1 Economica (2003) 70, Measuring Over-education By ARNAUD CHEVALIER University College Dublin and London School of Economics Final version received 20 November Previous work on over-education has assumed homogeneity of workers and jobs. Relaxing these assumptions, we find that over-educated workers have lower education credentials than matched graduates. Among the over-educated graduates we distinguish between the apparently over-educated workers, who have similar unobserved skills as matched graduates, and the genuinely over-educated workers, who have a much lower skill endowment. Overeducation is associated with a pay penalty of 5% 11% for apparently over-educated workers compared with matched graduates and of 22% 26% for the genuinely over-educated. Overeducation originates from the lack of skills of graduates. This should be taken into consideration in the current debate on the future of higher education in the UK. INTRODUCTION Participation in higher education in the United Kingdom has increased considerably in the past fifteen years; the proportion of a cohort attending tertiary education has soared from 15% in 1985 to more than 33%. As rates of returns have remained stable over the period (Chevalier and Walker 2001), it appears that the supply and demand of graduates have grown at similar rates. However, average returns may hide a great disparity in the outcomes faced by UK graduates. Various studies have shown evidence of an excess supply, since up to 40% of UK graduates have too much education for their job (Mason 1996; Dolton and Vignoles 2000). How to measure too much education has been a weak point of the over-education literature. The different definitions proposed all have caveats. This paper presents an alternative definition of overeducation. The countervailing facts of high returns and a possible large excess supply of graduates can be reconciled by relaxing the assumption that graduates are homogeneous in their skills endowment. 1 Widening the access to higher education has increased the heterogeneity of the skills of new graduates entering the labour market. 2 Hence previous studies on over-education, which implicitly assumed homogeneity of workers, have overestimated the true extent of the phenomenon. This paper proposes to divide over-educated workers by their skill level. First, I rely on a proxy variable concerning satisfaction with the match between education and job. I assume that employees in a non-graduate job who are satisfied with this match are only apparently over-educated, while those who are dissatisfied are genuinely over-educated. 3 This alternative definition of over-education also allows us to account for the possible heterogeneity of jobs, which has been largely neglected in most of the empirical work. In a second step, I use the longitudinal nature of the data to control for unobservable skills. The extent of over-education may indicate the need for a reformation of tertiary education in the United Kingdom. The number of places offered by UK universities is determined by the government, which also funds, almost
2 510 ECONOMICA [AUGUST exclusively, higher education as far as teaching is concerned. 4 Hence overeducation can be seen as a waste of public resources. On the other hand, if over-education stems from the heterogeneity of graduates and=or the lack of skills of some of them, it may be possible to argue for more funding or reforms of the higher education system. This debate can be seen in the light of one objective of the present government, namely to bring 50% of a cohort to higher education. For the empirical analysis, I use a survey of UK graduates conducted in 1996 on a population of graduates from two cohorts (1985 and 1990) described in Section II. Section I presents evidence on the use of satisfaction with education and job as a measure of over-education and compares the match definition of over-education with other measures. Two-thirds of the previously over-educated workers are found to be apparently over-educated. Following the previous literature, the empirical work concentrates on the determinants of over-education (Section III) and its impact on wages (Section IV). It has been consistently found that over-educated workers suffer from a pay penalty compared with workers with the same education in a matched job. (See Groot and Maassen van den Brink 2000 for a meta-analysis.) Using previous definitions of over-education, I estimate a 14% pay penalty, which is consistent with previous estimates for UK graduates (Battu et al. 1999; Dolton and Vignoles 2000). With the alternative measure, the wage gap reaches 26% for the genuinely over-educated but only 8% for the apparently over-educated workers. Over-educated graduates have lower education credentials, and genuinely over-educated workers differ from other graduates by their unobservable skills. Section V concludes with some policy recommendations. I. DEFINITIONS OF OVER-EDUCATION In the 1970s, a surge in the number of graduates in the United States triggered the first research on the demand for graduates in the labour force (Berg 1970; Freeman 1976). Freeman concluded that, as the excess qualified workforce has to settle for jobs that do not require a degree, the returns to education should plummet. Lower returns should reduce the investment in higher education, and the labour market should then return to an equilibrium point. Freeman s prediction never materialized, as returns to education remained high; however, college participation dropped in the 1970s (Card and Lemieux 2000). Similarly, in the United Kingdom, despite the recent evidence that between 29% and 47% of the workforce is over-educated (Green et al. 1999), returns to education remained stable between 1978 and 1996 (Chevalier and Walker 2001), implying that the demand for skills kept up with the supply. In Freeman s model of supply and demand, over-education is due only to a temporary disequilibrium; but Dolton and Vignoles (1997), analysing the early careers of a cohort of 1980 UK graduates, found that 62% of male graduates who were over-educated in their first job remained in a sub-graduate position six years after graduation. Sloane et al. (1999) report no evidence that for overeducated workers the quality of the match improves with a change of employer. On the contrary, over-educated workers tend to be more prone to dismissal and to have more unemployment spells. As the effect of overeducation on the individual tends to be permanent, it may be difficult to argue
