University of Birmingham Do you really need it? Educational mismatch and earnings in Ghana

This study examines the effects of educational mismatch on earnings using cross-sectional data from three rounds of the Ghana Living Standards Survey. There are gender-earnings differentials, with women suffering a penalty for undereducation that is almost twice that paid by men for having less than the required years of education, although the premium associated with overeducation is higher for women. While the penalty associated with deficit schooling decreases for men over time, this penalty increases for women. Women but not men experience significant increases in returns for attaining the required education. On the contrary, the lack of significance of returns to overeducation over time suggests that human capital beyond the required level is unproduc-tive and may represent a waste of public resources. K


| 3
DARKO AnD ABROKWA for the analyses (Section 3). Next, the estimation techniques are described in Section 4 followed by a discussion of the results (Section 5). Section 6 concludes.

| EDUCATION AND THE LABOR MARKET IN GHANA
Ghana's economic growth has been remarkably strong, with its lowest economic growth of 3.3% recorded in 1994 and highest of 15% recorded in 2011 (Aryeetey & Baah-Boateng, 2015). Despite these improvements in the economic performance of Ghana, a major developmental concern is the ability of the country to translate this strong growth performance into job creation. With an annual average job creation growth of 3.2% between 1991 and 2005, the country has not been able to keep up with the increase in the potential workforce, which averages approximately 3.8% per year. Job creation increased to 4% between 2005 and 2012, but annual growth among the working population decreased to 2.6%, with economic growth averaging 8% per year over this period (See Honorati & de Silva, 2016 for a further discussion on job creation in Ghana). The low level and quality of human capital make it difficult for the benefits of growth to be spread through the creation of high-skilled and high-earning jobs (Aryeetey & Baah-Boateng, 2015). The majority of jobs are low skilled and require a limited use of cognitive skills (Handel et al., 2016;Honorati & de Silva, 2016). The prevalence of low-skilled jobs is partly explained by the low education level among the employed although it is also possible that the prevalence of low-skilled jobs could lead to lower levels of educational attainment. Nonetheless, Honorati and de Silva (2016) report that two out of five workers have no more than a primary education (first 6 years of formal education), and only one in five has more than a basic education, that is, primary and junior secondary school education.
The education system in Ghana has undergone a number of reforms over the past few decades. Figure 1 illustrates the structure of the education system in the late 1960s and the reform of 1987. Prior to the 1987 reform, pretertiary education consisted of 6 years of primary education; 4 and 5 years of middle and secondary school, respectively; and 2 years of A-level school. The 1970s saw further reforms, such as the introduction of junior secondary school and senior secondary school initiated by the Dzobo Committee. In 1987, the education system was reformed with the introduction of 12 years of pretertiary education consisting of 6 years of primary education, 3 years of junior secondary education, and 3 years of senior secondary education. The main objective of the 1987 reform was to expand and improve the quality of education in Ghana and make basic education free and equitable. The reform helped reverse the decline in enrolments that characterized the early 1980s (World Bank, 2004). The reform also aimed to contain, and partially recover, costs and to enhance sector management and budgeting procedures.

DARKO AnD ABROKWA
A further review of the education system took place in 2007 under the New Patriotic Party, which led to the extension of the duration of senior high school from 3 to 4 years. However, this policy was short-lived, and in 2009, the government reverted to 3 years of senior high school (see Adu-Gyamfi, Donkoh, & Addo, 2016;Akyeampong, 2010;Akyeampong, Djangmah, Oduro, Seidu, & Hunt, 2007;Blunch & Hammer, 2018 for more discussion on the education system in Ghana).
There is evidence of educational mismatch in the Ghanaian labor market, with almost 50% of workers underqualified in 2005 (Honorati & de Silva, 2016). Figure 2 shows the incidence of educational mismatch by employment type from 1998-1999 to 2012-2013. The incidence of undereducation increased over this period and was particularly high among agricultural workers. Among wage workers, whose earnings are usually higher, the incidence of overeducation fell from 53.5% in 1998-1999 to 24.8% in 2012-2013, whereas undereducation increased from 37.9% to 44.1% in 2012-2013. The analysis detailed in this paper builds on this evidence.

