefficients as a dependent variable in the second-stage equation weighted with the inverse of their standard error. These empirical analyses provide evidence that the estimated effects and their statistical inferences reported in the paper are not driven by statistical properties of the estimated dependent variable in the secondstage equation. The estimation results of these analyses are available from the author upon request.

0.747 –

The coefficients presented in the table come from the second-stage estimation. In the first-stage estimation, the linear probability for each educational outcome is estimated to obtain provincial averages of each outcome, which are used as a dependent variable in the second-stage estimation. The first-stage estimation includes dummies for each province and variables that are listed in Tables A.1 and A.2. The control variables used in the second-stage regression are listed in Table A.3. All specifications are weighted. See text for details. Standard errors are reported in parentheses. * Mark that the estimated coefficients are significantly different from zero at the 10% level. *** Mark that the estimated coefficients are significantly different from zero at the 1% level.

−0.405* (0.248) 73 0.169 58.17 −0.061 (0.176) 73 0.215 – −0.171 (0.191) 73 0.761 – (0.103) 73

−0.368

−0.885 (0.309) 73 0.099 57.05 −0.152 (0.205) 73 0.105 – −0.046 (0.212) 73 0.746 – Recent migrant–native ratio −0.664 (0.159) Number of observations 73 R-squared 0.695 First-stage F-statistics –

First-difference Cross-section OLS 2000 Cross-section OLS 1990

*** ***

High school

First-difference-IV First-difference Cross-section OLS 2000 Cross-section OLS 1990

Middle school

Table 3 Estimates of effects of the recent migrant–native ratio on natives’ educational outcomes.

***

First-difference-IV

A. Berker / Economics of Education Review 28 (2009) 739–749

745

adverse effects of migrant inflows at the higher educational level. In Turkey, this selection effect might be compounded by the fact that the educational system’s financing is poorly designed to address the needs of students with varying backgrounds and that the Secondary School (High School) Student Selection and Placement Examination conducted at the end of eighth grade (when compulsory education ends) may enhance the educational inequalities associated with students’ and their families’ characteristics. As the international migration literature suggests, when their children’s education is at stake, native families may exhibit heterogeneity in responding to the migrant inflow (Betts, 1998; Betts & Fairlie, 2003; Gould et al., 2004). In particular, the level of family wealth may determine to what extent the family mitigates these adverse effects of migrant inflows on their children’s educational outcomes. Not surprisingly, the adverse effects of migrant inflows are expected to diminish as families devote more resources to their children’s education. Because the Census data do not provide any information about families’ wealth, I use the household head’s education to assign families into two different socio-economic groups, assuming that this variable is positively correlated to the family wealth: (1) the household heads with less than a high school education (low-SES households), and (2) those with at least a high-school education (high-SES households). As shown in Table 4, for the middle-school completion rate, the most adverse effect of migrant inflows is felt by native children from low-SES households, suggesting that high-SES households are more likely than their counterparts to have the means to reduce the adverse effects of migrant inflows. For the high-school completion rate, the estimated effect is negative but not significant, and it does not show much variation by children’s socio-economic status. 6.2. Other results: heterogeneity in estimated effects In this section, I determine the extent to which the recent migrant ratio’s estimated effects vary for different migrant and native groups and their gender. Given the definitional framework, where natives include both permanent natives and old migrants, the compositional effects may account for their estimation results, specifically for the high-school completion rate, reported in Table 3. Indeed, the findings shown in Table 5 suggest that for the high-school completion rate, because the negative estimated effects of the migrant ratio are far larger and more significant for permanent natives than for old migrants, grouping them together may reduce the estimated effects for natives in terms of both the magnitude and significance level. For the middle-school completion rate, the estimated effects become larger for both permanent natives and old migrants. As reported in Table 5, the finding that recent migrants are the least likely to be adversely affected by the migrant inflow deserves further analysis. In particular, when I differentiate the estimated effects of the migrant ratio by recent migrants’ gender, the estimation results reported Table 6 suggest that that the gender compositional effect may account for the lowest overall negative estimated effects obtained for recent migrants. For example, for the