3 2003] MEASURING OVER-EDUCATION 511 that over-education is due to a disequilibrium. These observed facts are not consistent with Freeman s assumption, but constant returns and overeducation can be reconciled by relaxing the hypothesis that the graduate population is homogeneous in its skill endowment. A large heterogeneity in the skills of graduates is in accordance with the persistence of over-education and with the non-promotion of over-educated workers. (Over-educated workers are trapped at the end of the job-search queue.) Most of the literature has ignored the issue of educational heterogeneity by defining over-education as departing from a norm, 5 and assuming the homogeneity in skills of all workers with an identical qualification, therefore overestimating the extent of over-education. Some definitions of over-education also implicitly imply that jobs are homogeneous in their skill requirement. Empirical work so far has relied on three main definitions of education required for the job. First, a job analyst definition of the skill=educational requirement for each occupation, as available in the United States (Dictionary of Titles), the Netherlands and Portugal is used. 6 Second, a measure of a worker s self-assessment of educational requirement is available in some surveys (e.g. Green et al. 1999). 7 Third, the distribution of education is calculated for each occupation; employees who depart from the mean (Verdugo and Verdugo 1988) or mode (Mendes de Oliviera et al. 2000) by more than some ad hoc value (generally, one standard deviation) are classified as over-educated. All of the above definitions of over-education suffer from major drawbacks. The job analyst definition is objective but is based on the assumption that all jobs with the same title have the same educational requirement. Also, one should keep in mind that a Dictionary of Titles is lengthy to compile, so the information collected might not be up to date by the time of release, especially in a rapidly changing work environment. Using selfassessment to define the job s educational requirements adjusts the measure of over-education to the specific skills needed for the job; however, it can be affected by classification error, as the researcher may not know how the worker s judgment was made. 8 The statistical definition of over-education has the advantage of always being available. However, it is probably the least desirable. As it is based on the observed distribution of education for a given occupation, it is sensitive to cohort effects, especially in the case of a rapid change in the educational level required for a given occupation. Also, the measure is sensitive to the level of aggregation necessary to obtain a reliable distribution of education, and, as the expert definition, it assumes that all jobs with the same title have identical skill requirements. Finally, when based on the mean of the educational distribution, this measure defines over-education as belonging to the upper tail of the education distribution. Defining overeducation as departing by more than one standard deviation from the mean results in similar proportions of over and under-educated workers being found around 16% of the population if education is measured in years and if the distribution of education per occupation is normally distributed (Hartog 2000). 9 It is worth noting that the choice of definition has a large effect on the incidence of over-education found but not on the pay penalty associated with over-education (Groot and Maasen van den Brink 2000). Finally, the main assumption of these measures of over-education is that all individuals with a
4 512 ECONOMICA [AUGUST given education level are perfect substitutes (Halaby 1994). I reject this assumption of homogeneity and propose an alternative definition of overeducation. Let us assume that there exist: (i) two types of graduates, clever (g) and under-achiever (u), and (ii) three types of job differing by their skill requirements: graduate (G), non-graduate jobs with intermediate skill level (upgraded job, U) and non-graduate job with low skill level (L). As in a matching model, graduates and jobs are ranked by quality. The possible outcomes are as follows: Skilled graduate Less skilled graduate Graduate job Perfect match (Gg) X Upgraded job Genuine over-education Apparent over-education (Ug) (Uu) Non-graduate job X Genuine over-education (Lu) Skilled graduates compete for graduate jobs and most of them obtain a graduate job: this is a perfect match (Gg). However, skilled graduates at the end of the queue might be offered only upgraded jobs (Ug) and hence would be genuinely over-educated. 10 Less skilled graduates are not offered graduate jobs, and they compete for upgraded jobs (Uu). In most of the literature this type of match is defined as a mismatch, as the worker appears to be over-educated for the position. However, owing to the lower standard of educational skills of u- graduates, we argue that this type of match can be considered appropriate. The less able from that group can only get a non-graduate job (Uu) and are genuinely over-educated. In this framework, skilled graduates in an upgraded job and less skilled graduates in a non-graduate job feel that their skills are under-used, and they are defined as genuinely over-educated, whereas less skilled graduates in an upgraded job are only apparently over-educated, as their skills match the requirement of the job. In time, genuinely over-educated graduates might be able to move into a perfect match job if they are a g-type or into an upgraded job if they are a u-type. However, apparently overeducated graduates will not be able to move to a perfect match position as they lack some essential graduate skills. Therefore, a bulk of the graduate population will remain over-educated, as observed by Dolton and Vignoles (1997). To reflect this dichotomy of the over-educated population, an alternative measure of over-education based on occupation and a measure of the job satisfaction is proposed. Using the Standard Occupation Code (2-digit), the following occupations are defined as graduate jobs: Occupations which require degrees [are] managers and administrators, professionals, plus from the associate professional category, computer analysts (Alpin et al. 1998). (See Table A1 for details.) This basic separation between graduate and nongraduate jobs is roughly equivalent to a Dictionary of Titles definition of overeducation. For graduates in a non-graduate job, the answer to the following question is used to classify them as genuinely or apparently over-educated: How dis=satisfied are you with the match between your work and your qualifications? The possible answers to this question are grouped into six
