Schooling Opportunities and Intergenerational Educational Mobility in Turkey: An IV Estimation Using Census Data

Abstract We estimate the intergenerational transmission of education in Turkey using micro-data from the 1990 and 2000 censuses and an instrumental variable (IV) approach. We construct a historical series of provincial enrolment rates by gender to isolate the environmental effect on parental education. The results reveal that intergenerational educational mobility increases over time through a stronger decrease in the transmission of paternal education. The improvement is larger for boys, and the transmission is higher for mother-daughter pairs and in the case of poorer educated parents.


Introduction
Educational mobility across generations is one of the major indicators of a country's performance in terms of social mobility and, by extension, equality of opportunity (fairness, meritocracy, inclusiveness) and economic efficiency. Today's most developed areas are characterised by relatively greater mobility. Based on evidence from 42 countries over 50-year trends, Hertz et al.'s (2007) findings suggest that the more equal a society is, the weaker the intergenerational correlation. Similarly, Causa and Johansson (2011) compare the more equal Nordic countries with southern European countries while Blanden (2013) contrasts Nordic countries with developing countries, southern Europe and France using various social mobility indicators.
The major difficulty in quantifying the magnitude of the intergenerational transmission of education lies in the identification strategy. Björklund and Salvanes (2011), Black and Devereux (2011) and Holmlund, Lindahl, and Plug (2011) identify three approaches that promise to overcome the issue of the omitted variable bias: twins, adoptees and instrumental variable (IV) estimates. All three seek to estimate the 'net' impact of parental education on the offspring's education. The first two approaches seek to differentiate the nature factors (endogenous genetic/ability factors that affect parental education and its transmission to the offspring) from the nurture factors, while the third approach consists in defining factors that affect parental educational environment, exogenous to their ability. Black and Devereux (2011) and Holmlund et al.'s (2011) surveys suggest that results are highly sensitive to data and the methodology used. Holmlund et al.'s (2011) empirical work applying all three methods to the single dataset of the Swedish register finds that the results are sensitive to the identification strategy and that the OLS estimates are greater than the IV estimates. Overall, although parental education has a causal effect on children's educational outcomes, the evidence is mixed concerning its significance, its magnitude, and its gender-differentiated dimension (mother versus father and daughters versus sons). Surveys including information on twins or adoptees that are used in measuring intergenerational transmission of education are rare, such that the IV methodology is the most commonly used; furthermore, 'the IV approach is preferable to twin/adoptee strategies as it isolates the effect of an exogenous change in education of parents' (Black & Devereux, 2011, p. 1527. Instrumenting for factors that are likely to influence parental education enables discussing policy options that could improve educational opportunities and thereby social mobility and economic development, in line with the equal-opportunity approach (Roemer, 1998;Roemer & Trannoy, 2015).
The opportunities (environmental factors) can be measured through a number of variables. Prior to intergenerational studies, exogenous variables related to the schooling environment have been exploited as instruments to examine the causal effect of education on various types of outcomes, starting with returns to schooling (for an early literature assessment see Card, 1999). Following Acemoğlu and Angrist (2001), a strand of literature has been using compulsory schooling laws (CSLs) to investigate the causal impact of education expansion on various outcomes. This is motivated by the fact that changes in compulsory education are assumed to be exogenous to ability. Similarly, several studies have used the 1997 Compulsory Education Law extending compulsory schooling from five to eight years to assess the impact of education in Turkey on various outcomes. 1 However, the 1997 reform is too recent to explore in assessing intergenerational outcomes. Although a number of changes have taken place before 1997, they did not have far-reaching effects. A significant issue was the gradual aspect of compulsory education attainment itself. Factors such as resource constraints, parental compliance and governmental enforcement capacity has varied over time and space which had heterogeneous effects on compulsory school attainment. In this respect, we follow the empirical literature that uses variation in treatment intensity of change in the educational environment (for the literature on the evaluation of schooling programmes in developing economies with varying implementation intensity see for example Berlinski & Galiani, 2007;Berlinski, Galiani, & Gertler, 2009;Chin, 2005;Duflo, 2001;King & Behrman, 2009). As we do not have a specific reform, we exploit the cohort-specific, time-and space-varying aspects of mandatory schooling attainment affecting all the parental cohorts subject to our study.