746

A. Berker / Economics of Education Review 28 (2009) 739–749

Table 4 Estimates of effects of the recent migrant–native ratio on natives’ educational outcomes by families’ socio-economic status. Middle school

Low-SES families High-SES families

High school

First-difference

First-difference-IV

First-difference

First-difference-IV

−0.157 (0.217) −0.176 (0.251)

−0.940 (0.327) −0.549 (0.362)

0.034 (0.191) −0.296 (0.256)

−0.412 (0.269) −0.370 (0.362)

***

See notes to Table 3. Standard errors are reported in parenthesis. *** Mark that the estimated coefficients are significantly different from zero at the 1% level. Table 5 Estimates of effects of the recent migrant–native ratio on educational outcomes by migrant status. Middle school

Permanent natives Old migrants Recent migrants

High school

First-difference

First-difference-IV

First-difference

First-difference-IV

−0.323 (0.227) −0.585** (0.285) 0.194 (0.290)

−1.191*** (0.352) −1.161*** (0.390) −0.057 (0.416)

−0.221 (0.222) −0.397 (0.242) 0.315 (0.337)

−0.878*** (0.332) −0.242 (0.320) −0.285 (0.493)

See notes to Table 3. Standard errors are reported in parenthesis. ** Mark that the estimated coefficients are significantly different from zero at the 5% level. *** Mark that the estimated coefficients are significantly different from zero at the 1% level. Table 6 Estimates of effects of the recent migrant–native ratio on educational outcomes by migrant status and gender. Middle school

High school

First-difference

First-difference-IV

First-difference

First-difference-IV

(A) Natives Male Female

−0.307 (0.238) 0.046 (0.226)

−1.273*** (0.363) −0.480 (0.325)

−0.131 (0.209) −0.012 (0.207)

−0.312 (0.289) −0.523* (0.298)

(B) Permanent natives Male Female

−0.443* (0.268) −0.168 (0.263)

−1.470*** (0.419) −0.877** (0.392)

−0.309 (0.217) −0.151 (0.270)

−0.717** (0.351) −1.015*** (0.409)

(C) Old migrants Male Female

−0.698** (0.344) −0.284 (0.232)

−1.690*** (0.487) −0.639 (0.405)

−0.334 (0.331) −.436* (0.261)

−0.114 (0.438) −0.383 (0.346)

−0.757 (0.556) 0.526 (0.532)

−0.183 (0.505) −0.376 (0.408)

−1.474** (0.751) −0.699 (0.585)

(D) Recent migrants Male Female

0.245 (0.368) 0.200 (0.370)

See notes to Table 3. The sample of analysis is restricted to males, females, respectively. Standard errors are reported in parentheses. * Mark that the estimated coefficients are significantly different from zero at the 10% level. ** Mark that the estimated coefficients are significantly different from zero at the 5% level. *** Mark that the estimated coefficients are significantly different from zero at the 1% level.

recent migrants’ high-school completion rate, the negative estimated effect for males (−1.474) is about two times higher than for their female peers (−0.699). A similar gender compositional effect is found for their middle-school completion rates, where the estimated effects run in the opposite direction, but not statistically significant. While the most adverse effect of migrant inflows on the highschool completion rate is felt by male recent migrants, taking all the estimation results together, there exists no strong evidence for the prediction that because recent migrants are likely to be clustered with old migrants and other recent migrants within a province, the estimated effects will be highest for recent and old migrants and to be lowest for permanent natives. As shown in Table 6, for both permanent natives and old migrants, examining gender differences in the estimated effects of the migrant ratio suggests that the migrant inflow’s adverse effects seem to be larger and more significant for males’ middle-school completion rate than for

that of their female peers. However, the opposite pattern is observed for the high-school completion rate, for which the estimated effects are significant only for permanent natives. The latter finding is consistent with findings from previous studies, which suggest that when a negative economic shock occurs at the household or community level, female children experience its burden disproportionately, in part, because of lower returns to females’ education and a cultural preference favoring male children.11 However, it is difficult to explain or reconcile the former finding with the existing literature, where internal migration’s educational consequences are not yet extensively analyzed along with the gender dimension. Finally, I evaluate whether the estimated effects are robust to the inclusion of the old migrant–native ratio in the regression models. For a given province, the stock of

11

See, for example, Rose (1999).