5 2003] MEASURING OVER-EDUCATION 513 categories, ranging from very dissatisfied to very satisfied. The very dissatisfied and dissatisfied answers are grouped to generate a dichotomous variable. 11 Over-educated workers who are satisfied with the match between their education and their work are defined as apparently over-educated, whereas those who are dissatisfied are genuinely over-educated. The genuinely overeducated population is composed of two groups that differ by their ability: clever graduates in upgraded jobs and under-achievers in low-skill jobs. The next section will show that less able students form the bulk of this category. II. DATA To conduct this study, a sample of two cohorts of UK graduates was used. The data were collected by a postal survey organized by the University of Birmingham in the winter of 1996 among graduates from 30 higher education institutions covering the range of UK institutions (see Belfield et al for details). Graduates from the 1985 and 1990 cohorts were selected, leading to a sample of 15,000 individuals. 12 Graduates from the Open University and the University of Buckingham were dropped, because of the different characteristics. 13 Furthermore, only first-degree graduates who were younger than 25 on graduation, were full-time employees in 1996, reporting their yearly wage, weekly hours of work and occupation code, living in the UK and without health problems were retained, resulting in a sample of 4844 observations. 14 The questionnaire covers a wide range of topics, including schooling, academic information, family background and employment history. Of particular use for this paper is the section on the satisfaction with the match between education and employment. Additionally, the survey has a longitudinal component, as respondents were asked about their employment situation at three points in time: one year after graduation, six years after and, for the older cohort, 11 years after. The survey allows us to compare various definitions of over-education. First, in the spirit of the Dictionary of Titles, graduate occupations are defined on the basis of occupational code (see Table A1). Additionally, the survey included the following question: Was the degree gained in 1985=1990 a requirement in the job specification for your main employment? This selfassessment measure was used by Battu et al. (2000). The measure of overeducation that I propose is a combination of the job title and the satisfaction regarding the match between education and job. The introduction of the satisfaction element allows us to introduce heterogeneity in graduates and jobs; graduates in non-graduate job who are satisfied with the match are defined as apparently over-educated. One limit of this definition is that it stipulates that graduates who, according to their occupation title, are in a graduate job cannot be dissatisfied with the match between their education and job. Figure 1 plots the distribution of the match satisfaction by occupational group (1-digit level), where satisfaction is measured on a 6-point scale, ranging from very dissatisfied (1) to very satisfied (6). About 15% of managers and professionals are dissatisfied with the match. This proportion is similar for technicians but much higher for all other occupation groups. The most dissatisfied are graduates in clerical occupations, over two-thirds of whom reported being in satisfaction category 1 or 2. By stipulating that graduates in a
6 514 ECONOMICA [AUGUST 50 % Manag. Prof. Techn. Clerc. Manuf. & serv. Sales Other FIGURE 1. Satisfaction match education=job by occupation. graduate job cannot be over-educated even if they are dissatisfied, we may underestimate the extent of over-education. On the other hand, it could be that this dissatisfaction reflects a mismatch of skills, for example an art student working as a financial expert, rather than a lack of skills. Table 1 compares the different definitions of over-education. Only 74% of graduates whose occupation title defined them as being in a graduate job claimed that a degree was a requirement to obtain the job. This illustrates the difficulty of defining a graduate occupation. The occupation title may not be perfect for defining a graduate job, but nor is the requirement definition. A TABLE 1 DEGREE REQUIREMENT AND OVER-EDUCATION Requirement Matched graduate Apparent over-edu. Genuine over-edu. Total Missing % 0.83% 0.71% 0.85% Not know % 0.62% 1.06% 0.54% A degree % 49.69% 15.60% 68.15% A diploma % 2.28% 1.42% 2.13% Not a requirement % 46.58% 81.21% 28.34% Total % % % % Note: For each entry, the second row represents the percentage of the column population.
7 2003] MEASURING OVER-EDUCATION 515 degree may not always be a formal requirement, but it may be implicit, especially when internally promoted. Battu et al. (2000) note that as experience is acquired a degree becomes increasingly less of a requirement, which suggests that, when one has held a position for which a degree was a requirement, one s subsequent posts implicitly assume that such a degree is held. Reliance on the requirement question may introduce a large measurement error in overeducation. Over-education ranges from 19% with the occupation definition to 32% with the requirement definition, which is line with the UK literature. Relying on the occupational title, we now split the over-educated population according to their degree of satisfaction between their education and their job. Nearly two-thirds of graduates in a non-graduate job are not too dissatisfied with the match and are defined as apparently over-educated. The heterogeneity of jobs is plain in Table 1: 50% of apparently over-educated graduates were in jobs requiring a degree, but this proportion is only 16% for the genuinely over-educated workers. Evidence on the heterogeneity of the graduate population is provided in Figure 2 and Table 2. Figure 2 pictures the distribution of answers to the following question: On reflection and in general, in what ways has your degree contributed to your getting an interesting job? It is expected that graduates in an upgraded position are reasonably positive about the contribution of their degree, whereas graduates in a low-skill job are much more critical. The distribution of answers for matched and apparently over-educated graduates is similar. The mode answer for these two categories is 5, reflecting a high satisfaction with the effect of the degree on getting an interesting job. Apparently over-educated workers are slightly less satisfied than matched workers, which would confirm that they are in jobs requiring some graduate skills. On the other hand, genuinely over-educated workers are highly dissatisfied with the effect of their degree more than 50% felt that their degree did not contribute to their getting an interesting job. Genuinely and apparently over-educated graduates are typically in jobs that have different requirements regarding the use of graduate skills % Do not know 1 Not at all A lot Matched graduate Apparent over-edu. Genuine over-edu. FIGURE 2. Contribution of your degree in getting an interesting job.
