Heterogeneity among migrants, education–occupation mismatch and returns to education

ABSTRACT Using nationally representative data for India, this paper examines the incidence of education–occupation mismatch (EOM) and returns to education and EOM for internal migrants while considering the heterogeneity among them. In particular, it considers heterogeneity arising because of the reason to migrate, demographic characteristics, spatial factors, migration experience and type of migration. The analysis reveals that there is variation in the incidence and returns to EOM depending on the reason to migrate, demographic characteristics and spatial factors. The study highlights the need of focusing on EOM to increase the productivity benefits of migration. It also provides the framework for minimizing migrants’ likelihood of being mismatched while maximizing their returns to education.


INTRODUCTION
Migrants comprise those who experience either a permanent or a semi-permanent change of residence (Lee, 1966). Several models predict and explain migration as a result of investment in human capital. By now, economists and demographers have a clear-cut answer to why people migrate. Migration prompts numerous benefits. It reduces poverty, increases income, enhances upward mobility, escalates savings and assets formation, fosters investment in human capital, reduces vulnerability, improves food security, stimulates land markets at origin, increases local wages and demand for local goods and services, improves the economy, and tightens rural labour markets (Bird & Deshingkar, 2013). Due to the costly nature of international migration, internal migration has a considerably large role in the process of upward mobility, development and economic growth. Internal migrants are defined as those who relocate within their home country, but outside their usual place of residence for a considerable duration. 1 Internal migration is an important mechanism by which labour resources are redistributed nationally on account of changing demographic and economic forces (Greenwood, 1997). Various researchers have studied the patterns of internal migration and found that individuals choose to migrate toward regions that pay higher income (Borjas et al., 1992) and have low unemployment rate (Herzog et al., 1993). Therefore, individuals by being spatially flexible attempt to maximize their earnings and lower their chances of being unemployed. However, the extent to which migrants are able to efficiently match their education with their occupation and the impact of mismatch on their income is still a debatable question. This issue is interesting to examine because in a given occupation, productivity (Tsang & Levin, 1985) and job satisfaction (McGuinness & Sloane, 2011) are higher in a case where the education required by an occupation matches the education attained by a worker as compared with a situation of mismatch. Further, migration has been considered as an important channel for achieving the highest level of productivity (Maxwell, 1988). Therefore, the exploration of the interlinkages between migration and education-occupation mismatch (EOM) 2 addresses the broader question of workers' ability to better use their human capital endowments by being spatially flexible. It also highlights the present state of the use of human capital endowments of internal migrants with respect to recipient regions. Moreover, since EOM has considerable consequences on income (Duncan & Hoffman, 1981;Verdugo & Verdugo, 1989), it can serve as an imperative indicator of migrants' success or failure at the destination. This paper, in general, contributes to this literature by analysing the extent of and returns to EOM for migrants.
The last decade has witnessed growing contributions from researchers toward analysing the relationship between migration and EOM. The current research on this issue has touched on two themes. The first segment explores the impact of migration on the likelihood of mismatch between education and occupation (henceforth, mismatched). This theme has further evolved in two strands. The first deals with international migration. Researchers have found that international migration leads to a higher likelihood of being mismatched (Aleksynska & Tritah, 2013;Dahlstedt, 2011;Nielsen, 2011;Wald & Fang, 2008) and bases this finding on the imperfect transferability of human capital (Nieto et al., 2015). The second strand considers internal migration. Researchers have hypothesized that internal migration should lead to a lower likelihood of being mismatched. This is because spatial flexibility leads to better access to jobs and consequently widens the adequate job opportunities available to a worker, which is grounded on Frank's (1978) theory of differential overqualification. The central aspect of this theory is that the incidence of EOM would be higher for workers with relative spatial inflexibility (Büchel & Van Ham, 2003). Besides, the problem of human capital transferability is of limited concern for internal migration due to similar culture, language, infrastructure, etc. Therefore, EOM may arise due to spatially constrained job search and could be reduced by being spatially flexible within a home country. 3 While many studies have supported this argument by observing a higher likelihood of mismatch among natives than migrants, others have argued that there exists no significant relationship between migration and probability of being mismatched (Hensen et al., 2009;Iammarino & Marinelli, 2015;Jauhiainen, 2011).
The second segment analyses the impact of migration on the returns to EOM. In the case of international migration, the literature has observed that migrants lose much more from not being correctly matched than do natives (Joona et al., 2014;Nielsen, 2011). While the literature on the first segment is quite rich, the literature on the second segment is scant in the case of international migrants and almost non-existent for internal migrants. Further, apart from the international and internal migration, the existing literature has also examined the differential impact of the likelihood of being mismatched by source region (Joona et al., 2014), destination region (Iammarino & Marinelli, 2015), source-destination pairs (Aleksynska & Tritah, 2013) and education level (Croce & Ghignoni, 2015) of the migrants. Therefore, barring a few studies, others have considered migrants as a homogeneous group which could be misleading. This is because migrants can have different socio-economic, demographic and other personal characteristics, which in turn can have a differential impact on their labour market outcomes.
In order to address this concern, this study expands the theoretical model developed by Simpson (1992) and adapted by Büchel and Van Ham (2003). The model relates a person's match status to the range of job opportunities. A person looking for a job in a particular local labour market has three options in the scenario of not finding an adequate job: first, not to accept any job (unemployment); second, to settle for a mismatched job; and third, to migrate or commute to another labour market for better prospects. Based on the model, Büchel and Van Ham (2003) observed that individuals who can widen the size of their labour market have a better chance of getting an adequately matched job. This paper argues that a series of decisions do not end here. Once an individual decides to migrate, there are other decisions that an individual has to take regarding location, type of migration, etc. These decisions again depend on various factors such as availability of job opportunities at home location, assimilation, infrastructure facilities, etc. Since the motivation to migrate varies for different choices, it raises an interesting question regarding the differences in the labour market outcomes, in general, and EOM, in particular, with respect to different kinds of migration.
