Economic geography and human capital accumulation in Turkey: evidence from micro-data

ABSTRACT This study examines the impact of market access on human capital accumulation in Turkey. Using individual-level data, the analysis explores the background of human capital accumulation, combining market accessibility, wages and human capital development. Upon the treatment of wages as an endogenous covariate of interest and overtime work as an exogenous source of variation, we find evidence that the impact of market access on human capital development vanishes in ways not predicted by the augmented New Economic Geography set-up for human capital accumulation. Findings confirm that economic policies may be effective in reducing regional variation in human capital endowments.


INTRODUCTION
Neoclassical theory used education as an important element of cross-country differences (Mankiw et al., 1992). Meanwhile, endogenous growth theory emphasized that human capital stimulates innovation and technological advances that further spur economic growth (Lucas, 1988;Romer, 1994). While seminal contributions of growth theories are integrated into regional science (Faggian et al., 2019;Manca, 2012;Ramos et al., 2012), recent advances in urban economics also highlight that education is an important element for city-level differences in resilience and economic growth (Glaeser et al., 2014;Glaeser & Saiz, 2004;Moretti, 2004). A rising consensus on the positive impact of human capital development on economic growth is an important motivation to explore the reasons behind the variation in human capital development. Different theoretical approaches elucidate how educational human capital is dispersed across space. The central discussion of the urban economics literature is the interconnection between city-level economic prospects and individuals' location decision. Glaeser et al. (2014), Mori and Turrini (2005), Berry and Glaeser (2005) and Carlsen et al. (2016) underline that more populated growing cities offer higher wages to skilled individuals and create more incentives to move to central locations. Similarly, Combes et al. (2008Combes et al. ( , 2012 and Behrens et al. (2014) point out that cities attract more educated people. While this approach discusses how educational human capital differences evolve over cities, it does not incorporate the impact of geography. However, geographical advantages and agglomeration economies are crucial aspects of the dispersion of economic activity and human capital accumulation. In recent modifications to the New Economic Geography (NEG) literature, Redding and Schott (2003) explore the endogenous accumulation of human capital with the help of agglomeration economies. Redding and Schott (2003) offer a model that links urban economics and the NEG literature. It rests on the assumption that returns to education affect human capital accumulation. However, as the model restricts interregional mobilities, urban economics mechanisms are not one-toone applicable. Instead, the general equilibrium framework of the NEG and the agglomeration of economic activity form the backbone of proximity and human capital accumulation relation. Central regions benefit more from different externalities of agglomeration economies, which in turn influences their profitability. As profitable areas are likely to offer higher wages, it is more beneficial for individuals to increase their human capital at central locations.
While the use of geography is not new to economic theory (Gallup et al., 1999;Krugman, 1991), Redding and Schott (2003) and Venables (2004a, 2004b) use market access and potential (henceforth market access) as a proximity measure to explain cross-sectional differences in the level and sources of economic activity. While the former uses geography to understand regional variation in human capital accumulation, the latter approach focuses on inequalities at large. Even though NEG points to high-profit opportunities in high-market access areas (Fujita et al., 1999), it does not fully explain how profitability can be transferred to earnings for the production factors. Redding and Schott (2003) highlight that high profitability at central locations enables firms and regions to transfer extra value-added to production factors (i.e., skilled premium to educated workers). This augmented version of the NEG set-up asserts that rising skill premium in high-market access areas will act as a motivation for individuals to invest more in educational human capital development. Evidence from different developed countries confirms that more remote regions accumulate less human capital (Diebolt & Hippe, 2018;Faíña & Lopez-Rodriguez, 2006;Karahasan & López-Bazo, 2013;López-Rodríguez et al., 2007. Baldwin et al. (2011) argue that NEG models and their variants are useful to examine the preferences of firms to locate in larger markets. Measured by the market access effect, this process creates a circular causation as market size is mostly shaped by income and jobs, both of which are conditioned at the level of education. These discussions explain how central locations realize higher education level and create various positive economies that attract firms and individuals. As Baldwin et al. argue, market accessibility opens new lines of discussions on migration and mobility.
Linking geography and human capital accumulation raises concerns for peripheral areas' isolation. However, other dimensions of human capital dispersion also deserve attention. For instance, the attractiveness of high market areas not only influences individuals residing in that region but also affects individuals in other regions. Redding and Schott (2003) argue that although factor mobility is negligible at the cross-country level, its impact can be visible for intra-country analyses. Crozet (2004) and López-Rodríguez et al. (2007) highlight the importance of mobility and posit that market access can potentially influence individuals' migration to central locations. However, their findings on European Union (EU) regions show less room for interregional mobility within the EU area. The impact of migration stands out as an important empirical matter, which reflects the arguments of urban economists that educated and skilled individuals locate in regions that offer higher wages (Glaeser et al., 2014). We believe that it highlights the importance of the triangular relationship among wages, human capital accumulation and market accessibility.