8 516 ECONOMICA [AUGUST More evidence on the differences in the characteristics of the three groups of graduates is available from Table 2, which reports the mean characteristics for the three groups of graduates defined. First, I concentrate on the differences in educational attainment, since I have posited that these differences lead to over-education. Matched graduates have better academic credentials than other graduates. Additionally, apparently over-educated workers have better education than genuinely over-educated graduates. The average A-level score of matched graduates is one and two points higher than for, respectively, apparently and genuinely over-educated workers. Similarly, 10 percentage points in the proportion having attended a university rather than a polytechnic separate each category of graduates. The proportion of matched graduates graduating with first-class honours is 50% and more than three TABLE 2 MEAN AND STANDARD DEVIATION OF SELECTED VARIABLES Matched graduate Apparent overedu. Genuine over-edu. Variable Mean Std dev. Mean Std dev. Mean Std dev. Personal characteristics Cohort Male White Educational characteristics A-level score A-level missing University Degree: first Degree 2: Medical Biology Agriculture Physics Mathematics Engineering Architecture Administration Languages Humanities Education Subject missing PhD Master Professional qual PGCE Employment characteristics Month empl Month unempl Size 100/ Size 500þ Observations
9 2003] MEASURING OVER-EDUCATION 517 times higher than for, respectively, apparently and genuinely over-educated employees. Within the over-educated population, apparently over-educated workers, have better credentials than genuinely over-educated workers, which suggests that the latter group is composed mostly of under-achievers in a lowskill position (Lu), rather than skilled graduates in an upgraded job (Ug). The skill differential observed between the various groups of graduates confirms that over-education originates not from disequilibria in the market for graduates, but from the lack of skills acquired by graduates at university. The remainder of Table 2 shows the expected results. Over-education is more likely to affect graduates from the younger cohort than the older one. Three hypotheses could explain this difference. First, older workers would have had more time to assess the labour market and to acquire, through on-the-jobtraining, some of the skills they were originally lacking. Second, workers might have revised their career expectations with time and accepted their situation as inevitable. Third, it may be that graduates from the younger cohort were less likely to acquire graduate skills while studying because of overcrowding or changes in the curriculum. Unfortunately, these conflicting hypotheses cannot be tested with the data. Finally, there is a slight gender differential, with women more likely than men to accept non-graduate jobs. Married women might be more constrained in their job search by family preferences and hence are more likely to be over-educated (see Frank 1978, and Battu et al for empirical evidence concerning the UK), or employers may discriminate against women. It is also worth noting that the labour market experience of the three groups of graduates is substantially different. Matched graduates have up to a full year s employment more and have experienced a third of the unemployment of genuinely over-educated workers, but this could be due to cohort effects. Educational differential III. DETERMINANTS OF OVER-EDUCATION I have hypothesized that over-education stems from heterogeneity in the skills of graduates. This hypothesis is tested by investigating the determinants of over-education in a simple model. Let us assume that over-education is determined by the following latent model, where X is a vector of educational achievement and is a normally distributed term reflecting the unobservable component of over-education. 15 To clarify the notation, I do not include a subscript for the individual: (1) OE* ¼ X þ This latent model is not observed. Instead, we have an ordinal variable: 8 1 if OE* < >< 1 (2) OE ¼ 2 if 1 OE* < 2 >: 3 otherwise: Thus, the determinants of over-education are estimated by using a multinomial logit, where the omitted category (OE ¼ 1) is being in a graduate job, while 2 and 3 represent, respectively, apparent and genuine over-education.