This paper aims to answer this question, and therefore estimates the differential returns to EOM while distinguishing the migrants by (1) reason to migrate, particularly in search for a job or to take up a confirmed employment or for other work-related reasons; (2) demographic characteristics, namely gender and marital status; (3) spatial factors, including source-destination pairs and distance travelled; and (4) migration experience as per years since migration. Our central hypothesis is that the internal migrants by being spatially flexible attempt to maximize the returns to their human capital endowments (particularly education) and minimize the penalty from being mismatched. But whether all kinds of migrants get similar results is primarily an empirical question and is yet to be explored. To the best of our knowledge, no one has investigated this aspect. To explore this dimension, the study uses data from India's National Sample Survey Office (NSSO) which were collected in the period 2007-08 and primarily concerned with employment, unemployment and migration particulars.
The contribution of this paper is twofold. First, it contributes to the understanding of heterogeneity among migrants and the consequent differential impact of EOM in case of a developing country. Prior literature has argued that developing countries do not bias the migration patterns by providing unemployment benefits and other fiscal programmes; and hence serve as a better case to study the issues related to migration (Stark & Bloom, 1985). India with its marked interregional disparities provides a compelling case study for this analysis. Further, in India it has been found that the migration rate generally shows an increasing trend with the increase in education level, which indicates that higher educated workers are more likely to migrate (NSSO, 2010). However, the extent to which migrants are able to use their human capital, particularly education, is not well understood. Second, by analysing the relationship between heterogeneity among migrants and the extent of and returns to EOM, we highlight the role of geographical limitations in affecting the opportunities to optimally gain the returns to attained education. 4 Only a few studies in India capture the incidence of EOM. One such study by Bahl and Sharma (2021) analyses the extent of EOM in the Indian labour market and further analyses the role of EOM in understanding intra-education wage inequality (dispersion) using a quantile regression approach. The authors do not focus on the issue of spatial flexibility in the regional context and do not have migrants as their interest group. The present study differs from theirs along multiple dimensions. We focus on whether internal migration affects the incidence and returns to EOM in the context of India. The prime motivation behind this analysis is to test whether spatial flexibility leads to improvement in the returns to education in the wake of EOM. Further, as a corollary, we also analyse whether migration reduces the penalty for surplus education and improves the returns to required and attained education. Lastly, as discussed above, we estimate the heterogeneity in returns to education and depending on the various characteristics of internal migrants.
Our study is an original contribution to the literature on EOM and spatial flexibility in the labour market. We contribute by highlighting the role of regional mobility and access to larger labour markets in reducing the mismatch and labour misallocation in the market while improving the pecuniary returns to education. Further, we emphasize that these returns are not uniformly accessible to all the migrants but are also affected by varying characteristics of the migrants themselves. This dimension needs to be understood better by the academicians and the policymakers for evidence-based targeted policymaking for the workers, especially migrants.
The rest of the paper is structured as follows. Section 2 describes the data source, measurement and empirical model. Section 3 explains the results. The final section concludes.

DATA AND METHODS
This section describes the data source, gives some background on internal migration and EOM in India, followed by measurement of EOM, and an empirical methodology.

Data source
In the case of India, the information regarding migrants and their labour market outcomes are available as part of the surveys conducted by the National Sample Survey Office (NSSO). The latest data collection period for this nationally representative survey is 2007-08 (64th round). The survey focuses on migration and its duration was from July 2007 to June 2008. Apart from this survey, till now no other dataset has provided the micro-data for the individual level analysis of the migrants and their comparison with non-migrants. The other labour force surveys as well as the Census of India either do not capture the migration information or only provide the aggregated information at the district or state level.
The survey is conducted across 35 states and union territories of India and covers 125,578 households (79,091 in rural areas and 46,487 in urban areas), enumerating a total of 572,254 persons. The Census of India 2001 has been used as a sampling frame for the survey. The NSSO survey contains detailed information on household characteristics such as religion, social group, land possessed, etc., along with the demographic characteristics such as education, age, gender, marital status, etc. The survey also provides information on the occupation and industry category of the workers.
Migration information is collected at two levels: household and individual. Considering the objective of this paper, the unit of analysis is at the individual level. The NSSO treats a member of the sample household as a migrant: if he or she had stayed continuously for at least 6 months or more in a place (village/town) other than the village/town where he/she was enumerated. The village/town where the person had stayed continuously for at least 6 months or more prior to moving to the place of enumeration (village/ town) was referred to as the 'last usual place of residence' of that migrated person. Shifting of residence within village or/town was not considered as an event of migration. (NSSO, 2001, p. 14) In other words, an individual is classified as a migrant if his or her last usual place of residence is different from the place of enumeration. The usual place of residence is one where an individual has stayed continuously for at least six months. For our analysis, the sample is restricted to the working-age population (15-59 years), which is consistent with the economically active age group as considered by the Government of India (NSSO, 2010). Further, we only consider workers employed for at least six months in the year preceding the survey, which has been measured by the principal activity status of the workers. Thus, we only consider the main workers (which is analogous to fulltime workers in an international context).
The main results consider the wage/salary of internal migrants who moved for work-related reasons, 5 unless otherwise stated. The daily wages are reported for each of the seven days preceding the survey, and we have calculated the average daily wage for the workers by summing their daily wages and dividing it by the number of days worked in the last seven days. Given that the sample is only for the period of one year, we have measured the wages in nominal terms and have not used any consumer price index deflator. Wages are measured in Indian rupees, which is a standard practice in the context of research on the Indian labour market. Additionally, while using the wages for estimation, we also drop extreme observations by truncating 0.5% at either end of the wage distributions. This ensures that our findings are not affected by the outliers on either side. We have not included self-employed workers in our analysis due to the non-availability of their earnings. We also provide estimates for nonmigrants in the working-age group who receive wage/salary. Thus, our final sample size comprises 15,434 migrants and 60,689 non-migrants. Table A0 in Appendix A in the supplemental data online contains the details about the sample considered in this study.