Using micro data from the 1990 and 2000 census surveys 2 provided by the Turkish Statistical Institute (TurkStat), we estimate the intergenerational transmission of education in Turkey. More precisely, our estimated outcome is the probability of a child aged 16-17 having completed at least one level of post-compulsory schooling, defined as having obtained a lower secondary diploma. Using various publications of the National Education Statistics, we construct a unique historical series of provincial enrolment rates at compulsory level, which we use as an instrumental variable (IV) that accounts for the parents' educational opportunity. We run our estimates by child and parent gender and separately for the years 1990 and 2000 to assess the evolution of the transmission and mobility.
Determinants of various types of educational outcomes have been widely studied for Turkey. 3 The only work that explicitly addresses the issue of intergenerational mobility of education in Turkey is Tansel (2015) who undertakes a cohort analysis using the 2007 Adult Education Survey. In conformity with ours, her estimates suggest that mobility has been slower for girls, and for children with parents having a poorer educational background. Our findings show that while the impact of paternal education is greater than maternal education in the standard probit estimates, it decreases in the IV estimates such that maternal education becomes more important; and the effect of maternal education becomes greater on daughters. These imply that the environmental improvement has been less favourable for the female population and that this gender inequality has been intergenerationally transmitted.
The next section describes the data and the construction of the IV, Section 3 presents and discusses the probit and IV estimation results. Section 4 concludes.
Schooling opportunities and educational mobility in Turkey 1397

Data
There are two practical issues in evaluating intergenerational mobility: the source of data containing information on child and parent education; and the method and data source related to the identification strategy. Most studies on the subject come from advanced economies, and mostly Nordic countries where data availability is largest. Many countries, including Turkey, lack the administrative data that would provide the most comprehensive and accurate information on child's and parents' education.
Given these issues, we have opted to estimate intergenerational mobility using census data that is the most representative alternative source to administrative data; and we have constructed local enrolment rates to instrument for parental education. 4 Each of the census data (1990,2000) consists of a 5 per cent nationally representative sample. The major advantage of the census survey is that it is the most representative population data at a disaggregated regional level, which is crucial to our understanding of educational environment as captured by the IV. On the flip side, it bears weaknesses in terms of health and socioeconomic variables, income, extensive information on family characteristics, and so forth.
We estimate the probability of a child having obtained at least one post-compulsory diploma (lower secondaryalso called upper primary or middleschool diploma). The census questionnaires allow identifying parent-child pairs based on co-residence, in other words we do not have information on family members residing outside the household, if any. Our dependent variable is a child with at least one parent present in the household, and aged 16-17, which corresponds to 1974-1975 or 1984-1985 cohorts respectively for the 1990 and 2000 censuses. 5 Table 1 gives information on the basic data restrictions. As school starting age is six or seven, 6 lower secondary school completion corresponds to eight years of education corresponding to 14 or 15 years of age. We set the upper limit of the child's age to [16][17] in order to have the largest sample of children co-residing with parents and avoid the selection bias. Above 16-17 years of age, children are more likely to leave the parental household to attend higher education, 7 or for other reasons: girls are likely to leave for marriage 8 and boys for compulsory military service. We consider boys and girls separately, and for each gender we estimate the impact of the father and the mother separately. 9 The age range is between 31-54 years for mothers and 31-64 years for fathers. We omit households with missing information on child or parent, with a child born abroad, and polygamous and single-parent households. We also drop children that are grandchildren in the household as parental information is unknown. Parental education is measured as the highest level of diploma obtained. Admittedly, this somewhat underestimates the actual level of schooling since drop-out information is missing in the data.