10 518 ECONOMICA [AUGUST The first two columns of Table 3 report the estimated marginal effects for the most parsimonious model, including personal characteristics and educational achievement measures. Graduates from cohort 1990 are 2.5 (3.1) percentage points more likely to be apparently (genuinely) over-educated than those from cohort White graduates are also at higher risk than non-white of being apparently over-educated but not genuinely over-educated. The evidence on the skill differentials as a determinant of over-education is mixed. In contradiction with Robst (1995), I did not find institutional effects, and graduates from former polytechnics are as likely as their peers from more traditional universities to obtain a graduate job. 16 While A-level scores and university achievement significantly affect the probability of being overeducated, the differences between apparent and genuine over-education are never significant (in contradiction with summary statistics). Having an A-level score between 11 and 15 compared with 0 5 decreases the probability of being overeducated by 2 4 percentage points. Going back to the model presented in Section I, these findings suggests that, while the selection between a matched job and a job for which a graduate is over-educated is clearly based on educational skills, the selection between apparent and genuine over-education is due to luck, as in a matching model, or to non-academic, possibly nonobservable, skill differential. Previously, I suggested that over-education may be more=less prominent for graduates of given subjects. The third and fourth columns of Table 3 report the estimated marginal effects for a model of over-education, including 12 dummies for subject of graduation. Math graduates are seven (five) percentage points less likely to be apparently (genuinely) over-educated than graduates from economics. The inclusion of subject dummies has no effect on our previous results. Subjects in high demand, such as medical science, mathematics, education and to a lesser extent engineering, are a safeguard against over-education, whereas students from biology, agriculture, languages and humanities are more at risk than economists of being over-educated. Once again, the differences between apparent and genuine over-education are not significant. In a third model, I included subsequent qualifications as well. The inclusion of these variables did not change the previous results. More qualification reduces the likelihood of over-education. Vocational qualifications and to a lesser extent a PhD are more effective at protecting workers against overeducation than academic qualifications. Again, this is in line with Mason s (1999) description of the UK labour market for graduates, where employers main concern is the employability of the new graduates. Having a vocational qualification appears to insure against over-education; on the other hand, it could be that employees in graduate jobs had the opportunity of gaining these qualifications whereas those who started as over-educated did not. 17 Including unobserved skills I tested the hypothesis that over-educated graduates differ in their level of unobservable skills by relying on the longitudinal element of the data. I did not use panel data analysis, as in Bauer (2002), for two reasons. First, the present data are a cross-section with some recall variables; thus, for most variables we
11 TABLE 3 MULTI LOGIT: DETERMINANTS OF OVER-EDUCATION (MARGINAL EFFECTS) Model 1 Model 2 Model 3 Apparent Genuine Apparent Genuine Apparent Genuine Cohort (0.008) (0.006) (0.007) (0.005) (0.006) (0.004) Male (0.010) (0.005) (0.009) (0.005) (0.007) (0.004) White (0.018) (0.016) (0.018) (0.011) (0.015) (0.008) A-level score (0.014) (0.010) (0.013) (0.009) (0.011) (0.007) A-level score (0.015) (0.014) (0.014) (0.014) (0.011) (0.011) University (0.011) (0.010) (0.010) (0.008) (0.008) (0.007) Degree: first (0.016) (0.018) (0.015) (0.015) (0.013) (0.012) Degree: 2: (0.008) (0.005) (0.007) (0.004) (0.007) (0.003) Medical (0.036) (0.028) (0.029) (0.024) Biology (0.018) (0.010) (0.014) (0.007) Agriculture (0.025) (0.015) (0.021) (0.012) Physics (0.014) (0.009) (0.011) (0.007) Mathematics (0.027) (0.015) (0.023) (0.011) Engineering (0.016) (0.013) (0.013) (0.010) Architecture (0.023) (0.015) (0.018) (0.012) Administration (0.023) (0.010) (0.019) (0.008) Languages (0.019) (0.006) (0.015) (0.005) Humanities (0.020) (0.006) (0.016) (0.005) Education (0.042) (0.012) (0.035) (0.011) PhD (0.021) (0.020) Other post-grad (0.013) (0.007) Professional qual (0.010) (0.007) PGCE (0.026) (0.015) Log likelihood Observations 4844 Note: Hubert White standard error and cluster analysis (by type of HEI). Emboldened statistics are significant at the 5% level. The omitted group is matched graduate. The default person is a female 1985 graduate from a former polytechnique institution who graduated with a 2:2 or lower in economics. A dummy for subject of degree missing was included but never found significant. A specification with regional dummies was also tested leading to quantitatively similar results.
12 520 ECONOMICA [AUGUST cannot observe variations in time. Second, I am not convinced that panel data allow one to purge the estimate of over-education of bias, since the identification relies on individuals that change category from one period to the next. The estimates would be unbiased if one assumed the exogeneity of the transition, a hypothesis that I do not support. Dolton and Vignoles (1997) have shown that graduates who made the transition seemed to catch up rapidly, so that soon no wage differential was observed between them and graduates who had always been matched. Thus, it is not surprising that Bauer (2002), using a panel of German workers, found that the negative effect of over-education on wages disappears, since the effect of over-education on pay is identified by workers making the transition to graduate job. In order to proxy the unobservable skills, I estimated earnings in the first job and assumed that the deviation between the expected and observed earnings is a proxy for the unobservable idiosyncratic characteristics affecting workers productivity. 18 This measure of individual characteristics is then introduced as an exogenous determinant of the current over-education status. I therefore assumed that unobservable skills remained constant over the six-year period. Formally, the model estimated has the following form: (3) ln(w 1 ) ¼ X1 X 1 þ S1 S 1 þ " 1 ; OE* ¼ X 2 þ ^" 1 þ : I restricted the sample to graduates from the 1990 cohort to avoid attributing cohort-specific effects to the idiosyncratic component of the residual. I also lost 850 observations for which no information about wages in 1991 was available. I neglected the effect of sample selection in the remainder of the analysis. Table 4 reports the transition from first to current over-education. By 1996, 50% of graduates who were over-educated in their first job had made the transition to a graduate job; 31% were apparently over-educated and 19% remained genuinely over-educated. 19 These figures are in line with the Dolton Vignoles (1997) results on a population of 1980 graduates (38% making the transition within six years). Note that less than 4% of graduates who were matched in their first job were over-educated six years later. Annual earnings in 1991 are grouped into 16 categories. As hours worked are not reported, the log annual earnings are estimated with a dummy for TABLE 4 TRANSITION WITHIN THE FIRST SIX YEARS OF GRADUATION (COHORT 90) Firstncurrent Matched Apparent over-ed. Genuine over-ed. Total Not a graduate job % 31.00% 18.86% Graduate job % 2.22% 1.44% Total % 11.2% 6.91% Note: For each entry, the second row represents the percentage of the row population.