In addition to the NSSO survey, we also use the All-India Survey of Higher Education 2012 (AISHE) and Socioeconomic High-resolution Rural-Urban Geographic (SHRUG) database. These two sources provide us with IV for our empirical model. The variables used from the AISHE are the number of colleges/universities and autonomous institutions within a district. Given that our sample survey is for the period 2007-08, we restrict the count of institutions till 2001, that is, their year of establishment should be 2001 or before. The year 2001 is selected to get enough lag. The SHRUG database has provided the following IV: (1) calibrated night-time light in 2001 at the district level; and (2) the SD of night-time light across settlements within a district. For a detailed discussion on these variables, see subsection 2.4. on the empirical model.

Internal migration and EOM in India
In this subsection we provide background about the phenomenon of internal migration and EOM in India. For brevity, we keep the discussion focused.

Internal migration in India
As per the estimates from the 2007-08 survey, around 26% (194 million) of the rural and 35% (94 million) of the urban populations identify themselves as internal migrants in India. Around 55% of this migration is on account of marriage. Coming to the streams of migration, ruralrural is around 62% of total migration, followed by rural-urban (19%), urban-urban (13%) and urban-rural (6%) (Chandrasekhar & Sharma, 2015). In India, internal migration is largely an intra-state phenomenon. Intra-district migration is around 59% of total migration and interdistrict (within state) has 29% share. Inter-state migration is only 12% of total migration.
In 2007-08, a total of 28 million migrants migrated for work related reasons. Of these, around 91% were in the working-age group 15-59 years, and around three-fourths considered urban areas as their destination.
Given these numbers, it can be said that internal migration is very large in India and very heterogeneous in terms of destination choices and reason to migrate. For a more detailed discussion on internal migration, see Chandrasekhar and Sharma (2015).

EOM in India
In the context of developing countries such as India, the misallocation of labour and subsequent mismatch between education of workers and occupational requirements remains a big concern. One extreme example of such instances is the application of around 3700 doctorates, 28,000 postgraduates and 50,000 undergraduates for the posts of 62 messengers in the Uttar Pradesh (a state of India) police with a minimum education requirement of class V (primary education -five years). 6 Such instances have become a norm instead of being an anomaly. Based on the study conducted by Bahl and Sharma (2021) using data from the NSSO's survey on employment and unemployment 2011-12, it has been estimated that around 19% (21% among men and 13% among women) of wage and salaried workers accounting for around 35 million were over-educated for the jobs. Additionally, the incidence of under-education for the same period was around 15% (26 million) of wage and salaried workers. These numbers are lower than many of the developed countries but remain quite high for the developing countries. The primary reason is the expansion in tertiary education along with the absence of job creation in the Indian economy.
In the next subsection we describe the measurement of EOM in our survey data.

EOM: measurement
This paper defines EOM as a gap between the attained years of education of an individual and the required years of education by her occupation. If attained education is greater than required education, a person is regarded as over-educated. If attained education is lower than the required education, the person is under-educated. A person is adequately educated when her attained and required education are aligned. Thus, to identify a worker's match status, we need two dimensions of crucial information, that is, attained years of education and required years of education with respect to the occupation. NSSO collects information on the level of general education (no formal schooling, below primary, primary, middle, secondary, higher secondary, graduate, and postgraduate and above). We convert the level of education into years of formal education for our analysis.
For the measurement of required education, the literature provides three methods: workers' self-assessment (WA), job analysis (JA) and realized matches (RM) (see Leuven & Oosterbeek, 2011, for a complete discussion). WA captures the aspect of mismatch from workers' perspective, JA focuses on employers' side and RM considers the perspective of a labour market. Thus, RM identifies the workers' match status by considering demand and supply-side factors of a particular job. Given the focus of this paper, RM serves as the appropriate method due to the following reasons. First, the focus is on the nationwide labour market outcomes of the migrants that can be best discussed using the measure that considers the entire labour market while estimating the required education. Second, in the case of developing countries, there is a significant dearth of surveys that capture the information required by WA and JA to measure the required education. Hence, RM remains the only applicable method. An added advantage is that the RM method is more common in the literature of spatial mobility and EOM (Nielsen, 2011;Poot & Stillman, 2016) and thus would give us leverage to compare our results with the existing studies.
RM defines required education of an occupation as a range, where the lower (upper) limit is calculated as the mean education attained by the workers minus (plus) 1 SD (standard deviation) in the education of workers working in a given occupation (Verdugo & Verdugo, 1989). Following this, the mean years of education for every three-digit National Classification of Occupation (NCO) 2004 codes have been computed using the sampling weights and a threshold of ± 1 SD from mean has been established to measure the range of required years of education. In this context, a worker will be categorized as adequately educated if his or her attained education falls in the range of ±1 SD from the mean, and a worker will be classified as over-educated (under-educated), if his or her attained education is higher (lower) than 1 SD above (below) mean. The mean and SD are measured only for the workers in the working-age group.
More precisely, suppose e i represents attained years of education of an individual i; e nat and s nat are the mean and SD of years of education, respectively, for the occupation calculated at the national level. Thus, an individual will be considered: Adequately educated if : e nat − s nat ≤ e i ≤ e nat + s nat Using this definition and data from NSSO (2007-08), 7 we find that 71% of the wage/salary employed workers in India are adequately educated. 8 Further, despite having overall lower education levels as compared with other countries (Tilak, 2018), the problem of over-education (16%) is more severe than under-education (12%) in India. While analysing EOM only for migrants, it is observed that migrants have a higher incidence of education-occupation match as compared with the overall group. About 77% of migrants receiving wage or salary are adequately educated. However, for work-related migrants, the proportion of the adequately educated drops to 70%. This highlights that individuals migrating for work-related reasons focus more on getting a job rather than an adequate one. This is indicated in the higher proportion of wage/salary employed (70%) among workrelated migrants as compared with overall migrants (22%).