A. Berker / Economics of Education Review 28 (2009) 739–749

747

Table 7 Estimates of effects of the recent and old migrant–native ratios on educational outcomes by migrant status. Middle school

High school

First-difference

First-difference-IV

First-difference

First-difference-IV

(A) Natives Recent migrant–native ratio Old migrant–native ratio

−0.159 (0.206) −0.038 (0.063)

−0.921*** (0.311) −0.051 (0.65)

−0.071 (0.176) −0.052 (0.054)

−0.449* (0.248) −0.059 (0.053)

(B) Permanent natives Recent migrant–native ratio Old migrant–native ratio

−0.342 (0.228) −0.071 (0.077)

−1.222*** (0.357) −0.052 (0.076)

−0.224 (0.224) −0.017 (0.072)

−0.896*** (0.333) −0.028 (0.072)

(C) Old migrants Recent migrant–native ratio Old migrant–native ratio

−0.588** (0.290) −0.006 (0.096)

−1.184*** (0.404) −0.021 (0.091)

−0.254 (0.222) −0.057 (0.046)

−0.206 (0.331) 0.033 (0.076)

(D) Recent migrants Recent migrant–native ratio Old migrant–native ratio

0.164 (0.290) −0.127 (0.104)

−0.147 (0.413) −0.069 (0.090)

0.331 (0.339) 0.077 (0.110)

−0.201 (0.483) 0.065 (0.106)

See notes to Table 3. Standard errors are reported in parentheses. * Mark that the estimated coefficients are significantly different from zero at the 10% level. ** Mark that the estimated coefficients are significantly different from zero at the 5% level. *** Mark that the estimated coefficients are significantly different from zero at the 1% level.

previous migrants may simultaneously determine families’ in-migration propensities and their labor market and their children’s educational outcomes by providing assistance in housing, finding a job, and providing education-related resources. Thus, omitting the old migrant–native ratio as a measure of the stock of previous migrants may contaminate the estimation results. Furthermore, as discussed in international migration literature, the adverse effects of migrant inflows on native outcomes, particularly their labor market outcomes, may be lessened by the outflow of the native population (Borjas, 1994; LaLonde & Topel, 1997). Although research on international migration provides mixed evidence for the presence of this possible mechanism through which the impacts of migrant inflows are reduced by natives’ movements within the country (Borjas, 2006; Card & DiNardo, 2000), because internal migration involves more intensely two-way population movements between provinces, the impact of natives’ out—migration might be greater when the effects of internal migrant inflows are examined. Because to some extent, by definition, the old migrant population captures the size of the immobile population in the province for the last 5 years prior to the Census, including the old migrant–native ratio in specifications also controls for the local population’s out-migration propensities induced by recent migrant inflows.12 As reported in Table 7, for each migrant and native group, the estimation results suggest that the recent migrant ratio’s estimated effects are robust to the inclusion of the old migrant ratio. That is, while except for recent migrants, the inflow of recent migrants continues to have a significant, negative impact on the middle-school completion rate, it only significantly reduces the high-school completion rate for permanent natives. Furthermore, findings

12 While the old migrant ratio might also be endogenously determined, the purpose of this empirical exercise is restricted to determining whether the recent migrant ratio’s estimated effects reported in the paper are driven by the exclusion of the stock of old migrants measured by the old migrant-native ratio from specifications.