This paper aims to test the causal impact of market access on human capital accumulation, considering the endogeneity of wages. Prior evidence highlights the impact of market accessibility on wage distribution. However, empirical studies linking market access to human capital are relatively limited. To our knowledge, no attempt has been made to link market accessibility to wages and education in the same framework. Therefore, this study uses micro-data and focuses on the background of Redding and Schott's (2003) model. The central aim is to critically evaluate the causal link between market access and human capital accumulation for a developing country: Turkey. Our central hypothesis is that individual factors, and specifically returns to education, have the potential to take out the influence of market accessibility, which is central to Redding and Schott's model. This study contributes to the existing literature in several ways. First, although regional disparities have been extensively investigated in Turkey (Dogruel & Dogruel, 2003;Filiztekin, 1998;Gezici & Hewings, 2007), our knowledge on the agglomeration of economic activity is relatively limited (Karahasan et al., 2016). Note that the historical origins of regional disparities have received interest within the international literature. Ezcurra and Rodríguez-Pose (2014) express the extent of spatial disparities and regional duality in Turkey. The peculiar developed west and underdeveloped east duality makes Turkey interesting to test Redding and Schott's (2003) model. Earlier evidence confirms that the distribution of human capital mimics the regional disparities in Turkey (Erdem, 2016). However, the determinants of this duality and the role of geography are relatively less investigated.
Second, from an international perspective, empirical studies mostly provide evidence from developed countries, leaving developing countries less investigated. López-Rodríguez et al. (2011) for Romania, López Rodríguez and Acevedo-Villalobos (2013) for Colombia, and López-Rodríguez and Runiewicz-Wardyn (2014) for Poland are few precedents that examine regional differences with the help of the NEG set-up for developing countries. However, these studies do not investigate the human capital aspects of regional disparities. To the best of our knowledge, López-Rodríguez et al. (2019) for Romania and Karahasan and Bilgel (2020) for Turkey are the only studies that use Redding and Schott's (2003) model for developing country examples. Based on an under-investigation of developing countries and the historical persistence of regional disparities, Turkey is an important example through which the influence of market access on human capital accumulation can be investigated. Moreover, the availability of micro-data to combine wages, market access and human capital accumulation enable us to test the causal claims of Redding and Schott's (2003) model, which has not been done so far for either developed or developing countries. Given the previously mentioned role of regional disparities and the underinvestigation of developing countries, Turkey is a suitable example for investigating whether market forces directly influence human capital accumulation or policy can still be effective in order to combat educational disparities. We argue that investigating a country with historical and persistent regional disparities is essential to compare and contrast market forces with policy implementation. Besides, lessons from developing countries will yield valuable information to sustain cohesion between developed and less-developed countries. We believe this will guide the policy-making scheme in developing countries that Economic geography and human capital accumulation in Turkey: evidence from micro-data have been suffering from policy ineffectiveness to combat regional disparities.
We also have to acknowledge several important features of Turkish labour markets, which make the Turkish case interesting. According to most recent figures of the Organization for Economic Co-operation and Development (OECD) and International Labor Organization (ILO), Turkey stands out in most of the labour market indicators within the Euro Area. For instance, the unemployment rate in Turkey is around 14%. This ratio is closer to the two Mediterranean countries heavily hit by the 2008 Global Financial Crises: Spain with 15% and Greece with 17%. However, it is well above those of the core countries such as France (9%) and Sweden (7%). More interestingly, the population share not included in education, employment and training is 26% as of 2019. This ratio is the highest within the Euro Area and also higher than the average of the upper-middle income countries of the world.
In addition to the overall labour market conditions, the institutional set-up and segmented structure of labour markets are crucial. The current state of the labour market regulations highlights a clear polarization for Turkey within Europe. Given its centralized administrative system, Turkey implements a nationwide and homogenous minimum wage across its territory. However, our prior knowledge confirms the existence of sizable disparities in the regional distribution of average wages (Elveren, 2010;Karahasan et al., 2016). This has been a long-lasting discussion for labour market regulations in Turkey. A careful comparison shows that as of 2019, Turkey has the lowest minimum wage within the Euro Area. Considering regional disparities, this homogenous minimum wage implementation is crucial. Additionally, based on recent figures for average hours worked (2013), Turkey has the highest working hours together with Greece in Europe. We have to highlight that we lack reliable information on the regional variation of productivity and average hours worked. However, given inflexible labour market regulations in a country with sizable regional differences, investigating the influence of market forces against that of public policy is a crucial element for investigating Turkey.