2.5. Empirical model 2.5.1. Ordinary least squares (OLS) estimation This paper uses augmented Mincerian (Mincer, 1974) wage equation to estimate the returns to wage-determining characteristics. Thus, our wage equation is as follows: where the dependent variable is logarithm of daily wage of individual i in region r; Edu i is our main interest variable, indicating years of education attained by an individual.
We have four sets of controls X 1i , X 2r , D r and M i . The first set of controls are at the individual level, which includes years of formal education, age and its squared term, 9 gender (male and female), marital status (unmarried, married and others), interaction of gender and marital status, social group (scheduled tribe, scheduled caste, other backward class and others), religion (Hindu, Muslim, Christian and others), occupation categories at a broad level 10 (legislators, senior officials and managers, professionals, associate professionals, clerks, service workers and market sales workers, skilled agricultural and fishery workers, craft and related trades workers, plant and machine operators and assemblers, and elementary occupations) and industry type (17 industry types including agriculture, fishing, manufacturing, etc.). 11 The survey also provides information on employment status (employed, unemployed and out-of-labour force) of migrants at their last usual place of residence. Since being employed in the past can positively impact the wages, we include the dummy for being a wage/salaried employee at the last usual place of residence.
The second set of controls are at the regional level focusing on local labour market characteristics. This includes location sector (rural or urban), share of international immigrants 12 (measured at the state level), log of working-age population (at the district level) and log of the labour force (at the district level). These factors control for the heterogeneity in the labour market size, potential labour market competition, as well as the interrelationship between international and internal migration.
The third set of controls use dummies for the fixed effect of origin and destination location. We use statelevel dummies for origin and destination locations.
The fourth set of controls are implemented in the empirical model for migrants to control for migrant heterogeneity in terms of reason for migration (to take a confirm job, search for work and other reasons), migration distance 13 (intra-district, inter-district within state, inter-state), years since migration and migration stream (rural-rural, rural-urban, urban-rural and urban-urban).
Further, to capture the differential returns to EOM, Duncan and Hoffman (1981) suggested segregating the years of education into required years of education, surplus years of education (if a worker is over-educated), and deficit years of education (if a worker is under-educated), that is: where Edu a i and Edu r i represent attained and required years of formal education, respectively, for an individual i; Edu s i represents surplus years of formal education (which is measured as Edu a i − Edu r i ); and Edu d i represents the deficit years of formal education (which is measured as Edu r i − Edu a i ). Considering this, our wage equation becomes as follows: where Edu r i represents required years of formal education; Edu s i represents surplus years of formal education; and Edu d i represents deficit years of formal education. Other variables are interpreted as before. In equation (3), d 1 , d 2 and d 3 depict the returns to required education, surplus years of education and deficit years of education, respectively. Therefore, if a person is adequately matched, the years of surplus and deficit education would become zero. Further, if a person is over-educated, the years of attained education would be divided between required education and surplus education. Similarly, if a person is under-educated, the years of attained education would be divided between required education and deficit education. Thus, for every observation, the years of required, surplus and deficit education would be largely uncorrelated. 14 The existing studies have usually found that (1) the returns to attained education are statistically lower than the returns to required education; (2) although surplus education yield positive returns, the returns are statistically lower than that of the required education; and (3) deficit education yields negative returns and this penalty is lower than return associated with required education (Hartog, 2000).
However, to estimate the unbiased coefficients in equation (3), there is a need to tackle a methodological issue of sample selection into employment (Heckman, 1979). This may arise as the wage of a worker is observable only after getting employed. But there is a possibility that some unobserved factors (such as ability) are positively related to both wages and likelihood of being employed. Therefore, considering only the employed individuals for analysis would lead to biased estimates. Another problem in our data is the unavailability of income information for self-employed individuals. However, the choice of selfemployment versus wage/salary employment is not random and disregarding the possible sample under self-employment may again lead to sample selection bias (Dolton & Makepeace, 1990). Further, the decision to migrate is also not random. Hence, there is a problem of triple-selection bias. To overcome this problem, a Heckman correction procedure is applied for all the three decisions, that is, the decision to work (to be or not to be engaged in economic activity), the choice of economic activity status (wage/salary or self-employment), and the decision to migrate (to be or not to be migrant).
To use this method, it is a prerequisite to identify at least one variable that does not affect the wages but influences the probability of participation (in our case, employment, self-employment and migration); such a variable is called an exclusion variable(s). We have used the number of dependent members (under 15 and over 59) in a household, household type (self-employed, regular wage/salary earning, casual labour, etc.), and household size as exclusion variables for the decision to work. This is in line with the literature that advocates the use of family characteristics as appropriate exclusion variables for the choice of work (Buchinsky, 2002). Further, we have used land possessed by the household as an exclusion variable for choice of economic activity status. This is on the grounds that land ownership improves the access to credit (Feder & Onchan, 1987), as it can be used as a collateral (Kaas et al., 2016). Further, self-employment being riskier than wage/salary employment, having access to a land can be considered as a safety cushion. We have used the migrant network indicators as the exogenous predictors for the decision to migrate (Bertoli, 2010;Bertoli et al., 2013;McKenzie & Rapoport, 2010). For the discussion on measurement of migrant network indicators, see section A2 in Appendix A in the supplemental data online. In short, to address the self-selection, we have estimated the first stage sample selection for the Heckman correction using the exclusion variables and then the resulting correction terms are used in the final wage equation.