| Data
The data used are from the GLSS rounds 4-6. The GLSS is a nationally representative survey initiated in 1980 by the Policy Research Division of the World Bank. Round 4 of the survey was conducted in 1998-1999 and covered 300 enumeration areas. Rounds 5 (2005Rounds 5 ( -2006 and 6 (2012-2013) covered 580 and 1,200 enumeration areas, respectively. The enumeration areas are stratified into 10 administrative regions, and each region is subdivided into rural and urban areas. Figure A1 shows the distribution of the sample across the 10 administrative regions. The total number of households surveyed was 6,000 in round 4; 8,687 in round 5; and 16,772 in round 6. F I G U R E 2 Incidence of educational mismatch by employment type, 1998-1999 to 2012-2013  The survey collects detailed information on individual and household characteristics, as well as information on current education and employment status. The sample consists of individuals aged 25-60 years who have ever attended school and reported sufficient information to calculate their level of education (years of schooling completed) and other measures relevant for the study. In the analysis of earnings, the sample is restricted to those in work at the time of the survey, including the self-employed. The upper limit of 60 years reflects the official retirement age in Ghana. This restriction resulted in around 38,680 observations. We also excluded family workers and apprentice and domestic workers from the analysis 4 because earnings for this group of workers are difficult to measure. For instance, apprentices in Ghana are not paid directly but instead pay their "masters" to be trained. Similarly, domestic workers often receive other forms of nonwage compensation. This resulted in a further reduction in the sample to around 32,399 observations. In addition, the final sample is restricted to individuals for whom records of highest grade completed were available because they had attended school at some point even if this was for less than a year (zero completed schooling). Although in some occupations those who never attended school (zero education) could be disproportionately represented, including this group raises a number of issues. First, assigning educational attainment a value of 0 (to indicate that they have not completed any grade) for those who never attended school would treat this group of workers as if they were the same as those who attended school but not completed a grade. However, these two groups are not the same. In the case of the latter, some form of learning may have still taken place even though no grade was completed. The two groups could also be different in terms of other observables. Adding the group that never attended school would also skew the distribution. We therefore exclude these individuals from our analysis. This additional restriction led to a further reduction in the sample to around 22,436 who were in work.
In the regression analysis for earnings estimated for only individuals who were in work, the sample was reduced further due mainly to missing values for earnings as well as some other included independent variables, resulting in a final sample of 18,396 in the earnings regressions. Table A1 provides the summary statistics for the main variables used in the analysis. Table A2 provides descriptions for the main variables.
Men were marginally more represented in the full sample than women but were strongly represented in the employed sample. The majority of workers were in agriculture and services and sales occupations. These two occupation groups accounted for approximately 50% of the total share of employment. Women were disproportionately located in services and sales. Figure 3 shows the distribution of earnings for men and women between 1998-1999 and 2012-2013. The distribution for men lies to the right of the distribution for women, consistent with a gender wage gap. The Kolmogorov-Smirnov test of significance (Wilcox, 2005) suggests that this difference is highly significant.