Finally, studies using aggregate regional data fail to control for policy-sensitive individual factors that are essential for regional development and agglomeration economies. Hence, more effort is required to examine the micro-aspects of localization and agglomeration economies (Duranton & Overman, 2005;Duranton & Puga, 2004). Baldwin and Okubo (2006) show that firm-level heterogeneity plays a dominant role in understanding the micro-aspects of agglomeration economies. Similarly, Mion and Naticchioni (2009) highlight that the microfoundation of agglomeration economies is a candidate to explore the so-called black box of agglomeration economies. To our knowledge, only a few empirical studies use individual-level data to understand the background of economic geography and agglomeration economies Based on the importance of individual heterogeneity, we focus on the endogenous dependence among wages, human capital accumulation and market access using micro-data. Although the theoretical model includes these three dimensions, the endogenous feedback, which we argue is an essential part of the model, has not been central to applied studies testing Redding and Schott's (2003) model. The interconnection between wages, market access and human capital accumulation deserves more attention and stands as an important challenge for testing the theoretical set-up of Redding and Schott's (2003) model.

THEORETICAL BACKGROUND
Agglomeration economies is an important element that explains the spatial distribution of economic activity. Krugman (1991), Krugman and Venables (1995), Venables (1996), Fujita et al. (1999) and Baldwin et al. (2011) argue that location decisions of economic agents are bounded by proximity to certain geographies. Redding and Schott (2003) augment the general equilibrium framework of the NEG models by allowing for the endogenous accumulation of human capital development.
Consider an economy composed of i [ {1, . . . , R} regions. Each region is endowed with L i consumers who have 1 unit of labour assumed to be initially unskilled. Individuals can endogenously decide whether or not to become skilled. The decision to become skilled is examined by comparing the wage premium and the cost of education: where w s i − w u i is the skilled premium (w s i and w u i represent wages to skilled and unskilled workers, respectively); Ω(z) ¼ h i /a(z) is the cost of education, where a(z) is the individual ability level; and h i is an institutional parameter (public provision of education) assumed to be homogeneous for regions of the same country. The cost of education is given as units of unskilled labour required for 1 unit of unskilled labour to become skilled. Rewriting equation (1), the skill-indifference condition is defined as: where a * i is the critical ability level. Any individual with an ability level higher than the critical level chooses to become skilled. Therefore, any individual with an ability level of a * i is indifferent between becoming skilled or staying unskilled (Redding & Schott, 2003).
The consumption side of the model follows the standard NEG formulation, where consumers prefer homogeneous agricultural products and different set of manufacturing goods. Within the model there are two sectors: agriculture exhibiting constant returns to scale; and manufacturing subject to increasing returns. 1 The manufacturing sector representative firm in region i maximizes the following profit function: where, P M ij is the price of the manufacturing good in region j of 1 unit produced in region i; α and β are input shares for skilled and unskilled workers; G i is the price index for manufacturing goods with input share of 1 − α − β; c i is the constant marginal input requirement; while c i F is the fixed input requirement, which is an inverse index for technological efficiency. Finally, x i = S R j=1 represents the total firm output in all markets. On the other hand, the model uses the iceberg transportation costs and underlines that for 1 unit of output to arrive from region i to j, T M ij > 1 units must be shipped. T M ij is assumed to capture all transportation costs from region i to j. Solving the profit maximization condition for the manufacturing sector results in the wage equation: where σ is the elasticity of substitution between manufacturing goods; and E j is the total manufacturing expenditure of region j, which includes both final and intermediate consumption. An important dimension of the model stems from the way geography is incorporated into the wage equation. Equation (5) shows the market capacity of an exporter region i and refers to the market access (MA). Similarly, equation (6) refers to the supply access (SA), where n j is the number of manufacturing firms and p j is the manufacturing prices in region j: Substituting equations (5) and (6) into (4) and taking the manufacturing price index to the right-hand side, we reach a new form of the wage equation: where ζ is a constant representing the impact of manufacturing intermediary goods. Returning to the skill indifference condition, one can link geography and human capital accumulation. Applying the zero-profit conditions in manufacturing sector by taking logarithms and totally differentiating equation (7), one arrives at the relationship between geographical position and human capital accumulation for the manufacturing sector: Redding and Schott (2003) argue that as regions become remote (decreasing MA) and given that manufacturing production is skill intensive, relative wages for skilled workers will be lower. From the skill-indifference condition, this implies a higher critical ability level for individuals to become skilled, which reduces the supply of skilled workers, thus limits the accumulation of human capital. While the theoretical model links market access with human capital accumulation over the mediating causal effect of wages, other forces can be effective. As argued in the urban economics literature, local demand for skilled individuals might be attracting more educated people in central locations. This can work over interregional mobilities. Moreover, urban-based amenities and/or rising income in central urban locations can attract more skilled and educated individuals. Besides, other individual factors can affect human capital accumulation. Our empirical set-up uses individual data that enable us to combine wages, human capital accumulation and market accessibility. Additionally, we include other individual characteristics that can influence human capital accumulation. Due to lack of micro-data on migration, we are unable to control for interregional mobilities. Therefore, we focus on the influence of wages and individual factors in understanding the robustness of market access and human capital relation rather than engaging in the evaluation of mobility.