Therefore, our final wage equation is as follows: where l are the correction terms for sample selection related to employment (versus not employed), wage/salary employment (versus self-employment) and migration (versus no migration), respectively. Other variables are interpreted as above. The estimates of the sample selection probit models for all three decisions are provided in Table A2 in Appendix A in the supplemental data online.
It can be noted that the exclusion variables for all the three decisions are statistically significant. Furthermore, apart from the exclusion variables, we have also controlled for years of education, age, age squared, gender, marital status, interaction of gender and marital status, social group, religion, state and sector. 15 All these variables appear to be statistically significant in influencing the probability of migration, employment and wage/salary employment.
Equation (4) is estimated separately for a varied group of migrants. The summary statistics for the wage/salary employed non-migrants and work-related migrants in the working-age group (15-59) are presented in Table  A3 in Appendix A in the supplemental data online, for brevity. Also, the incidence and average daily wages (in Indian rupees) of work-related migrants in the working age-group by different demographics are given in Table  A4 online.

Instrumental variables (IV) estimation
In our OLS model, after accounting for the sample selection bias regarding decision to employment, nature of employment (wage/salary versus self-employment) and migration, the concern for endogeneity of the main variable, that is, years of education persists. In the empirical literature on returns to education, IV estimation is the best way to address this concern. To correct the endogeneity of general education measured as the attained years of education, we require exogenous variables as instruments in our model. To prove their validity and strength, these instruments should meet the following conditions: (1) they should be uncorrelated to the error term, also called the orthogonality condition; (2) they must be correlated to the endogenous variable that needs to be instrumented; (3) they should not be part of the original model, also referred to as the exclusion criteria; and (4) the number of instruments should be at least equal or more than the number of endogenous variables. The finding of instruments in the observational data that simultaneously satisfy the above-mentioned four conditions remains a major problem in implementing IV models for identification.
For our IV estimation, we have identified the following IV based on the literature: (1) geographical variation in the number of colleges/educational institutions at the district level; (2) monthly per capita expenditure as a proxy for household status; (3) calibrated night-time light in 2001 at the district level; and (4) the SD of night-time light across the settlements within a district.
However, one additional statistical concern for our IV model is weak instrument condition. To overcome this, we have taken an approach suggested by Lewbel (2012). Hereafter, we will refer to the IV model as the Lewbel IV model.
Further, we have faced one additional problem while estimating the Lewbel IV model for the migrants. For non-migrants, the district of schooling and work remains the same, and therefore the number of colleges/educational institutions and night-time light can be used as the valid IV. However, the same is not true for all the migrants, especially inter-district and inter-state migrants. Also, the survey does not provide an identifier of their origin district. 16 Thus, using the above IV for inter-district and inter-state migrants will cause measurement error in the IV, leading to biased and inconsistent Lewbel IV estimates. Given this caveat, the IV used in our model remain only valid for the non-migrants and the intra-district migrants. Hence, in our results section we have only reported the results for these two groups.
The detailed equations for Lewbel IV models and corresponding discussion are provided in section A3 in Appendix A in the supplemental data online.
In the next section we present the results of sample selection models, followed by the OLS models, and finally Lewbel IV estimates along with the diagnostic tests for the IV models.

RESULTS
In this section we present the wage returns to education and EOM for non-migrants and work-related migrants segregated by their demographic characteristics, spatial factors and migration experience taking account of other wage-determining characteristics. For brevity, the estimates for only concerned variables are shown. 17 Before going ahead with the main results, the sample selection models (see Table A2 in Appendix A in the supplemental data online) are briefly discussed. While selecting the model for migration, we have found that the migrant network has an inverted-'U'-shaped relationship with the decision to migrate (Yamauchi & Tanabe, 2008). While analysing the employment selection model, we have found that individuals with a higher number of dependents in the household have a higher probability of entering the labour market. Further, individuals with higher land endowments are more likely to engage in wage and salary employment (see Table A2, column 3, online). This can be explained by the livelihood diversification strategy (Reardon, 1997). We have included the inverse Mills ratio from these three sample selection models in the later stage of econometric analysis and found that the Mills ratio for migration and employment are significant with negative and positive coefficients, respectively. This indicates that there is no significant sample selection bias depending on the type of employment. However, migration has negative sample selection bias and employment decision has positive sample selection bias, which has been corrected through this Heckman procedure.

OLS estimates
Having discussed these sample selection models, the main results are reviewed. First, we present the results for workrelated migrants and non-migrants (Table 1). On average, for every one year of formal education, the work-related migrants earn a wage reward of 3%, while for nonmigrants, the return is around 2%. Further, dividing the years of education into required, surplus and deficit years of education, it is observed that the results for both migrants and non-migrants are aligned with the literature (Hartog, 2000;Leuven & Oosterbeek, 2011). The returns to required education are higher than the returns to attained education. Also, while the returns to surplus education are positive, they are lower than the returns to required education. The returns to deficit education are negative, which highlight that the under-educated workers earn lower wages as compared with adequately and overeducated workers in a given occupation. In the subsequent analysis, the migrants have been segregated into various groups to find answer to the primary question of this paper.
First, we differentiate the migrants according to the reason to migrate (Table 2), which are (1) in search of a job (Table 2, column 2); (2) to take-up a confirmed job (Table 2, column 3); and (3) other job-related reasons (Table 2, column 3) which includes business, transfer of job and proximity to workplace. 18 Migrants who did not have a confirmed job before migration earn lower returns to education as compared with migrants who had one. Migrants in search of a job have lower bargaining power in the labour market (Blanchard, 1991) due to their spatial inflexibility and urgency to get a job for covering their living expenses at the destination. Further, migrants with a confirmed job get higher returns with respect to required and surplus education and higher penalty for deficit education. Migrants for other reasons (business, transfer, proximity to work) have higher returns than the other two categories for attained education, but get similar returns for required and surplus years of education. While the difference in the returns to deficit education is significant, all the groups of migrants earn statistically similar returns to their required and surplus education.