The instrumental variable, the provincial primary school enrolment ratio, is defined as: where g, r and t respectively stand for gender, province and year.  , 1990census 1974-19751936-19591926-1959Birth year, 2000census 1984-19851946-19691936-1969 Missing information on child or parent Child born abroad Households other than 1 mother and 1 father omittedpolygamous or single parent households Omitted children (16)(17) If child is grand-child (parent unknown) The enrolled primary school population is taken from the National Education Statistics publications. Primary school types vary across time, and we sum all types of public schools that include compulsory schooling. 10 Population by age comes from five-year Census of Population publications that provide age in categories. 11 We take the sum of categories 5-9 and 10-14, such that we are constrained to define primary school age population as 5-14 years of age. We then match parents with the primary school enrolment rates in the province in which the parent lived at the age of seven. As enrolment ratios are available every five years the match is undertaken by age brackets as given in Table 2. The table should be read as follows: parents aged between 31 and 34 years in the 1990 census are born between the years 1956-1959, they start school between the years 1963-1967 (assuming a school starting age of seven), and are assigned the local enrolment ratio for the year 1964-1965. Our instrument requires that parental residence at primary school age is known so that the proper local enrolment ratio can be matched with the parent when aged seven years. Consequently, in our two-stage estimations we only keep households where parental residence location at primary school starting age can be identified, which requires additional restrictions. The residence information is deduced from the current and birth residence location, and the recent migration experience (having migrated in the last five years) as given in Table 3: if current and birth residence location are the same then the parent is defined as 'born local'. If birth and current residence locations are not the same and that the parent has migrated five years ago then s/he is defined as a 'new migrant'. For the other cases, households are omitted as parental residence location at primary schooling age cannot be clearly identified or matched with local enrolment data ('birth abroad'). Finally, as geographical administrative units vary over time, we define a harmonised provincial classification on the basis of the 67 provincial breakdown (corresponding to the year 1985) for compatibility across census surveys and enrolment ratios. 12 Our instrument bears a number of limitations: it does not explicitly account for factors such as school/teacher quality, existence of joint classes, early drop-out, irregular attendance, class repetition, late entry, and so forth. In addition, we have opted not to use the number of schools (proxy for supply) or the pupil-teacher ratio (proxy for school quality) because the heterogeneity across schools is vast in  1956-19591951-19551946-19501941-19451936-19401931-19351926-1930School starting year 1963-19671958-19621953-19571948-19521943-19471938-19421933-1937Corresponding year of enrol. ratio 1964-19651959-19601954-19551949-19501944-19451939-19401934-1935 1966-19691961-19651956-19601951-19551946-19501941-19451936-1940School starting year 1973-19771968-19721963-19671958-19621953-19571948-19521943-1947Corresponding year of enrol. ratio 1974-19751969-19701964-19651959-19601954-19551949-19501944-1945  Schooling opportunities and educational mobility in Turkey 1399 terms of classes and students by school, and because a historically low pupil-teacher ratio in less developed regions is rarely an indicator of quality. Given the prevalence of mixed public primary schools, the number of schools fails to capture the gender-differentiated dimension of primary school attainment. Another difficulty resides in distinguishing between supply (quantity and quality), demand (compliance) and compliance enforcement capacity (non-dissuasive and/or uneven application of sanctions). 13 In many developing countries access to educationeven at compulsory levelsis related to all these three factors that exhibit temporal and spatial heterogeneity. Potentially, households' demand for compulsory education may not be exogenous to (parental) education, and in this sense our results cannot be interpreted as pure causal effects. Nevertheless, we believe our instrument provides a more realistic measure of the factors that are likely to affect parental education and thereby its transmission, beyond the expansion of supply per se. More, these possibly endogenous demand factors are also likely to be subject to (exogenous) change (variation in governmental enforcement, settlement size, local development and labour market conditions may affect norms and values). Figure 1 gives the evolution of the average enrolment rate and the inequality of primary school enrolment rates across provinces, by gender, using education statistics and censuses (as defined in E g;r;t ð Þ ). Figure 2 provides similar evidence accounting for schooling years of provincial population by five-year cohorts, by gender, using the 2000 census data. The inequality figures show that, although decreasing, there has been a significant heterogeneity across provinces and cohorts, and a persisting overall gender gap in terms of both primary school enrolment rates and schooling years in general. Note that there is also a significant heterogeneity in terms of gender differences at the provincial level.