in Table 7 indicate that the old migrant ratio does not generate any significant impact on children’s educational outcomes, even for recent migrants and the old migrants themselves. 7. Conclusion This study provides evidence that the migrant influx is negatively associated with natives’ educational success, particularly that of permanent natives. This result might have an important implication for Turkey, where interprovincial mobility is high and where striking regional disparities in educational resources and thus educational success persist. Thus the Turkish educational system, in which the central government uniformly determines each school’s financial and educational resources, might fall short in taking into account province- or school-level peculiarities, such as the share of school-age children from poor families or the share of migrant children in a school, when determining the amount of resources assigned to each public school. Indeed, the evidence suggests that poor families are more likely to fail to mitigate the adverse effects of having a higher share of school-age migrants in the province where they live. However, to reach definitive conclusions, this negative relation between the migrant inflow and natives’ educational outcomes must be empirically verified using more disaggregated data, such as neighborhood- and school-level data, which are not yet available for Turkey. While these analyses will augment our knowledge about the various channels through which the migrant inflow may influence educational outcomes, they also will help policy-makers design public policies to offset any adverse effects of the migrant inflow on local areas’ educational production. Acknowledgments I am thankful to two anonymous referees for their very helpful comments that significantly improved the paper;

748

A. Berker / Economics of Education Review 28 (2009) 739–749

Table A.1 Descriptive of statistics of variables used in the first stage regression analysis of likelihood of completing middle school: age group 16–19. Variables

Age Female Household head’s education: no formal schooling Household head’s education: primary school graduate Household head’s education: middle-school graduate Household head’s education: high-school graduate Household head’s education: 2-year or 4-year university graduate Household head’s education: missing Number of household members aged 0–5 Number of household member aged 6–20 Number of household member aged 21–64 Number of household member aged 65 or more Number of observations

Natives

Permanent natives

Old migrants

1990

1990

2000

Recent migrants

1990

2000

2000

1990

2000

17.412 (1.100) 0.495 (0.499) 0.210 (0.407)

17.473 (1.106) 17.399 (1.098) 17.467 (1.103) 17.431 (1.103) 0.496 (0.499) 0.504 (0.499) 0.501 (0.499) 0.480 (0.499) 0.156 (0.363) 0.226 (0.418) 0.164 (0.370) 0.188 (0.390)

17.488 (1.110) 17.520 (1.105) 17.723 (1.097) 0.489 (0.499) 0.461 (0.498) 0.459 (498) 0.145 (0.352) 0.148 (0.355) 0.100 (0.300)

0.594 (0.490)

0.534 (0.498)

0.592 (0.491)

0.519 (0.499)

0.598 (0.490)

0.554 (0.497) 0.484 (0.499)

0.355 (0.478)

0.076 (0.266)

0.112 (0.316)

0.079 (0.270)

0.122 (0.328)

0.073 (0.260)

0.099 (0.299) 0.097 (0.297)

0.090 (0.286)

0.069 (0.254)

0.127 (0.332)

0.066 (0.249)

0.132 (0.339)

0.074 (0.262)

0.119 (0.324)

0.177 (0.382)

0.341 (0.474)

0.047 (0.211)

0.069 (0.253)

0.034 (0.182)

0.059 (0.237)

0.065 (0.247)

0.081 (0.273)

0.090 (0.286)

0.113 (0.316)

0.0005 (0.0238) 0.0002 (0.016) 0.0005 (0.024) 0.0004 (0.020) 0.0005 (0.023) 0.0001 (0.011) 0.0004 (0.021) 0.0001 (0.008) 0.478 (0.870)

0.416 (0.870)

0.516 (0.913)

0.461 (0.949)

0.421 (0.800)

0.358 (0.752) 0.513 (0.864)

0.348 (0.746)

3.105 (1.715)

2.841 (1.842)

3.207 (1.792)

2.969 (1.957)

2.952 (1.582)

2.676 (1.630)

2.911 (1.664)

2.816 (1.885)

2.522 (1.298)

2.546 (1.442)

2.548 (1.393)

2.603 (1.957)

2.482 (1.139)

2.473 (1.223)

2.193 (1.250)

2.037 (1.556)

0.126 (0.382)

0.133 (0.408)

0.143 (0.410)

2.603 (1.589)

0.099 (0.333)

0.111 (0.359)

0.074 (0.287)

0.057 (0.255)

126,861

171,114

74,080

95,322

52,781

75,792

12,495

14,059

The listed variables are used to estimate provincial average of second level primary school completion rate in the first-stage estimation. Standard deviations are reported in parentheses.