For spatial factors, the migrants by source-destination pairs, that is, rural-rural, rural-urban, urban-rural and urban-urban, are separated (Table 3). The rural and urban areas vary in terms of job opportunities, infrastructure, etc. This necessitates the consideration of source-destination pairs while analysing the returns to education. We have found that irrespective of the destination, workers from urban areas earn higher returns to education. This could be due to better quality of education in the urban areas as compared with the rural areas (Agrawal, 2014). Also, workers in urban areas are better informed regarding the availability of job opportunities, and hence can choose the more appropriate alternative. Besides, workers who move from urban areas to urban areas witness the highest penalty for deficit years of education. We have also found differences in the returns to be significant across all the groups. Further, workers moving to rural areas experience a more prominent difference between their returns for required and surplus education as compared with other groups. One plausible explanation is that jobs in rural areas are not as skill-intensive as jobs in urban areas, and consequently wages are determined more by the job characteristics rather than the education of workers. Thus, having surplus education does not add many returns in terms of wages.
Further, to estimate the differences in the returns as per the distance travelled in migration, we have divided the migrants into three groups (Table 4): intra-district (column 2), inter-district but within state (column 3) and inter-state (column 4). 19 It is found that workers who move inter-district but within the state earn the highest return on education. This phenomenon can be  explained using the contrasting results found in case of the international and internal migrants (Devillanova, 2013). Workers who choose to migrate only within their district have lower intensity of spatial flexibility, and thus get lower returns to education as compared with workers who migrate inter-district. But as distance increases, the negative impact of imperfect portability of human capital endowments begins to dominate the positive impact of spatial flexibility on wages. Workers moving to a different state may find it difficult to transfer their human capital due to variation in the quality of education, lack of language proficiency, cultural differences, etc. (Krishna, 2004); and therefore earn lower returns. Further, since education of the inter-state migrants is not directly transferable, they also earn a lower penalty for deficit years of education. However, we have found that differences in the returns to surplus and deficit education are not significant among these groups. Therefore, it can be concluded that workers who are spatially flexible and have better portability of their human capital endowments earn the highest rewards for the required education. Next, conditional on age, we have found that the longer a migrant has been at a particular destination, the higher are the returns to education (Table 5). This is in  line with the literature that claims that migrants' wages improve with the time spent at destination (Borjas et al., 1992;Yamauchi, 2004) on account of the assimilation process. The researchers have also found that the difference in the likelihood of being mismatched between migrants and non-migrants decreases with the increase in the stay of migrants (Aleksynska & Tritah, 2013;Nieto et al., 2015). We have also observed that the returns to required and surplus education for the group are higher with respect to the highest migration experience. However, the differences in the returns are insignificant across groups. Therefore, to summarize, while migrants get differential returns to attained, required, surplus, and deficit education on account of their respective reason to migrate and spatial factors, the migration experience does not lead to differential returns. 20 3.2 Lewbel IV estimates 3.2.1. Test for endogeneity Before going ahead with the two-stage least squares (2SLS) IV model, the endogeneity of the variableattained years of education for migrants and non-migrants was tested using the Durbin-Wu-Hausman test of endogeneity (Cameron & Trivedi, 2010). For nonmigrants, we have observed that the attained years of education is endogenous, as the null hypothesis that the variable is exogenous is rejected. The Durbin-Wu-Hausman test of endogeneity F-statistic (p-value in parentheses) values are 8.26 (0.004). For migrants, the Fstatistics value is 14.75 (0.0001), which indicates that attained years of education is endogenous. Additionally, for the intra-district migrants also, we find that years of education is endogenous (4.0, p ¼ 0.046).
Next, we have tested the endogeneity of variables: required years of education, surplus years of education and deficit years of education. Among the non-migrants, we have failed to reject the null hypothesis for deficit and surplus years of education, which means that these variables are exogenous but the required years of education is endogenous. The F-statistic (p-values in parentheses) values of required, surplus and deficit years of education for non-migrants are 1.87 (0.10), 0.03 (0.85) and 0.3 (0.82), respectively; for migrants, the F-statistic values for required, surplus and deficit years of education are 2.04 (0.15), 2.26 (0.13) and 1.58 (0.11), respectively. This indicates that these three variables are exogenous in our sample. For the intra-district migrants, F-statistic values for required, surplus and deficit years of education are 0.14 (0.71), 2.6 (0.10) and 4.4 (0.06), respectively. Thus, for the smaller sample of intra-migrants also, we observe that these three variables are exogenous.
The endogeneity tests indicate that IV model is required for the attained years of education. However, for the required, surplus and deficit years of education, an IV model is not needed due to the exogeneity of these variables in our model.

Results from Lewbel IV estimates
It is clear from the endogeneity test that only attained years of education is endogenous. Further, due to measurement error issues, we have only run the Lewbel IV regression for non-migrants and intra-district migrants.
With respect to the IV estimation for attained years of education for migrants and non-migrants, it is observed that OLS estimates are different from Lewbel IV estimates (Table 6). For the non-migrants, Lewbel IV estimates are marginally smaller than OLS estimates. For Table 5. Returns to education and education-occupation mismatch (EOM) for wage/salary employed migrants in working-age group: by years since migration. the migrant workers, the Lewbel IV model only for the intra-district migrants has been estimated because there is no information about the origin district of in-migrants, so use of IV for these migrants will be incorrect. We know the district of origin and destination for the intra-district migrants only by definition. Therefore, we can use the IV for this subsample of migrants. We observe that Lewbel IV estimates for intra-district migrants are higher than OLS estimates, highlighting that OLS estimates were downward biased ( Table 6). The yearly returns to attained education for the intra-district migrants is around 4% as compared with OLS estimates of around 3%. Ideally, we would want the IV estimation for the full sample available to us, that is, for both non-migrants and migrants. However, information about the district of origin is not available for all the migrants, we are only aware of their state of origin, except for intra-district migrants. Further, aggregating the IV at the state level would make them correlated to other state level variables, thus violating the exogeneity condition. In addition, it would reduce the variation in the IV, thereby reducing their predictive power leading to large standard errors and imprecise/biased estimates. Moreover, if we consider the current district's information for the IV estimation of migrants, it will lead to measurement error because the place of residence may not be the same as the place of education for them. Given this caveat, the IV estimation is only used for the intra-district migrants and non-migrants, whose origin district is the same as the destination district, and therefore, we can be sure about the validity of the IV.