We use a number of controls from the information that is available in both censuses; Table 4 provides summary statistics for all the variables, according to the sample used in the estimations. These controls include siblings by gender, current residence location type (city centre, district, village), house ownership (interacted with residence dummiesdistrict centre or villagein order to control for house value), migration information (being born at the local residence location is opposed to having moved within the last five years), province fixed effects and an indicator of local employment prospects. One important factor that potentially affects educational decisions is the probability of finding a job: one is more likely tobe encouraged tocontinue studies if among the population that hold the targeted diploma level (post-compulsory) there are a large number of wage earners. More particularly, in the case of women, a female employment indicator may capture a kind of neighbourhood effect that counteracts social norms and values inhibiting female labour-force participation. The indicator of employment prospects is constructed as the share of non-agricultural wage-earners having a post-compulsory diploma in working age population (aged 15-65 years) having a post-compulsory diploma at the local level (district) 14 and by gender in order to account for the latter effect. Our calculations using census data suggest that the level of this indicator is strikingly low: respectively  6 1934-35 1939-40 1944-45 1949-50 1954-55 1959-60 1964-65 1969-70 1974-75 Mean enrol. rate, female Inter-provincial enrol. rate Gini, female Mean enrol. rate, male Inter-provincial enrol. rate Gini, male 20.2 and 4.7 per cent for men and women in 1990, and 27.7 and 7.1 per cent for the year 2000. The rates are somewhat increasing which implies a certain improvement in returns to education however it remains low, especially in the case of women.

Estimation and results
We first run a standard probit model to estimate the probability of attending post-compulsory education, given in Equation (2): where S c i;j is the probability of having obtained at least a lower secondary level diploma. This indicator takes the value of 1 if the child i in household j has obtained a post-compulsory diploma and takes 0 if s/he has not. S p i;j denotes parental education of child i in household j. X i;j denotes a set of controls and v c i;j is the error term. Parental education (S p i;j ) falls into five categories: no diploma, 15 primary, lower secondary, upper secondary school and post-secondary diplomas; corresponding years of schooling (continuous parental education) are respectively 0, 5, 8, 11 and 15.
To deal with the issue of endogeneity we adopt an instrumental variable framework where we follow two-stage estimation procedures. When using continuous parental years of education we estimate a two-stage IV-probit model; when using categorical parental education we estimate a twostage residual inclusion (2SRI) regression, following Terza, Basu, and Rathouz (2008) who show that for non-linear models the 2SRI procedure yields consistent estimates in the second stage.
The first stage is defined as: where S p i;j and X i;j are the same as in Equation (2). E p i;j is the primary school enrolment ratio in the province r (E g;r;t ð Þ ) of child i's parent when the parent is aged seven (for the correspondence between the parent's year of birth, school starting age and provincial primary school enrolment ratio see Table 3). In the case of the IV-Probit model this stage is an OLS estimation using continuous parental education, whereas in the case of the 2SRI model it is an ordered probit estimation using categorical parental education. The second stage of the IV-probit model is defined as:  5 1925-29 1930-34 1935-39 1940-44 1945-49 1950-54 1955-59 1960-64 1965-69 Average years of sch. Inter-provincial years of sch.Gini Inter-provincial years of sch. Gini, female Inter-provincial years of sch. Gini, male Average years of sch., female Average years of sch., male whereŜ p i;j is the estimated value of S p i;j obtained from the OLS in the first stage. The second stage of the 2SRI model is defined as: where u p i;j correspond to the residuals of the ordered probit estimated in Equation (3). Tables 5 and 6 provide the estimation results separately for each census, for all children (with a dummy for the child gender) and for each of the mother-son, father-son, mother-daughter and fatherdaughter pairs. All estimations include province specific fixed effects which are expected to capture unobservable factors across provinces that are likely to affect both the parent and the child. Table 5 gives the results of the probit and IV-probit estimations where parental education is converted into schooling years, the sample used is subject to both the basic and additional restrictions as given in Tables 1 and 2, and standard errors are clustered at the province-level. Table 6 displays the summary results of the probit and 2SRI estimations where parental education is categorical, using the same sample and the same set of variables as in Table 5, here standard errors are bootstrapped. We also run estimates without the instrument using the basic restrictions only (see Online Appendix Tables A1 and A2 respectively for the continuous and categorical parental education). 16

Parental education
The standard probit estimations using continuous parental education (Table 5) suggest that overall, paternal education is greater, which is in conformity with the majority of the estimates of parental correlates in Turkey suggesting a larger effect of the father's education on children. From 1990 to 2000 the impact of parental education on boys decreases (the marginal effect of maternal education is 0.032 and 0.023, that of paternal education is 0.042 and 0.027 respectively in 1990 and 2000), whereas that on girls remains fairly stable (the marginal effect of maternal education is 0.030 and 0.030, that of paternal education is 0.033 and 0.031 respectively in 1990 and 2000). Consequently, intergenerational educational mobility seems to have increased only in the case of boys, through a decrease of both maternal and paternal education transmission. Also, note that the marginal effects of parental education on daughters seem less gender-differentiated.