13 2003] MEASURING OVER-EDUCATION Fraction zres91 5 (a) Matched graduates Fraction zres91 5 (b) Apparent over-education Fraction zres91 (c) Genuine over-education FIGURE 3. Distribution of normalized residuals by over-educated group.
14 522 ECONOMICA [AUGUST TABLE 5 DETERMINANTS OF OVER-EDUCATION, INCLUDING UNOBSERVABLE SKILL No unobserved skills Unobserved skills Apparent Genuine Apparent Genuine Male (0.001) (0.001) (0.001) (0.001) White (0.004) (0.001) (0.005) (0.001) A-level score (0.002) (0.001) (0.002) (0.001) A-level score (0.002) (0.001) (0.003) (0.001) University (0.001) (0.001) (0.002) (0.001) Degree: first (0.003) (0.002) (0.003) (0.002) Degree: 2: (0.001) (0.001) (0.001) (0.001) Medical (0.005) (0.014) (0.006) (0.015) Biology (0.002) (0.001) (0.003) (0.001) Agriculture (0.003) (0.001) (0.004) (0.001) Physics (0.003) (0.001) (0.003) (0.001) Mathematics (0.004) (0.002) (0.005) (0.002) Engineering (0.003) (0.001) (0.003) (0.001) Architecture (0.003) (0.001) (0.004) (0.002) Administration (0.003) (0.001) (0.004) (0.002) Languages (0.003) (0.001) (0.004) (0.001) Humanities (0.003) (0.001) (0.004) (0.001) Education (0.006) (0.002) (0.007) (0.002) PhD (0.026) (0.013) (0.029) (0.014) Other post-grad (0.003) (0.001) (0.003) (0.001) Professional qual (0.002) (0.001) (0.002) (0.001) PGCE (0.027) (0.002) (0.029) (0.002) Unob. skill (0.0008) (0.0004) Observations 2229 Log likelihood
15 2003] MEASURING OVER-EDUCATION 523 full-time employment among other exogenous variables covering the human capital and job characteristics. 20 This specification explains 29% of the variation in pay in the first job. The main determinants are job-specific, as the sample is fairly homogeneous in educational attainment. The residuals from this equation are used to calculate a z-score which I assume proxies the unobservable skills of individuals in their first job; the z-score is a measure of the skills differential to the average skills for individuals with the same observable characteristics. Figure 3 reproduces the distribution of the unobservable skills for the three groups of graduates. The distributions are fairly similar for currently matched and currently apparently over-educated graduates, while genuine over-educated graduates have lower unobservable skills. Table 5 reports the estimates of the determinants of over-education for the restricted sample of graduates. In the first column, I use the same specification as in model 3 of Table 3. The significant determinants are similar for both populations, but on the restricted sample the various effects are somehow reduced. I introduce the measure of unobservable skills. Its effect is small but significant; graduates with a higher score are less likely to be genuinely overeducated, but, as reported in Figure 4, there is no difference between matched and apparently over-educated graduates. The selection into the different types of job appears to be based on both educational achievement and unobservable skills. Graduates with better educational achievement obtain matched jobs; for the less talented graduates, the selection between upgraded and non-graduate jobs is based on their unobservable skills. This two-selection process confirms the theoretical model, that the graduate population is not homogeneous in skills % Pay/hr ( ) Matched graduate Apparent over-edu. Genuine over-edu. FIGURE 4. Distribution of pay per hour by over-education group.