| Educational mismatch measures
Three main methods (as noted in Section 1) have been used to derive measures of educational mismatch. The subjective method (self-assessment) is based on worker assessment of the level of educational attainment that is required to perform a job. This assessment is then compared with the worker's actual level of educational attainment to determine if they are matched or mismatched. Several drawbacks have been associated with this method. For instance, workers may not have the required level of information about the job, and therefore, their responses could be inaccurate (Leuven & Oosterbeek, 2011). Workers may also overstate the requirements of their job (Borghans & de Grip, 2000;Hartog, 2000), which could lead to a bias in their estimates (McGuinness, 2006;McGuinness, Pouliakas, & Redmond, 2018). Studies such as Dolton and Vignoles (2000), Belfield (2010), Daly et al. (2000), and Battu, Belfield, and Sloane (2000) have used the subjective (self-assessment) method to measure educational mismatch.
The second method is the objective approach (job evaluation), which is based on evaluations by professional job analysts. The information generated from such analyses is documented in the Dictionary of Occupational Titles and subsequently the Occupational Information Network (BLS, 1999). While the objective approach is based on experts' assessments, a major drawback is that updates are carried out infrequently, largely because of the associated costs of collecting these data (Hartog, 2000;McGuinness et al., 2018;Verhaest & Omey, 2012). Furthermore, because the classifications are based on opinions, albeit of experts, this approach can still suffer from subjectivity bias. This method has been used by a number of authors including Sloane, Battu, and Seaman (1999), Green and Zhu (2010), Cohn and Ng (2000), and Verhaest and Omey (2009).
The final method is the realized matches (statistical) method, which relies on education and occupation information, both of which are readily available in many nationally representative surveys (Flisi, Goglio, Meroni, Rodrigues, & Vera-Toscano, 2017;Hung, 2008). An educational mismatch is determined if the worker's education level is above the mean or modal value for the occupational group (Kiker et al., 1997;Verdugo & Verdugo, 1989). A worker is overeducated if he or she has more years of education than is required. Conversely, a worker is undereducated if he or she has insufficient education to be hired for the job. An advantage of the realized matches approach is that the educational standard of the status-specific occupational benchmark is updated regularly to capture changing educational requirements over time. This method has been extensively used in the literature (Bauer, 2002;Chiswick & Miller, 2009De Oliveira, Santos, & Kiker, 2000;Hartog, 2000;Herrera & Merceron, 2013;Mateos-Romero & Salinas-Jiménez, 2017;Ng, 2001;Quinn & Rubb, 2006;Rubb, 2014;Robst, 1994;Salinas-Jiménez, Rahona-López, & Murillo-Huertas, 2013;Verhaest & Omey, 2006;Yeo & Maani, 2017). Given the availability of information on education and occupation in the F I G U R E 3 Gender differences in earnings, 1998-1999  GLSS data, the realized matches method is appropriate for this analysis. Using this method, the mismatch variables for the pooled sample are constructed by taking the modal value of years of completed schooling for individuals who ever attended school across the different occupations. However, we acknowledge that the realized matches method is imperfect. For example, as pointed out to us by an anonymous reviewer, the modal worker may be overeducated or undereducated in some occupations. For this reason, we also conduct the analysis using an alternative measure of educational mismatch (described later).
Measures of educational mismatch are usually decomposed to capture overeducation, required education, and undereducation (Duncan & Hoffman, 1981). This decomposition, often known as ORU, is useful for a number of reasons. ORU is able to combine information on attained education and required education while maintaining the continuous nature of both dimensions (Korpi & Tåhlin, 2009). In addition, the pattern of results for research that has used this approach has proven to be consistent (Rubb, 2003). We use the modal value of completed years of education for each major occupational group to derive measures of required education, overeducation, and undereducation. Compared with the mean approach, the modal approach reduces sensitivity to outliers and provides a more accurate measure of adequate education. In addition, the required years of education derived from this measure refer to the actual education levels of workers in a job, rather than fractions of years of education. Figure 4 shows the distribution of educational mismatch utilizing these measures. The proportion of overeducated workers remained fairly unchanged over the period. However, there was a continuous increase in the proportion of undereducated workers. For instance, between 1998-1999 and 2012-2013, the percentage of workers who were undereducated increased from 38% to 55%. Interestingly, the share of workers with the exact level of educational attainment for their occupations fell throughout the period from 43% in 1998-1999 to 25% in 2012-2013. These trends raise concerns about the education system and the labor market given recent advances in educational enrolment and participation. Figure 5 shows the incidence of educational mismatch by gender over the three survey rounds. As shown, for both men and women, the proportion of workers in jobs that match their educational attainment consistently fell. In 1998In -1999.2% and 41.9% of men and women, respectively, had the required education for their jobs. These figures fell to 28.7% and 23.9% in 2012-2013. Undereducation increased throughout the period for both men and women although the incidence of undereducation F I G U R E 4 Incidence of educational mismatch over time was higher among women; the incidence of female undereducation increased from 46% in 1998-1999 to 60% in 2012-2013. The increase in undereducation, especially among women, suggests that mismatches in the Ghana labor market remain a challenge to effective labor market performance. In terms of overeducation, men are more likely to have more than the required level of education needed for their jobs than women. Figure 6 shows mismatches in years of schooling by gender for the three rounds of the survey. There are clear differences in educational attainment between men and women. The difference in average years of required education increased from approximately 10 and 10.2 in 1998-1999 (round 4) to approximately 10.3 and 10.8 in 2012-2013 for men and women, respectively. As the mismatch variables are constructed by pooling the sample and taking the modal values for each occupation, the increases in the years of required education are driven largely by changes in the occupational composition of individuals. With regard to years of overeducation, men accessed more schooling than women. For women, the extra years of schooling above the required level remained almost unchanged throughout the period. In contrast, women tended to have more years of deficit schooling than men although this fell from 3 years in 1998-1999 to 2.5 years in 2012-2013. The opposite was true for men: years of deficit schooling increased from almost 1.5 in 2005-2006 to 1.8 years in 2012-2013.