Data and sample
We use individual-level data obtained from the quadrennial Earnings Structure Survey (ESS) administered by the Turkish Statistical Office (TurkStat) in 2014 (ESS-Turkstat, 2014). The survey provides information on employee earnings and wages along with age, gender, tenure, occupation, education, geographical region and sphere of economic activity. The survey is designed to produce estimates stratified by firm size, geographical region (NUTS-1) and type of economic activity. 2 The sampling method of the ESS consists of two stages. The first stage involves the selection of the sample business establishments using stratified simple random sampling; and the second stage involves the selection of wage-earning respondents from the sample business establishments. The 2014 ESS was conducted on 17,137 business establishments with at least 10 employees, of which 11,190 replied. From these establishments, data on a total of 164,204 wage-earning employees were retrieved.
Overall, 19.6% of the respondents had at most a primary education, 16.6% had a primary and a secondary school education, 27.3% had a high-school education, 8.5% had a vocational school education and 28.0% had a university education. 3 The geographical distribution of the ordered education levels by quantile is given in Figure 1(a). The highest education levels prevail in the Istanbul region and in Central and Northeastern Anatolia, by and large corresponding to a high-school education on average.
Economic geography and human capital accumulation in Turkey: evidence from micro-data The sample shows a volatile and overly right-skewed wage distribution with a range of about 94,860 TL. Figure  1(b) shows the average gross monthly wage of the respondents. The highest average wage levels prevail in the regions of Istanbul, Western Anatolia and Eastern Marmara. These regions are also characterized by dense industrialization and irregular urbanization, especially in the Eastern Marmara region.
Figure 1(c) shows the geographical distribution of market access at the NUTS-1 level. Western Turkey diverges from the eastern and south-eastern Turkey, confirming a dual economic structure in terms of market accessibility. Although the Western Marmara region has high market access, the region stands out as having low education levels and low wages compared with its surrounding regions.
A careful inspection of the spatial distribution of human capital accumulation, wages and market access (Figure 1a-c) (see Table A1 in Appendix A in the supplemental data online) confirms the spatial heterogeneity in Turkey. However, a more detailed observation shows that the similarity between market access and wage distribution is stronger compared with the spatial distribution of education. These descriptive findings highlight potential motivations. First, the spatial heterogeneity in Turkey and the extent of isolation across the eastern peripheral regions make Turkey central from an economic policy perspective. Second, the contradiction among education, wages and market accessibility makes the Turkish case even more interesting and suitable for our critical approach to link these three pillars, essential for regional planning and development.

IDENTIFICATION STRATEGY
In order to evaluate Redding and Schott (2003) critically and to test whether the impact of market accessibility is robust to the inclusion of individual characteristics and returns to education, we construct a design that combines wages, MA and human capital accumulation in the same framework. While this approach distinguishes our set-up from prior studies using aggregate regional data, it enables us to investigate human capital accumulation by considering the influence of individual factors, returns to education and market accessibility.