Lastly, we can draw two conclusions from our IV model. First, for the non-migrants, the Lewbel IV estimates are marginally lower than OLS estimates. For the migrants, attained years of education is endogenous and the Lewbel IV estimates for migrants are higher than OLS. This indicates that the estimates for the attained years of education for migrants are higher than nonmigrants, and there are higher returns to being spatially flexible. In other words, the intra-district migration increases the returns to human capital and leads to higher productivity gains for the labour markets. Second, we have found that EOM indicators, that is, required, surplus and deficit years of education are exogenous in our sample, and therefore, OLS estimates are unbiased for these variables. This means that the majority of our results remain unchanged and the core results of our study are robust.

CONCLUSIONS
This paper analyses the extent to which migrants can utilize their education, especially in the context of their occupation. The utilization is tested using the EOM framework. In particular, we have investigated the incidence of EOM and returns to education and EOM for migrants considering the heterogeneity among them.
The main results are as follows. First, while the incidence of EOM (under-and over-education) does not differ much between migrants and non-migrants, migrants with different reasons to migrate, demographic characteristics, spatial factors and migration experience witness markedly different rates of EOM. Second, migrants with a confirmed job at the destination before migration earn higher returns to attained, required, and surplus education as compared with migrants who first migrate and then search for a job. Third, workers (females and higher educated groups) who get higher returns to required and surplus education also get higher penalties for their deficit years of education. Fourth, we have unveiled that irrespective of the destination, workers from urban areas earn higher returns to education, but they are the ones who pay the highest penalty for having deficit education. Further, while the labour market penalizes or rewards EOM equally in terms of statistics irrespective of the distance travelled, the returns to attained and required education vary significantly. Workers who choose to migrate only intra-district (relatively spatially inflexible) get lower returns to education as compared with workers who migrate inter-district (spatially flexible). But as the distance increases, the negative impact of imperfect portability of human capital endowments begins to dominate the positive impact of spatial flexibility on wages. Next, migration experience and assimilation does not lead to differential returns to education and EOM. We have also used the Lewbel IV model to correct the endogeneity of the attained years of education for the intra-district migrant and non-migrants in our sample. The Lewbel Note: ***Significant at 1% level and **significant at 5% level. We only report the second stage estimates for the interest variable. The detailed estimates are provided in Table A14 in Appendix A in the supplemental data online. Source: Authors' calculation based on National Sample Survey Office (NSSO) employment, unemployment and migration survey, 2007-08.
IV estimates are marginally higher for the intra-district migrants, whereas lower for the non-migrant as compared with the OLS estimates. Lastly, we have found that the required, surplus and deficit years of education are exogenous in our sample and therefore, do not require IV estimation.
The key implications of this study are as follows. First, by highlighting the differential in the labour market outcomes for the varied migrants, the study stresses the need that while analysing the decision to relocate, individuals should consider these differences in the returns to attain the maximum benefits. For example, it would be beneficial for the individuals to have a confirmed job at a destination before they decide to migrate. Also, to achieve the optimal gains to education, an individual should move to a place where he or she will be able to adequately transfer the human capital. Second, while spatial flexibility is seen as an important policy tool to redistribute the human resources for achieving higher productivity and growth for a national economy as a whole (Maxwell, 1988;Sahota, 1968), the under-utilization of human capital endowments of migrants could reverse such effects. The literature has found that apart from lower wages, EOM can lead to various other adverse consequences such as lower productivity (Tsang, 1987), higher attrition (Verhaest & Omey, 2006), etc. Hence, adequate attention has to be paid on the migrants' EOM to achieve the desired results.
The generalizability of results has always been a concern for country specific studies, that is, whether the results have external validity. With respect to this, two sets of arguments can be provided. First, the results are not specific to the nature of internal migration being measured. Similar results are also obtained by a number of studies as cited in the results section (Aleksynska & Tritah, 2013;Devillanova, 2013;Hensen et al., 2009;Nieto et al., 2015). Further, we have explored several dimensions of internal migrants based on their characteristics (reason for migration, migration distance, years since migration and migration stream). Thus, we provide more comprehensive analysis of internal migrants and EOM. All the characteristics observed or measured in the migration surveys can be adopted with respect to country-specific characteristics. Thus, the findings from India may be useful depending on the nature of internal migration. Additionally, in our study, we have controlled for a number of regional and local labour market, individual, household, migration specific variables, thereby providing a robust set of results. Thus, our study will surely be useful to researchers from other countries, mainly developing countries, who would be doing research on internal migration. Second, it can be argued that India accounts for around 17% of the world population and is the second-largest country by population. Further, out of the estimated 740 million internal migrants worldwide in 2009 (Klugman, 2009), around 280 million (estimates based on NSSO, 2010, as per the 2007-08 survey) happen in India. This is slightly more than one-third of the total internal migrants in the world. Hence, our data is representative of these internal migrants. Therefore, understanding their labour market outcomes will provide insights for studies on internal migration especially in developing countries.