16 524 ECONOMICA [AUGUST IV. OVER-EDUCATION AND WAGES The bulk of the literature has reported that over-education is associated with a pay penalty. (See Groot and Maasen van den Brink 2000 for a meta analysis. 21 ) For UK graduates, Dolton and Vignoles (2000) or Battu et al. (1999) estimate that the pay penalty for over-education ranges from 15% to 20%. In this dataset, the annual pay is reported in 16 categories. An hourly wage is constructed and its distribution for the three groups of graduates is reported in Figure Substantial pay differentials are observed between the three groups of graduates. The distributions of earnings for matched and apparently overeducated workers are roughly similar: the two distributions have identical mode, but for matched graduates the upper tail is heavier while for the apparently over-educated the lower end of the distribution has a higher density; the difference between the two distributions is the largest for hourly wages of between 6 and 7. On the other hand, the distribution for genuinely over-educated workers lies to the left of the previous two. Graduates in a graduate occupation earn a median pay of per hour. The pay gap reaches 10% and 33% for, respectively, apparent and genuine over-educated workers. This large pay differential confirms that genuinely over-educated graduates are likely to be of the less-skilled type, settling for jobs that have not been upgraded. The determinants of log hourly pay are estimated with the following equation, where O is the vector of dummies defining over-education; estimates of this model are presented in Table (4) ln(w) ¼ X X þ S S þ O O þ ": The results are as expected: males, and graduates with more labour market experience and a professional qualification earn more, while a teaching qualification (PGCE) is associated with a pay cut of 12%. Holding experience constant, graduates from the 1990 cohort earn more than those who graduated five years earlier. Graduates of higher educational ability also earn more, but the type of institution where education was achieved does not matter. An individual with an A-level score greater than 10 and a first earns 14.5% more than his peer who scored less than 5 and graduated with a 2:2 or below. In order to compare the present results with previous studies on the effect of overeducation on earnings, a single dummy for not being in a graduate job is included (model 1). A pay penalty of 14% is estimated, in line with previous estimates. However, since the over-educated population is heterogeneous, this is a misleading estimate. Model 2, therefore, separates the apparently overeducated and the genuinely over-educated. The pay penalty compared with matched graduates is estimated at 8% and 26% for apparently and genuinely over-educated graduates, respectively. In models 3 and 4, occupation dummies and a dummy for holding a non-permanent job are included. The pay penalty coefficients are reduced, sensibly, by the introduction of these variables, but the results remain broadly similar. The large difference observed in pay between the two groups of over-educated workers reinforces the view that the overeducated worker group cannot be considered homogeneous. The previous estimates of the pay differential between graduates are based on the underlying assumption that all graduates are similar in their skills, where skills include motivation and other unobservable characteristics
17 Table 6 OLS IN PAY PER HOUR: ALL GRADUATES No unobservable skills Restricted sample Model 1 Model 2 Model 3 Model 4 No skill Skill Non-grad. job (0.017) (0.015) Apparent over-ed (0.014) (0.027) (0.016) (0.019) Genuine over-ed (0.025) (0.038) (0.022) (0.028) Cohort (0.026) (0.027) (0.025) (0.025) Male (0.011) (0.011) (0.010) (0.010) (0.013) (0.013) White (0.017) (0.017) (0.016) (0.017) (0.026) (0.027) A-level score (0.010) (0.012) (0.011) (0.012) (0.012) (0.013) A-level score (0.015) (0.016) (0.016) (0.017) (0.021) (0.020) University (0.017) (0.017) (0.017) (0.017) (0.018) (0.017) Degree: first (0.017) (0.018) (0.017) (0.017) (0.032) (0.033) Degree: (0.009) (0.009) (0.008) (0.008) (0.010) (0.011) PhD (0.034) (0.036) (0.032) (0.034) (0.100) (0.097) Master s degree (0.016) (0.016) (0.017) (0.016) (0.019) (0.020) Professional qual (0.012) (0.011) (0.011) (0.011) (0.014) (0.014) PGCE (0.021) (0.020) (0.021) (0.020) (0.032) (0.034) Months employed (0.001) (0.001) (0.001) (0.001) (0.003) (0.003) Employment 2 / (0.001) (0.001) (0.001) (0.001) (0.003) (0.003) Unemployment (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) Size (0.013) (0.014) (0.014) (0.014) (0.020) (0.019) Size 500þ (0.010) (0.011) (0.012) (0.011) (0.017) (0.018) Unobs. skill (0.008) Constant (0.055) (0.054) (0.051) (0.054) (0.094) (0.086) Subject dummies F ¼ 26:06 F ¼ 16:70 F ¼ 18:23 F ¼ 17:46 F ¼ 36:16 F ¼ 33:42 Region dummies F ¼ 84:29 F ¼ 83:51 F ¼ 87:43 F ¼ 86:39 F ¼ 16:73 F ¼ 16:86 Occupation dummies F ¼ 31:53 F ¼ 17:95 Contract type F ¼ 24:38 F ¼ 23:92 Observations R-squared Note: Hubert White standard error and cluster analysis (by type of HEI). Emboldened statistics are significant at the 5% level. When skills are included, the standard errors are obtained by bootstrap (500 reps.)