Our outcome of interest that represents human capital accumulation is the individual's education level (E) measured on an ordinal scale with J possible ordered outcomes, j ¼ 1, … , J. There is some underlying, unobserved continuous latent outcome E * that determines the observed values of the discrete outcome E. The outcome equation can be written as: where m j 's are the cut points that reflect the predicted cumulative probabilities at covariate values of zero. The latent outcome variable E * is defined as: where W is the individual's wage; X are the covariates of the outcome equation that include individual, employment and firm characteristics such as age, gender, tenure, type of employment, whether the individual works full-time, whether the individual has a permanent employment, weekly working hours, size of the firm, a dummy variable on collective bargaining agreement to capture the effects of improvements in the socioeconomic status of the employees and in physical and non-physical working conditions and quadratic terms of age and tenure to capture possible non-linearities; K consists of dummy variables on ISCO-08 occupations and on NACE Rev. 2 branch of economic activity and regional dummy variables; and ε is the idiosyncratic error term. The wage equation is given by the following model: where Z is a superset of X that additionally controls for excluded instruments; MA is the market access of the ith region; and υ is the error term. This wage equation is similar to that defined in Redding and Venables (2004b) and Boulhol and De Serres (2010). Note that in our empirical set-up, we first estimate equations (10) and (11) by a recursive bivariate ordered probit model, which enable us to consider the influence of wages and market access within the education equation. This step tests Redding and Schott's (2003), which does not include the influence of wages. Next, we estimate an ordered probit model, which only accounts for market access in the educational human capital accumulation model. This is a direct test of Redding and Schott's model (E * = X b + dW + hK + 1). Comparing the results from the recursive bivariate ordered probit and the ordered probit models enables us to evaluate critically the outcomes of Redding and Schott's model. 4 In order to better apprehend our empirical set-up, we need to refer to the endogeneity issue. While Redding and Schott (2003) assume that wages induce more human capital accumulation, human capital accumulation may drive wages up by the virtues of the Mincerian earnings equation and therefore wages may be endogenous. Moreover, human capital accumulation can result from region-specific amenities and labour market attractions. Omission of these factors, which can be strongly linked with market accessibility, will influence the causal claims of the theoretical model. This problem requires the use of an instrument that can plausibly be viewed as randomly moving around wages. Under the possible endogeneity scenario as we posit, unobservable confounders may be important in the determination of wages and human capital accumulation and thus, higher wages are likely to be correlated with the individual's education level. If the model is empirically supported, then the processes that determine human capital accumulation and wages cannot be thought independent.
Measuring proximity is at the centre of the theoretical model. Theoretically, SA and MA are highly correlated accessibility measures. Therefore, we only use MA in our empirical analyses. While Redding and Venables (2004b) measure market access from a gravity equation taking into account geographical distance and export flows across countries, Bosker and Garretsen (2010) use distance and regional demand potential (i.e., income, value-added, population). The data required to calculate MA based on a gravity equation are not available in Turkey. Additionally, we would like to compare our results with those of prior studies ( Redding and Schott's (2003) model. Therefore, we use the distance weighted purchasing power approach offered by Harris (1954): where Y j is the per capita income of region j; and D ij is the motorway distance between any pair of regions i, j, retrieved from the General Directorate of Highways of the Republic of Turkey. Economic geography and human capital accumulation in Turkey: evidence from micro-data Ignoring the ordered nature of the education variable for the moment, equation (10) is a reverse specification of the Mincer equation (Mincer, 1958(Mincer, , 1974. Incorporating equation (11) into equation (10) gives the theoretical set-up of Redding and Schott (2003) and the empirical construct of López-Rodríguez et al. (2007). While Mincerian earnings function models the logarithm of earnings as the sum of years of schooling and linear and quadratic terms of labour market experience, we posit that wages provide an incentive to invest in human capital; hence, education is a function of wages and possibly of linear and quadratic terms of labour market experience, among others. The challenge with this view is that higher education levels are likely to drive up wages by the virtues of the Mincer equation. Therefore, wage is likely to be endogenous in equation (10) due to reverse causality in the same spirit that schooling is endogenous in the Mincer equation. This implies that the unobservable factors that determine human capital are likely to be correlated with the unobservable factors that determine wages (cov(1, y) = 0). A failure to handle this reverse causality in equation (10) yields biased and inconsistent estimates.
If the correlation between the error terms of the education and wage equations is zero (ρ ¼ 0), then equation (9) can be estimated by a generalized ordered probit. If r = 0, then the unobservable determinants of wages are said to be correlated with the unobservable determinants of education, indicating that wage is endogenous. The possibility of endogeneity requires a joint estimation of equations (10) and (11) to obtain consistent and asymptotically efficient estimates (Bilgel & Karahasan, 2018).
To account for the possibility that wages may be endogenous to education, a source of exogenous variation that can plausibly be viewed as randomly moving around wages should be identified (Bilgel, 2020). This source of exogenous variation helps model identification. It should be (strongly) correlated with wages (i.e., relevant); should have a direct impact on wages; and should not have a direct impact on education (i.e., excluded) or the latent errors, ε, of the model (i.e., clean). We use the time (hours) spent at work over and above regular hours (overtime) and overtime payments (in Turkish lira) as excluded instruments. First, there is no evident relation that suggests that overtime, either in hours or in monetary terms, would directly affect one's education level. Second, individuals who work overtime should otherwise earn higher wages, and that overtime should affect human capital only through its effect on wages, although we cannot formally test this. 5

FINDINGS
In order to test the causal claims of Redding and Schott (2003) and to understand whether wages wipe out the impact of market access, we include both wages and market accessibility in the same framework. This calls for a substantial effort to deal with the endogeneity of wages (reverse causation and possible omitted factors that can influence human capital distribution).