Although the study aims to provide a rich description of internal migration and EOM, the limitations are inevitable. The lack of information on skills and quality of education makes it difficult to differentiate migrants with similar levels of education. Therefore, it may be possible that over-educated workers are under-skilled, which makes them earn lower wages. However, this dimension cannot be explored due to the lack of data on the skill level or quality of education. Another important aspect is to account for other sources of human capital formation apart from formal education in these models, especially for the developing countries, where informal training and on the job learning remains important in determining the match between workers and their jobs. 21 However, converting these various dimensions into a single index value for the purpose of comparison is a tedious task. Lastly, there is an unavailability of information on spatial aspects of the labour market. For example, having geocoded labour market datasets for such analysis would help in understanding the spatial dependence across various local labour markets, and their implications for misallocation of labour and corresponding returns to human capital. Information on the spatial location of the migrants at the destination can help us understand the mechanisms through which EOM are corrected or accentuated. Further, the probability of employment and consequent match status is affected by the prevailing conditions in the local labour market. Therefore, the spatial information on labour market aspects can further enrich this study. Lastly, in the context of IV models, we can only use the non-migrant and intra-district migrants' sample for our analysis due to lack of information about the district of origin of the migrant individuals. This has restricted our analysis for Lewbel IV estimate to the subsample and we cannot estimate models for heterogeneity along other dimensions in our study.
boundaries of a country (National Sample Survey Office (NSSO), 2010). This study focuses on India, and hence the definition provided comes from the Census of India and NSSO. The definition is also used by the Government of India for administrative and public policy purposes. Other countries also use similar definitions with slight variation in the duration considered. 2. EOM refers to the incongruity between the attained education of a worker and the required education by her occupation (Duncan & Hoffman, 1981). 3. Although this phenomenon is also present in the case of international migration, it is suppressed by the negative impact of imperfect portability of human capital. 4. It has been argued in the literature that widening the labour market of a worker can improve the returns to education. Thus, in the past studies it was hypothesized that migration will lead to better returns to education as this would enable the individuals to have access to a larger labour market. However, in this paper we argue that not all migrants are the same and may still face geographical limitations which may lead to suboptimal returns to education. 5. We categorize a migration to be work related if the reason to migrate is stated as any of the following: in search of a job, to take up a confirmed job, business, transfer of service/contract or proximity to place of work. 6. See https://economictimes.indiatimes.com/news/ politics-and-nation/over-93000-candidates-including-3700-phd-holders-apply-for-peon-job-in-up/articleshow/ 65604396.cms?from=mdr/ (last accessed on 31 October 2021). 7. For more detailed descriptive statistics for migrants and EOM, see Table A1 in Appendix A in the supplemental data online. 8. Section A1 in Appendix A in the supplemental data online discusses the sensitivity of these estimates to various threshold cut-offs being taken for measuring incidence of EOM. 9. The age and age squared term variables are used to capture the effect of learning on the job, and experience in the labour market, especially when no explicit indicators of these aspects of human capital are available in the labour market surveys. We are thankful to an anonymous reviewer for highlighting this aspect. 10. The NSSO provides occupation codes at three-digit levels, which can be integrated into nine broad categories. While we use occupation codes at three-digit level to measure the EOM, we are controlling for broad categories in our estimation. 11. For the complete list, see http://mospi.nic.in/sites/ default/files/main_menu/national_industrial_ classification/nic_2004_struc_detail.pdf/. 12. This indicator is estimated at the state level to avoid the measurement error (attenuation bias) due to very low sample size at the district level. We are thankful to an anonymous reviewer for this suggestion. 13. India is a federal country with three tiers of administrative bodies, which affect human mobility through various place-based policies, that is, district, state and central government. Based on this identification of boundaries, we can identify three types of migrants in our survey data. One, intra-district migrants who migrate within the boundaries of the district which is the smallest administrative unit observable in the data used. Second, inter-district (within-state) migrants who cross the district boundaries but remain within the state in the process of migration. Third, inter-state migrants who cross the state boundaries. Based on these three definitions, generally, intra-district migrants travel the least distances while inter-state migrants travel the most distances. 14. Interestingly, in the literature on EOM, studies have never discussed the concern of potential correlation between required, surplus and deficit years of education. We are thankful to an anonymous reviewer for raising this concern. In our dataset we have observed that the pairwise correlation of required years of education with surplus and deficit years of education is −0.08 and −0.09, respectively, which is small. However, the pairwise correlation between surplus and deficit years of education is 0.5. This will increase the standard error of the estimated coefficient by the variation inflation factor (VIF = 1(1 − r2), where r is a pairwise correlation) of 1.33. Thus, the concern of multicollinearity in this case is not present. 15. Although the employment decision is highly gender specific, in our model we do not run separate probit models for males and females because the sample of female migrants is quite small in our data. However, we control for gender and other demographic variables in the model to address this issue. 16. Card (1993) had past information about the geographical location of the household, which was used to identify the corresponding distribution of schools for the households. However, such information is not available in our survey. 17. For the results for the full set of controls, see Tables A5-A10 in Appendix A in the supplemental data online. 18. Section A4 in Appendix A in the supplemental data online also provides the estimates for returns to education for other than work-related migrants, which include education, forced migration (due to natural disaster, sociopolitical problems and displacement for development projects), marriage, tied movers (due to mobility of parents/ family members) and others. 19. The district is the smallest geographical unit at a subnational level. In 2007-08, India comprised 35 states and union territories, 87 National Sample Survey (NSS) regions and 588 districts. 20. We have also analyed the returns to education and returns to EOM for married migrants and by gender. The results are discussed in sections A5 and A6, respectively, in Appendix A in the supplemental data online. 21. We are thankful to an anonymous reviewer for highlighting this important aspect of labour markets in the developing countries. ORCID Shweta Bahl http://orcid.org/0000-0002-9238-1570 Ajay Sharma http://orcid.org/0000-0001-5150-299X