However, the IV-probit estimates which take into account the educational environment, suggest that the paternal impact is lower, whereas the marginal effects of maternal education are fairly the sameexcept on daughters in 1990compared with the standard probit estimates.
As such, the IV-probit estimates provide a different picture where the marginal effects of maternal education are greater than paternal education, and where mobility increases for both boys (the marginal effect of maternal education is 0.032 and 0.022, and that of paternal education is 0.032 and 0.016 respectively in 1990 and 2000) and girls (the marginal effect of maternal education is 0.039 and 0.029, and paternal education is 0.024 and 0.016 respectively in 1990 and 2000), 17 through a decrease in the transmission of both parents' education, but mostly the father's. These results imply that the improvement in the educational environment mostly benefited the male population and that the decrease in the transmission has been faster in time. The impact of maternal education is more persistent despite some improvement in opportunities. 18 The IV-probit results further suggest that the mother's education has a greater impact on both daughters' and sons' educational outcomes, and that this impact is more persistent especially on daughters. Consequently, there is still room for policies targeting the enhancement of women's educational attainment that potentially affect their child'snamely the daughter'seducation and raise intergenerational educational mobility of the female population.
In order to account for possibly non-linear effects we run an IV estimation using a 2SRI strategy where parental education is categorical in the first stage (Table 6). We here find a threshold effect: Schooling opportunities and educational mobility in Turkey 1403 Table 5. Marginal effects of the probability of having completed at least lower secondary school having a parent with a five-year education or less significantly decreases the probability of postcompulsory schooling and this negative effect is larger in the case of maternal education. Again, the effect of transmission decreases across years for both sons and daughters. The decrease in the transmission of paternal education is greater, and the level of transmission is strongest in the case of mother-daughter pairs. For boys, having a mother ( By way of comparison, using the 1994 Household Income Expenditure Survey, Tansel (2002) estimates the probabilities of obtaining primary school, middle school (lower secondary) and high school diplomas respectively at ages 14-19, 16-19 and 19-20, without using an instrument. She finds that the impact of both parents' education is greater on daughters (except one case), meaning lesser mobility for girls. She also finds support for a greater maternal impact on daughters, however, the impact of the mother's education relative to the father's varies across specifications, and are not always found to be significantly different. Her estimates using categorical parental education also suggest that having poorer educated parents decrease girls' and boys' educational outcomes. Using the 2007 Adult Education Survey and controlling for cohorts, Tansel (2015) finds that intergenerational mobility increases for younger cohorts. As in her previous study, she finds that the mobility is slower in the case of the female population and that the transmission of education is stronger when parents are poorly educated. We seek to capture the impact of schooling expansion by instrumenting for parental education and consistently find a greater impact of maternal education, and an improvement in the intergenerational educational mobility largely due to a decrease in the transmission of paternal education through the expansion of schooling opportunities that mostly benefitted the male population.

Other factors
We find a negative correlation of siblings that increases with the number of siblings, on both genders for both years, which is a sign of resource constraints. The effect decreases in time, especially for girls. These effects are gender biased: the negative externality of male sibling(s) is higher than that of female sibling(s), and overall, the negative impact of siblings remains higher on boys (of about one third compared to girls). However, as our family unit definition rests upon the co-residence criterion, we are unable to capture the actual number of siblings, hence their composition. Using the 1998 Turkey Demographic and Health Survey (TDHS), Dayıoğlu, Kırdar, and Tansel (2009) estimate the impact of siblings on the enrolment of urban children aged 8-15 using twins-per-birth as an instrument and show that the sibship size is insignificant, but that sibling composition (order) in poorer households matters.