We report the instrumental variables (IV) estimates for the education equation where the ordinal nature of education level can be safely ignored. Our sole aim is to assess the relevancy, cleanliness, excludability and endogeneity of the instrument. Table 1 reports the results where wages are instrumented by the overtime hours and overtime payments. The diagnostic test on endogeneity, instrument relevance, weak identification and instrument validity are reported at the bottom of Table 1. The instrument relevance and endogeneity of wages are assessed using the heteroscedasticity consistent version of the Anderson canonical correlation LM statistic (Kleibergen-Paap LM statistic) and the endogeneity test along with the weak identification-robust inference test results (Moreira, 2003).
Column (1) of Table 1 reports the first-stage results obtained from a regression of wages on market access, employment characteristics, individual characteristics, geographical location of the establishments classified by the NUTS-1 level and excluded instruments. The Anderson-Rubin weak instrument-robust inference test result given at the bottom of Table 1 suggests that the excluded instruments are strongly correlated with wages. 6 The endogeneity test result confirms our expectation that wages are highly endogenous to education levels.
Having multiple instruments allows us to test for overidentifying restrictions. The instrument validity test, assessed by the Hansen J-statistic reported at the bottom of Table 1, indicates that the instruments are uncorrelated with the unobservable factors of education and that they are correctly excluded from the education equation. The IV diagnostics provide evidence that the excluded instruments are valid and can be used to isolate the causal effect of wages on education.
After assessing the endogeneity of wages and the conditions that the IV need to satisfy, we construct our empirical specification via a recursive bivariate-ordered probit model whose results are reported by the wage and education equations, respectively in columns (1.1) and (1.2) of Table 2. Both equations host the same control variables except that the wage equation additionally includes two excluded instruments (overtime hours and payments) to help model identification. From column (1.1) of Table 2, overtime payments increase wages, but overtime hours exert a statistically significant and negative impact on wages. 7 The size and the statistical significance of the error correlation reported at the bottom of Table 2 shows that the unobservables of the education equation and the unobservables of the wage equation are negatively correlated, confirming that wage is endogenous to education levels.
While the NEG framework asserts that MA affects human capital development, we have enough reasons to believe that this mechanism works through returns to education (i.e., wages might influence individuals' educational investment). Therefore, it is possible that the influence of MA on education emanates from the signal that returns to education sends to individuals. A measure of MA is included in both education and wage equations in order to isolate any potential confounding effect of MA from the effects of wages on education levels. Our results show that a 1% increase in MA increases wages by about 0.09%; however, it does not exert a statistically distinguishable effect on education levels while wages continue to influence the decision to invest in education.
Column (2) of Table 2 shows the consequences of omitting wages (therefore ignoring cross-equation error correlation) from the model and reports the effect of MA on human capital accumulation employing a conventional ordered probit. Our a priori expectation is that the omission of wages from the model, if it truly affects education levels, should return a biased and otherwise positive parameter estimate of the impact of MA. The results are in line with expectations; MA exerts a statistically significant impact on human capital accumulation, only upon the omission of wages from the model. The exercise given in column (2) of Table 2 signifies the cruciality of the inclusion of wages as a predictor of human capital accumulation. Remarkably, columns (1.1) and (2) yield interesting empirical comparisons. Results of the wage equation (column 1.1) for Turkey are similar with those obtained by Redding and Venables (2004a) and Boulhol and De Serres (2010). Additionally, results reported from the education equation (column 2) are again similar to the empirical   Note: N ¼ 164,023. The outcome variable is the level of education (ordered). The natural log of average monthly wage is instrumented by overtime hours and overtime payments. The under-identification test reports the Kleibergen-Paap rk Lagrange multiplier (LM) statistic and p-value for the null hypothesis of under-identification. Weak identification test reports the Cragg-Donald-Wald F-statistic and p-value for the null hypothesis of weak identification. The Stock and Yogo (2005) weak identification test critical values for 10% and 15% maximal instrumental variables (IV) size are 19.93 and 11.59, respectively. The conditional likelihood ratio of Moreira (2003) reports the weak identification-robust inference likelihood ratio and p-value for the null hypothesis that the coefficient on wage is zero. The Hansen J-statistic reports the chi-square and p-value for the joint null hypothesis that the instruments are uncorrelated with the error term and correctly excluded from the equation. The redundancy test reports the chi-square and the p-value for the null hypothesis of instrument redundancy. The endogeneity test reports the chi-square and p-value for the null hypothesis that wage is exogenous. Standard errors shown in parentheses are clustered at the NUTS-1 level and are robust to arbitrary heteroscedasticity. For the fact that the number of clusters is less than the combined number of exogenous regressors and excluded instruments, the covariance matrix of orthogonality conditions is not of full rank. Therefore, age, tenure, NUTS-1 regional dummies (12 regions) and dummy variables on ISCO-08 occupations and on statistical classification of economic activities (NACE) are partialled-out for the covariance matrix to have full rank. All specifications use sampling weights provided by ESS-Turkstat (2014) and include 9 − 1 ¼ 8 dummy variables on ISCO-08 occupations and 17 − 1 ¼ 16 dummy variables on NACE. *, ** and ***Statistical significance at the 10%, 5% and 1% levels, respectively.