| Alternative education measure
Realizing some of the limitations inherent in measuring educational mismatch using, for instance, self-reported measures, and the possibility for the representative worker to be overeducated or undereducated in some occupations, we adopt an approach as an additional measure of (mis)match that reflects the outcome of supply and demand in the labor market (Boualam, 2014 DARKO AnD ABROKWA used to construct a matching index for each individual at the educational attainment level, for employment type and occupation. This index is defined as.
where and N educ,emp,occ represents the total number of workers with a given educational attainment, educ, in a given employment type, emp, within an occupation, occ. N occ is the total number of workers in a given occupation. share educ,emp,occ represents the share of workers in a given occupation and employment type with a given educational attainment. share educ,emp,occ is rescaled by dividing by the maximum value, share max occ , that can be observed for all educational attainment levels within each occupation. A coefficient close or equal to 0 indicates that this type of educational attainment is rarely observed within the occupation, and a coefficient close or equal to 1 indicates a match.
An advantage of this approach is that it does not assume that only one level of educational attainment is ideal for a given occupation. Different matches may be observed in any given occupation. In addition, the measure is continuous, thus allowing us to observe a full range of values reflecting

| Estimation technique
We examine the returns to education using augmented Mincer-earning-type regressions. As the sample used in these estimations consisted of only employed individuals for whom earnings were observed, any estimates that failed to consider selection into employment would yield biased estimates, partly because the sample of employed individuals may not be a random subset of the population. The willingness of individuals to supply labor may therefore not be randomly assigned. We address this issue using the Heckman selection procedure (Heckman, 1979) by first deriving the inverse Mills ratio (IMR), also referred to as Heckman's lambda, from the employment participation equation estimated using probit. The selection term (IMR) is then included as an additional explanatory variable in the earnings equation. The estimated employment participation equation is of the form.
where Employed indicates the employment status of the individual, with 1 indicating that the individual is currently employed and 0 if not employed. x i is a vector of individual characteristics, including gender, age, and marital status. j i is a vector of the round, location, and region dummies. g i consists of variables that are not included in the earnings regressions but satisfy the exclusion restrictions. The restriction requires that these variables should directly affect employment probability but should not have a direct effect on earnings. We use the number of young children (less than 6 years) in the household and the local average share of working individuals. Our rationale for including these variables is that young children may impose a time constraint and can therefore affect the individual's participation in work (Huber & Mellace, 2014;Mulligan & Rubinstein, 2008), particularly for women. Older children, on the contrary, are less dependent and less likely to affect the respondent's employment decision. Similarly, the local share of workers is a possible indicator of community economic development so that individuals in communities with a larger share of workers will be more likely to find jobs. 5 Pagán (2002) finds that a number of working individuals have positive and significant effects on employment propensity. These variables can affect the decision to work but will not directly affect earnings. The employment participation equations are estimated jointly and separately for men and women.
The second stage of the estimation focuses on earnings. The earnings regression includes the Heckman selection term, IMR (or lambda). The equation estimated is of the form.
where lny i is the log of weekly earnings of individual i. REQEDUC, OVEREDUC, and UNDEREDUC are the required education, overeducation, and undereducation, respectively, all in years. IMR (lambda) is the Heckman selection term derived from Equation 2. z i is a vector of individual characteristics, including gender and marital status. z i also includes the number of hours worked per week, experience (calculated as age minus years of schooling minus compulsory school start age [6 years]), and its square and employment type. j i is as previously defined and is a vector of round, location, and region dummies. The data collection round and location dummies capture any possible effect of time on earnings. j i also includes occupational dummies.
Finally, we estimate the returns to education using the alternative index measure (match educ,emp,occ ) of educational match as specified in Equation 1 as an independent variable. The estimated equation is of the form.
All parameters are as previously defined.