House ownership is the only variable that could proxy the wealth effect in the model. 19 Although it is negative and significant in the full sample, the effect is limited and mainly exists for girls. The negative significant effect could be indicating that the benefits (returns) of schooling outweigh the costs for girls in poorer households. Such a strategy could help improve lifetime wellbeing not only through employment income (labour market), but also by securing better mating opportunities (marriage market), which may be an important motivation in a low female labour force participation rate environment. Most studies find a positive relation between household income and child education (for example, for Turkey Smits & Hoşgör, 2006;Tansel, 2002). The difference may be due to the fact that household ownership is only a proximate indicator of wealth, and that we are unable to deal with the income effect. Nevertheless, our contrasting results, especially in the case of girls, may be due to the fact that we instrument for parental education and/or that the census data, which allows controlling for a larger and more disaggregate level of geographical units, may be capturing a larger number of local (for example normative) effects. For instance, Dildar (2015) who analyses the relation between conservatism and female labour force supply in Turkey using the 2008 TDHS, finds that household wealth has a negative impact on female labour force supply. To the extent that the latter also affects child education, these issues deserve further investigation using appropriate identification strategies.
Being born locally as opposed to being a recent migrant does not have a significant effect in 1990. The effect of being a native turns positive and significant for the 2000 census and only for boys. We have to keep in mind that the restricted model only captures the immigration effect for those who moved recently (no later than five years). This is an expected outcome, as the 1990s are marked by forced migration and population displacements due to armed conflicts in the eastern regions arising from the Kurdish issue (Kurban, Yükseker, Çelik, Ünalan, & Aker, 2007); but also by poor regional policies and decreases in agricultural subsidies affecting other less-developed regions. Overall, these may have contributed to the migration of mainly disadvantaged households, perpetuating their deficiencies in terms of educational outcomes in their new location of residence.
In conformity with previous findings of studies on the determinants of educational outcomes in Turkey, the effect of residing in a village is negative and strong but its effect is decreasing. Compared to boys, the marginal effect of residing in a village is higher for girls (twice as high) and the declining trend from 1990 to 2000 is less pronounced. 20 It is reasonable to assume that private returns to education remain lower in rural areas where low-skilled farming activities prevail. Furthermore, it may be that education supply remains low and that the cost of schooling is relatively higher in terms of commuting and the relative income of rural households, which is likely to be lower. In rural areas, marginal returns to education are lower for girls, who are more likely to be caught up in household chores and engaged in unpaid family work. Alongside policies aiming at decreasing schooling costs, rural development policies that would contribute to the development of non-rural wage activities may also increase the expected returns to education and post-compulsory attainment in rural areas.
Compared to residing in a provincial centre, residing in a district (or county) has a positive and significant coefficient. The results suggest that there have been significant improvements in the development of smaller urban centres particularly for girls: living in a district centre as opposed to living in a provincial centre had no impact on their schooling outcomes in 1990, but becomes significant and positive in 2000. For the overall effect, it can be argued that district centres have the optimal scale and provide a better school environment than villages or cities. Villages are sometimes very remote and lack enough students or teachers to have proper classes at each level. Cities, on the other hand, are more segregated in terms of class size and school quality. Depending on the scale of suburban area and migration inflow, a city centre might have over-sized classes, an insufficient number of teachers for separate classes, and higher drop-out rates.
Another factor that potentially affects education demand is the expected returns to schooling, which depends on the local development level and more specifically on labour market conditions. We construct an employment prospects indicator to capture such effects. In the narrower sense, it can be interpreted as the aspirations of post-compulsory graduates in terms of job opportunities in the local non-agricultural labour market. Besides cultural and regional factors, local labour market conditions provide a material motivation for the schooling of girls and may also decrease the role of other factors. Tansel (2002) and Rankin and Aytaç (2006) also consider local labour market conditions in assessing educational outcomes. Tansel (2002) considers the shares of employment in industry and services separately at the provincial level. She mostly finds an insignificant impact, but when significant, employment in industry and services respectively has a positive and a negative impact. Rankin and Aytaç (2006) use province level information on the share of employment in agriculture and the ratio of male to female employment, and find insignificant results. The census data allow us to construct a more precise indicator where we can add a gender dimension along with the activity type and employment status (non-agricultural wage-earners), and education level (having obtained post-compulsory diploma) at the district level. We find that the indicator has a significant impact across all specifications. We can interpret this result as an incentive effect since the marginal effect of the indicator on girls' probability of post-compulsory schooling doubles that of boys. The finding also suggests that comparing two regions where the employment prospects gap is increasing, the marginal effect would be very substantial, and that the development of local labour markets and the increase in the participation of women, could enhance girls' post-compulsory schooling. 21 It appears that in Schooling opportunities and educational mobility in Turkey 1409 countries with significant gender differences, constructing gender-specific indicators of local labour market conditions might yield more suggestive results, and contribute to a better understanding of gender differences in educational outcomes.