evidence from developed countries (López-Rodríguez et al., 2007;Redding & Schott, 2003). However, once the education equation in column (1.2) is considered, where MA, wages and human capital accumulation are used together, our results turn out to be different from the vast literature. We believe this finding is crucial as it is the first attempt to use these three pillars within the same construct. 8 Overall, our empirical analyses indicate that NEG's expectation on the impact of MA is statistically insignificant once returns to education are considered. However, wages continue to be an important determinant of educational human capital accumulation.
Returning to the original set-up of Redding and Schott (2003), our results point out that, unlike expectations, a failure to account for returns to education has consequences for the individual's educational investment decision. Without using wages, market access has a huge impact on human capital dispersion. However, the inclusion of wages shows that the impact of market access is stronger on wages, but not on the educational human capital accumulation.

DISCUSSION
The theoretical discussions on how human capital is geographically dispersed have received interest in the urban economics and the NEG literatures. In this paper, we investigate the contributions of Redding and Schott (2003) to understand the endogenous accumulation of human capital. A strong channel between geographical

1260
Burhan Can Karahasan and Fırat Bilgel proximity and firm profitability is defined within the model; however, more effort is required to understand individuals' incentives to invest in education. Given this concern, we focus on the interaction among geographical proximity (market access), human capital accumulation and wages. We use micro-level data to control for these three pillars within the same empirical set-up. This enables us to critically question the causal channels of Redding and Schott's model. Using individual-level data results in a loss of locality due to the representation level of the data. However, it would not be possible to focus on the interaction among market access, human capital accumulation and returns to education unless individuallevel data are used. We find that MA is statistically positively associated with wage distribution after controlling for a host of other factors. This finding is consistent with the strong co-movement between MA and wages for the pre-2000s at NUTS-3 disaggregation (Karahasan et al., 2016). Therefore, the impact of MA on wages at a broader regional disaggregation continues after the 2000s in Turkey even at the individual level. Additionally, these results are on a par with prior evidence testing the influence of proximity on regional inequalities among other developed countries (Boulhol & De Serres, 2010;Head & Mayer, 2011). However, unlike prior findings for Europe (Diebolt & Hippe, 2018;Faíña & Lopez-Rodriguez, 2006;Karahasan & Bilgel, 2020;López-Rodríguez et al., 2007;López-Rodríguez & Manso-Fernandez, 2017), MA is not effective in explaining the variation in educational human capital development in Turkey. On the contrary, our findings are similar to those obtained for Spain, where structural differences in production and spatial spillovers across regions are used as two potential explanations for the weak influence of MA on human capital accumulation (Karahasan & López-Bazo, 2013).
In Redding and Schott's (2003) model, a higher MA is associated with higher education level. This explains why central locations are more attractive and dominated by highly educated people. MA is historically persistent in developing countries and usually policy insensitive. Therefore, the emphasis of the NEG model on MA brings additional concerns on the effectiveness of local policies to combat human capital-based social inequalities. However, our findings highlight that the impact of MA on educational human capital accumulation is not so strong, but rather sensitive to other factors that are supposed to affect human capital development. Once the regional distribution of wages is considered, the impact of MA disappears. In return, our study provides significant empirical evidence that returns to education matters for individuals' decision on educational investment at different levels. Therefore, regional policies are still in effect. Unlike the remarks on the policy-insensitivity of MA, individual factors that are sensitive to social policy as well as returns to education that are sensitive to market-based policy are crucial aspects of educational human capital development. While policy implications have limited scope to influence MA, some policy tools can still influence the level of economic activity (i.e., local innovation and taxation policies) or the bilateral distance between locations (i.e., infrastructure improvement). These measures can be taken into account to consider the historical rigidity of MA. This can potentially cancel out the policy neutrality. Policy sensitivity of MA deserves a separate discussion on the nature of local public policy beyond the scope of our paper. Detailing the background of these policies is an essential part of our future research agenda. 9 There are other alternative explanations as to why certain regions are clustered with more educated individuals. The urban economics literature points out that urbanized areas demand relatively more skilled individuals and city size affects urban agglomeration, both of which attract more educated individuals (Behrens et al., 2014;Glaeser & Saiz, 2004). 10 This can influence individual mobility across regions (Crozet, 2004). Moreover, public provision of education can be heterogeneous as well. Redding and Schott (2003) consider the institutional set-up of education as homogenous within the territories of the same country. However, possible institutional divergence can also be a source of mobility as individuals may prefer to move to locations with better institutional set-up for education. While we acknowledge these potential channels, we opted to focus on the causal framework of Redding and Schott's (2003) model. With the availability of more detailed individual-level data (i.e., migration, different labour market indicators, quality of education and institutional capacity of the education system at a local level) different empirical set-ups can be constructed to understand which individual characteristics motivate the accumulation of human capital at the regional scale. Our results so far confirm that there are potential determinants of the accumulation of human capital development other than MA.