18 526 ECONOMICA [AUGUST affecting productivity. Assuming that over-educated workers are somehow less skilled than matched workers, the estimated pay differential for being overeducated is biased upwards, as it includes returns to skills that are specific to the better group of graduates. Formally, " in (4) partly measures the endowment in unobservable (to the econometrician) skills. Since skills and over-education are correlated, the estimates of o is equal to (5) ^ o ¼ o þ cov(o;") Var(O) : By including a measure of unobservable skills, this bias is reduced, as the overeducation dummy (0) and " 2 are now independent. Formally, the estimated model as the following form: (6) ln(w 1 ) ¼ X1 X 1 þ S1 S 1 þ " 1 ; ln(w 2 ) ¼ X2 X 2 þ S2 S 2 þ O O þ ^" 1 þ " 2 : The last two columns of Table 6 report the determinant of pay for the subsample of 1990 graduates who were working in For this younger cohort, I did not observe a gender pay gap, and the effect of work experienced is wrongly signed. These findings are not surprising, since for young graduates no gender pay gap is expected, and differences in past employment capture differences in educational decisions after the degree and job search strategy. The effect of over-education on this population is also slightly reduced, which indicates that the impact of over-education on older workers is probably higher than on younger (Dolton and Silles 2001). The inclusion of unobservable skills does not alter these conclusions. Increasing unobservable skills by a standard deviation increases pay by 7%. The effect of over-education on wage is reduced by about 8% when the measure of skills is included, 24 to 4.8% and 21.6%, respectively, for the apparently and genuinely over-educated. These estimates may be considered as lower bounds on the effect of overeducation on pay, since, as in most of the literature, I have assumed that overeducation is an exogenous variable. This is obviously a stringent assumption, especially since I have hinted that genuinely and apparently over-educated graduates may differ in their unobservable skills. One exception in the literature that relaxes this assumption is Dolton and Silles (2001). Using a survey conducted among alumni from Newcastle University, these authors estimate that the impact of over-education on the first and current job wage is 18% and 30%, respectively. Instrumental variables or accounting for selection lead to estimates that do not differ statistically from those obtained by OLS, mostly because of their lack of precision. The difficulty is to find an instrument that affects over-education but not wages directly. Dolton and Silles rely on mobility between region of education and region of employment, which may be a weak instrument. Taking our estimates as a lower bound is nevertheless informative. Returns to a degree, compared with A-levels, are typically around 30% in the United Kingdom (Chevalier and Walker 2001); genuinely overeducated graduates then suffer from a pay penalty compared with other graduates, which substantially reduce the returns from their investment. This is an important consideration to keep in mind as the debate on the financing of
19 2003] MEASURING OVER-EDUCATION 527 universities and the increase of university fees has recently surged (Greenaway and Haynes 2000). V. CONCLUSION Previous work on over-education has largely neglected the heterogeneity of the graduate population and jobs, which substantially biases the effects of over-education on pay. The group of graduates traditionally defined as overeducated can be divided between the apparently and the genuinely overeducated. The apparently over-educated are paid between 5% and 10% less than well-matched graduates, while genuinely over-educated graduates suffer from a pay penalty ranging from 22% to 26% compared with matched graduates. The sorting between the groups of graduates appears to be due to two phenomena. Educational achievement is used to separate good graduates and under-achievers; good graduates obtain graduate jobs, while underachievers compete for upgraded and non-graduate jobs. The second selection is based on unobservable characteristics of graduates. The unobservable skills of genuinely over-educated graduates are substantially worse than those of the two other types of graduate. As a large number of graduates do not seem to be able to acquire graduate skills while at university, it is worth pondering whether mis-skilled graduates best satisfy employers needs for intermediate skills jobs. The mis-qualification of the workforce is costly for society and for individuals. In the absence of alternative vocational qualifications, a degree can be viewed as a signalling device (Spence 1973); individuals with subgraduate qualities may make a rational choice by going to university in order to reveal their characteristics and obtain an upgraded job. This is possible because society but not the individual bears the cost of providing the signal. The provision of less academic tertiary education qualifications could lead to a more cost-efficient sorting. Whether over-education has increased over time is debatable, but it is possible that the large increase in the intake of students that took place since the mid-1980s in the United Kingdom, associated with a reduction in the cost per student, has led to less personal attention being given to students, and contributed to the increase in heterogeneity in the skills of graduates. Additionally, the irrelevance of the university curriculum vis a` vis the job market in some subjects might also lead to some over-education; I noted large variations by subject in the probability of over-education. Thus, an increase or reallocation of money within higher education may reduce over-education. Finally, a substantial proportion of the surveyed population reported being dissatisfied with their education job match despite being in a graduate job. The meaning of this dissatisfaction and its effect on the productivity and earnings of matched graduates would be worth exploring. 25 APPENDIX Table A1 defines graduate occupations based on occupational code. Table A2 gives details on the creation of the sample.
20 528 ECONOMICA [AUGUST TABLE A1 GRADUATE OCCUPATIONS Managers and administrator 10 General managers/administrators of national/local governments, large companies or organizations 11 Production managers in manufacturing, construction, mining and energy industries 12 Specialist managers 13 Financial institution and office managers, civil service executive officers 14 Managers in transport and storage 15 Protective service officers 16 Managers in farming, horticulture, forestry and fishing 17 Managers and proprietors in service industries 19 Managers and administrators others Professional occupations 20 Natural scientists 21 Engineers and technologists 22 Health professionals 23 Teaching professionals 24 Legal professionals 25 Business and financial professionals 26 Architects, town planners and surveyors 27 Librarians and related professionals 29 Other professionals occupations Associate professional and technical occupations 32 Computer analyst/programmers TABLE A2 THE CAREERS OF HIGHLY QUALIFIED WORKERS All respondents (Diploma, first degree, post-graduate) 15,530 First degree only 9,534 Age on graduation ,322 Complete career history 7,255 Full-time work in ,220 Hours ( > 30), pay and occupation in ,681 Work in the UK 5,427 Employee 5,005 Disable 4,883 Satisfaction match education/job 4,844 ACKNOWLEDGMENTS Financial support from the European Commission under the TSER programme PL for the PURE project is gratefully acknowledged. This paper is a substantially revised version of Chapter 4 of my PhD thesis (University of Birmingham, 2000). I thank Peter Dolton, Colm Harmon, John Heywood, Gauthier Lanot, Stan Siebert, Hilary Steedman, Tarja Viitanen, Ian Walker and participants at the Keele University Economics Workshop, the EEEG Education Workshop (2000), the IZA Labor Summer School (2000)