Table A.2 Descriptive of statistics of variables used in the first stage regression analysis of likelihood of completing high school: age group 18–20. Variables

Natives 1990

Age Female Household head’s education: no formal schooling Household head’s education: primary school graduate Household head’s education: middle-school graduate Household head’s education: high school graduate Household head’s education: 2-year or 4-year university graduate Household head’s education: missing Number of household members aged 0–5 Number of household member aged 6–20 Number of household member aged 21–64 Number of household member aged 65 or more Number of observations

2000

Permanent natives

Old migrants

1990

1990

2000

Recent migrants 2000

1990

2000

18.920 (0.827) 18.939 (0.818) 18.927 (0.834) 0.548 (0.497) 0.539 (0.498) 0.559 (0.496) 0.217 (0.412) 0.164 (0.370) 0.234 (0.423)

18.944 (0.824) 18.908 (0.816) 18.932 (0.810) 18.961 (0.824) 19.05 (0.810) 0.544 (0.498) 0.532 (0.498) 0.534 (0.498) 0.534 (0.4989) 0.488 (0.499) 0.174 (0.379) 0.191 (0.393) 0.152 (0.359) 0.129 (0.335) 0.079 (0.269)

0.580 (0.493) 0.526 (0.499)

0.577 (0.493)

0.511 (0.499)

0.583 (0.492) 0.545 (0.497) 0.454 (0.497)

0.281 (0.449)

0.078 (0.268) 0.111 (0.315)

0.079 (0.270)

0.121 (0.326)

0.076 (0.266) 0.099 (0.299) 0.093 (0.291)

0.075 (0.264)

0.080 (0.271)

0.0775 (0.2679) 0.136 (0.343)

0.084 (0.278) 0.124 (0.329)

0.231 (0.421)

0.456 (0.498)

0.030 (0.171)

0.063 (0.243) 0.078 (0.268) 0.090 (0.286)

0.107 (0.309)

0.131 (0.337)

0.043 (0.203) 0.065 (0.247)

0.055 (0.228)

0.0005 (0.021) 0.0004 (0.020) 0.0005 (0.022) 0.0005 (0.023) 0.0003 (0.019) 0.0002 (0.015) 0.0004 (0.020) 0.0001 (0.012) 0.535 (0.913)

0.469 (0.935)

0.579 (0.959)

0.525 (1.025)

0.469 (0.835) 0.397 (0.796) 0.550 (0.867)

0.323 (0.716)

2.901 (1.763)

2.723 (1.877)

2.999 (1.845)

2.860 (1.995)

2.753 (1.619)

2.544 (1.695) 2.626 (1.655)

2.768 (1.990)

2.557 (1.400)

2.612 (1.587)

2.590 (1.503)

2.695 (1.753)

2.507 (1.227)

2.505 (1.333)

2.059 (1.3119) 1.844 (1.645)

0.132 (0.391)

0.139 (0.420)

0.150 (0.420)

0.157 (0.456)

0.103 (0.339)

0.115 (0.366)

0.066 (0.273)

0.042 (0.221)

87,765

125,495

51,650

69,934

36,115

55,561

9544

13,037

The listed variables are used to estimate provincial average of middle-school completion rate in the first-stage estimation. Standard deviations are reported in parentheses.