Conclusion
We have estimated the intergenerational transmission of education in Turkey using census micro data for the years 1990 and 2000. We instrument parental education by constructing a historical series of compulsory education enrolment rates at the provincial level. As such, we propose a strategy that is replicable in other countries where administrative data is unavailable, but census micro data is available and includes parental birth province. We propose that the instrument has the benefit of controlling for parent's educational environment in a more realistic way, particularly in countries where (i) de jure educational reforms are unevenly implemented de facto, (ii) there is no reform but where compulsory school attainment varies at the sub-national level over time and by gender (iii) there are (unobservable) heterogeneities across schools that cannot be precisely matched with parents and their gender.
We find that the impact of parental education on the likelihood of completing a post-compulsory education level is decreasing from 1990 to 2000. Consequently, intergenerational educational mobility has increased over time. The IV estimations suggest that the transmission of parental education is lower than the standard probit estimates, especially for the 2000 census. Moreover, the impact of the father's education decreases whereas that of the mother's is more persistent. Comparing IV estimation results from the 1990 to 2000 censuses, we argue that the transmission of parental education decreases, however, the transmission remains higher for daughter-mother pairs.
When parental education levels are considered in categories, the results suggest some non-linearities: having a mother with an education level lower than post-compulsory constitutes an important obstacle to the offspring's probability of completing post-compulsory education. This marginal threshold effect is greatest for mother-daughter pairs. Among other factors, living in rural areas and localities with poor labour market conditions are important impediments to post-compulsory schooling, especially that of girls. Given the current persistence of gender differences in post-compulsory schooling, these findings suggest that the decrease in the gap through greater schooling (and other) opportunities for the female population may additionally have intergenerational benefits and eventually improve their social mobility.
Turkey. The same figures are −0.259 (mother) and insignificant (father) for Greece in 2001, and −0.350 and −0.295 for Turkey in 2000. 17. The increase in the IV-estimate for mother-daughter pairs in 1990 may be due to the persisting heterogeneities in compulsory schooling that we cannot account for and which were salient during the early decades, along with the stagnant female enrolment rates and inter-provincial inequalities (notably in the 1950s, see Figure 1). 18. There is an ongoing debate on the differentiated impact of intergenerational transmission of education according to parental gender and the methodology adopted. Typically, the literature suggests that twin-studies find a greater paternal influence as opposed to IV-studies where the maternal influence is often found to be larger (for a recent reappraisal, see Amin, Lundborg, & Rooth, 2015). Our strategy accounts for the environmental factors, as such it does explicitly differentiate between the nature (genetic) and nurture effects. Nevertheless, assuming unobserved genetic influence constant for a given sample, the variation in the impact of parental education estimates between the standard probit and IV estimates can be interpreted as the impact of the environment on the nurture effect. The latter is found to be more persistent in the case of mother-daughter pairs (on the developmental impact of allocating greater resources to the female population see for example Duflo, 2012;Schultz, 2002). 19. In all models, we interact the effect of house ownership with the residence information to account for rural-urban differences. Thus, the marginal effect reports the net effect of house ownership interacted with residence dummies. 20. Kırdar et al. (2015) show that the 1997 reform has been particularly efficient in enhancing school attainment for the newly mandated grades and decreasing the rural-urban gap. However, its impact has been weaker for girls' post-compulsory attainment in rural areas. 21. These findings may overlap with the rural development issue; unfortunately, we are not able to add the rural-urban breakdown to our indicator as it critically reduces the number of observations in each cell.