| Potential estimation issues
Equation 3 assumes that the education variables are exogenous. However, the literature on returns to education shows that education is endogenous due to certain unobserved factors that influence the individual's ability to acquire more years of schooling (while not having an independent effect on an individual's earnings) as well as measurement errors (Bauer, 2002;Harmon, Oosterbeek, & Walker, 2003;Korpi & Tåhlin, 2009). As such, all three measures of education may be endogenous and therefore must be accounted for using instrumental variables (Leuven & Oosterbeek, 2011). To apply the instrumental variables technique to produce acceptable estimates, they must satisfy two conditions: relevance and exogeneity. Regarding relevance, there should be a correlation between the instrument and endogenous regressor. Exogeneity requires no correlation between the instrument and error term. (3) Previous studies address this issue using retrospective data on the place of residence during childhood, economic problems, and disruptions in the family of origin, number of siblings, and family background (Dearden, 1999a;Dolton & Silles, 2008;Harmon et al., 2003;Korpi & Tåhlin, 2009). Other studies use panel data to estimate fixed effects regressions to control for unobserved, time-invariant factors (Bauer, 2002;Dolton & Silles, 2008).
A key challenge faced with the GLSS data is that there are no retrospective questions on childhood circumstances, such as the number of schools available in a neighborhood, family background, ability, and economic conditions, that can be used as potential instruments for education in the main estimation equation (Equation 2). 6 Dearden (1999b), however, examines the accuracy of OLS coefficients against results that attempt to eliminate biases from OLS estimates, such as measurement error and endogenous education choices. The conclusion from Dearden (1999b) is that failure to control for ability and family background characteristics that influence education choices will bias OLS estimates upward, whereas measurement error can only lead to a downward bias. Given our inability to fully address possible endogenous education choices, we proceed cautiously with the OLS estimations that we believe are fairly reasonable estimates of the true returns to education and generate effects that are, in most cases, of the same direction as one would have achieved with IV estimates. We are mindful of this caveat and therefore refrain from making any causal claims from our findings and instead interpret all coefficients as associations.

| Employment participation and returns to education for working individuals
This section presents the results from employment participation regressions and OLS selectivitycorrected earnings regressions. The results from the first-stage participation regressions are presented in columns (1)-(3) of Table 1. An immediate observation is that the coefficients for the number of younger children are negative and significant in the pooled regression and also for women. The share of working individuals at the local level is also positive and significant. The results clearly suggest that especially for women, having more young children decreases the probability of being employed, as evidenced in the literature. For men, this is not the case. In addition, living in a location with a larger share of individuals working significantly improves employment probability. The coefficient on the gender dummy is negative and significant, suggesting that women are less likely to be employed than men.
The estimates for the OLS selectivity-corrected earnings regressions in columns (4)-(6) of Table 1 provide pooled and separate earnings regressions for men and women. Including indicators of required, overeducation, and undereducation separately allows for different effects on earnings. The dependent variable in both models is the natural logarithm of weekly earnings. Across all specifications, the selectivity term, IMR (lambda) is positive and mostly significant. This indicates that individuals with the average characteristics of those who select into employment earn higher earnings than would individuals drawn randomly from the population with comparable characteristics.
For brevity, we focus on the education mismatch measures. In the pooled regression (column 4), the coefficients of required education, overeducation, and undereducation have the expected signs and are statistically significant at all levels. Each year of schooling beyond that required increases earnings by 8.7%. Despite the positive returns to overeducation, this effect is less than the earnings associated

| 15
DARKO AnD ABROKWA with individuals that have the required years of schooling (0.29). Regarding undereducation, there is a penalty of 3.7% for each year of deficit schooling. The direction of the effect is the same for both men and women (columns 5 and 6) although the penalty received for deficit schooling is larger for women than for men. Women with fewer years of schooling experience an earnings loss that is almost twice (4.6) the loss suffered by their male counterparts (2.9). The direction of the effects of the education mismatch variables is in line with the literature (Hartog, 2000;Kiker et al., 1997). The gender differences in returns to undereducation are also consistent with some previous research (Salinas-Jiménez et al., 2013) but contrary to others. Columns (7)-(9) show the results of the estimations that use the alternative index measure of educational (mis)match. The OLS results indicate that an increase in educational match increases earnings by 0.22. The coefficients are larger for men than for women, in line with the coefficients for required education in columns (5) and (6), with earnings increasing by 0.26 for men compared with 0.2 for women. As in column (4), the coefficient on the gender dummy is negative, confirming that women earn lower wages. The signs of the other variables are similarly uncontentious.