CONCLUSIONS
We investigated the impact of MA on human capital development differences among the regions of Turkey. We moved a step further by using wages, human capital accumulation and MA within Redding and Schott's (2003) model. To our knowledge, an attempt to incorporate these three pillars with the help of individual-level data is missing in the literature. Our identification strategy was predicated on the question of whether people increase their human capital levels by investing more in education because they live in central locations or returns to education motivate individuals' incentives to invest in human capital. As the former case is more policy insensitive, our study underlined the need for this decomposition.
While our preliminary analyses show similarities in terms of regional human capital patterns and market accessibility, our empirical results show that the use of wages and MA in the same set-up changes the expectations of the augmented NEG framework. The link between wages and human capital wipes out the impact of MA on human capital accumulation. Strikingly, we identified a positive link between MA and wages; however, we were unable to report any significant connection between MA and education. Therefore, even if wages have had an influence on education accumulation, this would not necessarily imply that the background is the market accessibility. Given the large rigidity and policy neutrality of MA, our results confirm that there are still certain policy tools in order to close the regional gaps in human capital accumulation in Turkey.
These findings broach new discussions about the true cause behind the dispersion of educated and skilled individuals. Our evidence confirms that it is not solely proximity that motivates educational human capital development. Besides, geographical proximity and MA are vital parts of wage distribution. Skill sorting and location decision of individuals at high-wage-offering locations indicate that the decision to invest in education is a combination of income opportunities and geographical advantages. While the former is relatively more policy-sensitive, the rigidity of geographical advantages can also be improved by implementing policies that enable easier access to markets and supply sources. Detailing down the right policy mix should be on the agenda of researchers.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.

NOTES
1. Since our central focus is on the human capital accumulation and supply side of the model, we skip the presentation of the agriculture sector. For the full solutions and the overall equilibrium, see Redding and Schott (2003, section 2). 2. The ESS is administered at the regional level only for 2014. There are alternative sources for individual wage data in Turkey; however, they fail to provide the required individual information for our identification strategy. Therefore, we only estimate our model for 2014. Prior evidence confirms the historical persistence of educational human capital disparities and market accessibility (Erdem, 2016;Karahasan et al., 2016). Therefore, we have enough reasons to believe that our results can be generalized for the Turkish case, shaped by persistent regional disparities. 3. See Section A.1 in Appendix A in the supplemental data online for descriptive statistics. 4. We estimate two variants of the education equation. The first (column (1.2) of Table 2) is an augmented version of Redding and Schott (2003) and uses wages and market access. The second (column (2) of Table 2) is the direct estimation of the Redding and Schott set-up that uses market access only to explain educational human capital distribution. 5. Another threat to identification is the possible endogeneity of MA (Boulhol & De Serres, 2010). As our central concern is the identification of wage, human capital accumulation and market access triangle, we waived further investigation of endogeneity of MA. However, the results that account for the endogeneity of MA by following the routine approach offered by Boulhol and De Serres (2010) are available from the authors upon request. 6. A suggested measure to assess the explanatory power of the excluded instruments is the first-stage F-statistic > 10 ( Bound et al., 1995;Staiger & Stock, 1997). A recent study shows that if an F > 10 threshold is used, the required critical value for a t-test at the 5% significance level to have a correct size/coverage is 3.43 instead of the usual 1.96 (Lee et al., 2020). Given the critical value of 3.43, the confidence interval for the coefficient on the natural log of monthly wage in the second stage of Table  1 becomes [0.011:1.747]. 7. We did not take the natural logarithm of overtime payments and hours because 80% of the sample have an overtime value of zero. 8. We compute the average marginal effect (AME) for wages at a regional level to assess the possible geographical variability of its impact on human capital accumulation. The results show sizable geographical variability, reminding one of the importance of returns to education for policy construct to combat regional disparities. See Section A.2 in