A. Berker / Economics of Education Review 28 (2009) 739–749

749

Table A.3 Descriptive of statistics of variables used in the second-stage regression analysis of educational outcomes: the 1990 and 2000 Census. Variables

1990

2000

Change

Recent migrant–native ratio (ages 6–20) Log employment rate (ages 25–64) Log primary and secondary graduates (ages 25–64) Log high school graduates (ages 25-64) Log university graduates (ages 25–64) Mean household size Log province population

0.110 (0.046) 9.219 (1.399) 9.180 (1.410) 7.596 (1.459) 6.989 (1.647) 5.716 (1.089) 11.111 (0.989)

0.082 (0.032) 9.383 (1.497) 9.529 (1.413) 8.249 (1.414) 7.681 (1.602) 5.388 (1.400) 11.321 (1.082)

−0.028 0.164 0.348 0.653 0.691 −0.327 0.209

Each variable is weighted by the number of children aged 6–20 years in the province. Standard deviations are reported in parentheses.

to this article’s responsible editor, Dan Goldhaber, for his kind and generous patience during the submission of the paper; to I˙ nsan Tunalı and I˙ smail Erol for their insightful comments during the early stage of the research; to Nebile Korucu and Bengi Yanık for their research assistance; and finally to Donna Maurer for her excellent editorial help. I gratefully acknowledge funding from The Scientific and Technological Research Council of Turkey (TÜBI˙ TAK), the Migration Research Program at Koc¸ University (MI˙ REKOC¸), and Abant I˙ zzet Baysal University. Appendix A. See Tables A.1–A.3. References Altonji, J. & Card, D. (1989). The effects of migration on the labor outcomes of natives. NBER Working Paper No. 3123. Bartel, A. P. (1989). Where do the new U.S. immigrants live? Journal of Labor Economics, 7(2), 371–391. Betts, J. R. (1998). Educational crowding out: Do immigrants affect the educational attainment of American minorities? Discussion Paper No. 98-04, University of California. San Diego: Department of Economics. Betts, J. R., & Fairlie, W. R. (2003). Does immigration induce ‘native flight’ from public schools into private schools? Journal of Public Economics, 87(5/6), 102–987. Betts, J. R. & Lofstrom, M. (1998). The educational attainment of immigrants: Trends and implications. NBER Working Paper No. 6757.

Borjas, G. J. (1994). The economics of immigration. Journal of Economic Literature, 32(4), 1667–1717. Borjas, G. J. (2006). Native internal migration and the labor market impact of immigration. Journal of Human Resources, 41(2), 221–258. Card, D., & DiNardo, J. (2000). Do immigrant inflows lead to native inflows? American Economic Review, 90(2), 360–367. Dincer, B., Özaslan, M., & Kavasoglu, T. (2003). A report on socio-economic development index for provinces and regions in Turkey. Ankara: Turkish Republic State Planning Organization Publication No. 2671. Gedik, A. (1996). Internal migration in Turkey, 1965–1985: Test of some conflicting findings in the literature. Working Papers in Demography No. 66, Research School of Social Sciences, The Australian National University, Canberra. Gould, E. D., Lavy, L., & Paserman, M. D. (2004). Does immigration affect the long-term educational outcomes of natives? Quasi-experimental evidence. NBER Working Paper No. 10844. LaLonde, R., & Topel, R. (1997). Economic impact of international migration and economic performance of migrants. In M. R. Rosenzweig, & O. Stark (Eds.), Handbook of population and family economics (pp. 799–850). North-Holland: Amsterdam. Mete, C. (2004). Education finance and equity in Turkey. Paper commissioned for the Turkey ESS. Washington, DC: World Bank. Munshi, K. (2003). Networks in modern economy: Mexican migrants in the U.S. labor Market. The Quarterly Journal of Economics, 118(2), 549–599. Rose, E. (1999). Consumption smoothing and excess female mortality in rural India. Review of Economics and Statistics, 81(1), 41–49. Staiger, D., & Stock, J. (1997). Instrumental variables regression with weak instruments. Econometrica, 65(3), 557–586. Tansel, A. (2002). Determinants of school attainment of boys and girls in Turkey: Individual, household and community factors. Economics of Education Review, 21(5), 455–470. World Bank. (2005). Turkey education sector study: Sustainable pathways to an effective, equitable, efficient education system for preschool through secondary school education. Washington, DC: World Bank.