| Educational mismatch and differences in returns over time
This section provides a brief analysis of differences in the effects of mismatch over time. This is done by including interaction terms between the survey round indicators and the three measures of education. Table 2 presents the results of these estimations. We comment only on the interaction terms.
In columns (1)-(3), the interactions between the survey rounds 5 and 6 and required education indicate an increase in earnings relative to round 4 although only marginally for the pooled sample. Only women experienced a significant increase in returns to required education in rounds 5 and 6 compared to round 4. Interestingly, there were no significant returns to overeducation for both men and women over the three periods although the effects were positive. Overall, there was a negative effect of undereducation on earnings, shown by the interaction between the survey rounds and undereducation. The negative effect of undereducation was larger for men than for women over the period. However, while the penalty associated with undereducation decreased for men between survey rounds 5 and 6, the penalty suffered by women for being undereducated increased. Columns (4)-(6), which used the index measure of match, show that earnings for both men and women increased over time with increased educational match.
As a sensitivity check, we generated another measure of match by comparing the individual's level of educational attainment with the level of education required by occupational group (ILO, 2014) using the definitions in the International Standard for the Classification of Occupations (ILO, 2012) and the associated skill level according to the International Standard Classification of Education (UNESCO Institute for Statistics, 2012). To derive the measure, the major occupation groups were grouped into four, and a level of educational attainment was assigned to each. 7 A match occurs if workers in a particular group have the assigned level of educational attainment. Those with a lower (higher) level of education for the occupation are considered undereducated (overeducated). This approach to determining a mismatch has been used elsewhere in the literature (Kalfa & Piracha, 2017;Mateos-Romero & Salinas-Jiménez, 2017). The results from these estimations are presented in Table A3. The results are consistent with the results in Table 1; overeducated individuals receive a wage premium compared with those with the assigned level of education, whereas undereducated individuals pay a penalty relative to those who have the required education level for their occupational group. Note: IMR = inverse Mills ratio; robust standard errors in parentheses.

| CONCLUSIONS
We have examined gender differences in educational mismatch for employed workers in the Ghanaian labor market and the implications of mismatch on earnings. The empirical analysis was based on cross-sectional data from three rounds of the GLSS, 1998-2013. The realized matches approach to measuring educational mismatch was used in addition to an alternative method that does not assume that only one level of educational attainment is ideal for a given occupation.
The results from the selectivity-corrected OLS earnings regressions are broadly in line with the literature. First, there are positive returns to surplus education although the returns are smaller than the returns received by workers with the required years of education. The premium for overeducation was found to be greater for women. Second and in contrast to the results for surplus and required education, workers with fewer years of schooling than that required for their jobs received a penalty. Women received a penalty that was twice the penalty received by men for having fewer years of schooling than required. The results from the alternative measure also showed that individuals with a better match were associated with higher earnings although the effects were larger for men than for women. Last, we considered differences in the effects of educational mismatch over time. We found that while the penalty associated with deficit schooling decreased over time for men, it increased for women. In addition, women rather than men were associated with a significant increase in returns to required education. The lack of significance of returns to overeducation over time indicates that there are disconnects in the linkages between Ghana's educational system and the labor market; the value of human capital beyond the required level may erode over time, representing a waste of public resources.
The policy significance of these results does not lie only in the fact that they provide evidence of educational mismatch. The results also indicate that educational mismatch, in particular undereducation, is more prevalent among women. There are possible repercussions for the Ghanaian labor market. The fact that employers are hiring workers with fewer years of education than required for their job indicates a possible skill shortage. To improve the labor market position of Ghanaians, particularly women, better access to quality education is needed to ensure that people leave school with the required skills and knowledge. It is also important for individuals to consider their educational investment decisions carefully. Accurate knowledge of occupation-specific requirements would help individuals to choose the right amount of education. Although it is important to create jobs, there is a need to improve the effectiveness of the search and hiring methods of firms. Specifically, it is imperative to reduce reliance on personal and social networks in labor market recruitment, commonly referred to as "who-you-know" and shift to a system based on merit. This shift could help ensure that more individuals are employed in jobs that match their level of education.
The results presented in this study indicate associations rather than causations, mainly because of our inability to identify strong instruments that could have been used to identify possible causal mechanisms of the effects of education on earnings. Although other studies have also proceeded in this way, it is important for future studies in developing countries to address issues of endogeneity to enable a better understanding of how a mismatch in employment impacts earnings. In addition, the use of panel data to unravel some of the unobserved heterogeneities would provide a further understanding of the phenomenon. A further limitation of this study is that the sample used consists of only individuals who ever attended school. Future research could benefit from the consideration of those who never attended school. This group may have particular characteristics that are important for extending the understanding of earnings differentials. In addition, the use of years of completed schooling to construct measures of educational mismatch is unable to capture aspects of education quality, which has mostly been found to induce higher years of schooling (Castelló-Climent & Hidalgo-Cabrillana, 2012;Dearden, Ferri, & Meghir, 2002;Harmon & Walker, 2000). This limits our ability to make any inferences about how the mismatch variables could be affected by education quality. Further research is needed to understand the role of education quality in mismatches in the labor market. Last, future surveys of developing country labor markets should provide more detailed measures of education quality, skills, and information on job requirements that are comparable with labor market surveys in most developed countries. This would enable a more comprehensive analysis of how different kinds of educational and skill mismatches, as well as education quality, affect individual earnings across countries and between regions. Most importantly, these surveys need to be nationally representative, and coverage should be for all sectors.

ACKNOWLEDGMENTS
The authors are grateful to the Ghana Statistical Service for making the Ghana Living Standards Survey available. The authors are indebted to two anonymous reviewers for their helpful suggestions and to Fiona Carmichael and Ferran Vendrell-Herrero for insightful comments.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are from the fourth (1998-1999), fifth (2005-2006), and sixth (2012-2013) rounds of the Ghana Living Standards Survey available from the Ghana Statistical Services at https://www.stats ghana.gov.gh/gssda tadow nload spage.php.

Christian K. Darko
https://orcid.org/0000-0002-1665-2594 ENDNOTES 1 This body of evidence is associated with three main methods to measuring mismatch. Belfield (2010), Daly et al. (2000), and Dolton and Vignoles (2000) use the subjective/self-assessment method, whereas Hartog (2000) and Kiker et al. (1997) use the realized matches method. Cohn and Ng (2000) employ the objective method. Rubb (2003) and Groot and Van den Brink (2000) use a meta-analysis that consists of the different methods. See McGuinness (2006) for an overview of the literature on overeducation. 2 The sample of countries analyzed included Benin, Burkina Faso, Cameroun, Congo DR, Cote D'Ivoire, Madagascar,Mali,Niger,Senegal,and Togo. 3 The countries analyzed were Armenia, Bolivia, Georgia, Ghana, Kenya, Macedonia, Lao DPR, Sri Lanka, Ukraine, Yunnan Province (China), and Vietnam. 4 In a separate analysis that is not reported, we included this group of workers. The results from these estimations remained fairly unchanged. 5 In separate results not reported here, we excluded the local share of working individuals as an exclusion restriction.
The results from these estimations were similar. 6 We made attempts to address endogeneity concerns by experimenting with a number of potential instruments for the education variable, that captured exogenous variations in developmental progress and education policy, as well as family background. The measures included women's property rights at birth (see, e.g., Allendorf, 2007;Mishra & Sam, 2016 on women's property rights and children's educational attainment); local average years of schooling at year of birth; and rainfall at age 15 years. The 1987 education policy reform was also used as an instrument as education policy reform has been used in previous research as instrument for education (Denny & Harmon, 2000;Meghir & Palme, 1999). Postestimation tests from these estimations showed that the instruments were weak (failing to satisfy the relevance criteria) and therefore could not be used in instrumental variables regression. 7 Major group 9 is assigned primary or less educational attainment, major groups 4-8 are assigned junior secondary and middle school educational attainment, major group 3 is assigned secondary educational attainment, and major group 1 is assigned post-secondary and tertiary educational attainment.  Note: Robust standard errors in parentheses.