Over-education: Is it voluntary for some individuals?

Transcription

1 Over-education: Is it voluntary for some individuals? David J. Black Melbourne Institute of Applied Economic and Social Research, The University of Melbourne January 2012 Abstract Over-education research considers the returns to investments in education. But, to date, the literature has examined only one such return: wages. Based on estimated wage penalties, it then infers instances of over-education represent labour market failures. However, given a broader assessment of returns, this may not be true. This paper recognises that many job attributes can affect individuals utility levels (e.g., hours, job security and required effort) and so some over-educated individuals may have actually obtained jobs that maximise their (expected) utility and, therefore, achieved their preferred outcome. Such voluntary over-education is estimated. Then, to validate the resultant estimates, the hypothesis that the voluntarily over-educated, but not involuntarily over-educated, trade wages for improvements in other job attributes is empirically tested. To do so, linear fixed effects, fixed effects ordered logit and random effects probit estimators are used to model the relationship between voluntary over-education and job attributes. Australian panel data for period 2001 to 2008 are used, where around 19 per cent of males and 23 per cent of females are identified as over-educated. This paper finds roughly 17 per cent of over-educated males and 21 per cent of over-educated females are voluntarily over-educated. It then finds these individuals trade wages for improvements such as greater job security, working preferred hours, greater job flexibility and reduced stress. Overall, they are also more satisfied with their achieved work-life balance. JEL Classification: I21, J22, J24, J28, J31 Keywords: human capital, over-education, labour market mismatch, compensating wage differentials This paper is a preliminary draft and constitutes a part of my Ph.D thesis; please do not cite without my permission (black@unimelb.edu.au). Thanks to Jeff Borland and Mark Wooden for valuable comments and supervision. This paper uses unit record data from the Household, Income and Labour Dynamics in Australia (HILDA) Survey. The HILDA Project was initiated and is funded by the Australian Government Department of Families, Housing, Community Services and Indigenous Affairs (FaHCSIA) and is managed by the Melbourne Institute of Applied Economic and Social Research (Melbourne Institute). The findings and views reported in this paper, however, are those of the author and should not be attributed to either FaHCSIA or the Melbourne Institute.

2 1. Introduction Modern societies devote considerable resources to the education of individuals. 1 And, with increasing numbers completing secondary and tertiary qualifications and governments continuing to emphasise the importance of education, these outlays have been, and will likely continue, rising over time. Spending on education is, of course, typically justified by the argument that education represents an investment for individuals and societies. Since many of the returns to education are realised via the employment of individuals (e.g., increased earnings, higher labour force participation rates and increased worker productivity), labour markets are consequently critical for such investments. This means labour market failures can adversely affect the returns that accrue to individuals and societies. In particular, their failure to facilitate the full utilisation of the human capital individuals derive from education a phenomenon referred to as over-education (Duncan and Hoffman, 1981; Green, McIntosh and Vignoles, 1999; McGuinness, 2006) would diminish such returns. Over-educated individuals would not receive the full returns, particularly with respect to increased earnings, to their investments in education. 2 For societies, over-education would result in an under-utilisation of the human capital available in the workforce, leading to productivity levels, economic growth rates and living standards below their potential. With returns diminished, individuals and societies may have then over-invested in education (i.e., the costs of under-utilised education may exceed its benefits). 3 Over-education, therefore, would have implications for government policy and contemporary research into the economic returns to education (Card, 2001; Leuven and Oosterbeek, 2011). Over-education can only occur when jobs have productivity ceilings thresholds at which output becomes unresponsive to individuals possessing more human capital that prevent individuals from fully utilising their human capital (from education) in performing the job. Given the neoclassical view of labour markets, such inflexibility may arise in the short-run as labour markets adjust to changes in the supply or demand for labour (e.g., an increase in supply of individuals with a university degree). It may also occur if factors impede (perfect) competition in labour markets; specifically, when the conditions of perfect information and perfect factor mobility fail to hold. These market imperfections can lead to situations in which firms do not, or cannot, adjust jobs to 1 In Australia, for example, federal and state governments allocated 16 per cent of their total expenditure (approximately $71 billion) in to funding education (ABS, 2011). Meanwhile, legislation dictates individuals spend at least 11 years in schools, though the vast majority today choose to spend additional years undertaking education. 2 It is assumed human capital represents individuals productive capacity and wages reflect their labour productivity achieved in the workplace (i.e., firms pay no wage premium for unused human capital). These are standard assumptions in economics research, and the basis for human capital theory (see, for example, Becker (1964), Preston (1997), Card (1999) and Dearden, Reed and Van Reenen (2006)). 3 Over-education and over-investment in education are two separate, though related, issues: over-investment concerns comparisons of costs and benefits of education, whilst over-education merely considers the benefits. Thus, being overeducated does not necessarily imply an over-investment in education. - Page 1 -

3 fully utilise individuals human capital and, at the same time, individuals are willing to accept jobs for which they are over-educated. 4 An implication of productivity ceilings is that jobs may be regarded as requiring a certain level of human capital to be performed. Assuming levels of education (or qualifications) can be used to quantify such required human capital levels, instances of overeducation can be identified: an individual is over-educated if their education (or highest qualification) level exceeds that considered necessary to perform their job. Alternatively, an individual whose education matches their job s required human capital level is deemed well-matched, while an individual with education below the required level is under-educated. Over-education has been the subject of considerable economics research, particularly since the mid-1990s (see McGuinness (2006) and Leuven and Oosterbeek (2011) for extensive reviews). This literature mainly consists of empirical studies estimating the incidence of over-education and its effect on individuals wages. Typically, 15 to 30 per cent of employed individuals are found to be over-educated and, compared to the well-matched, they incur wage penalties ranging from 5 to 30 per cent (Groot and Maassen van den Brink, 2000; Hartog, 2000; Sloane, 2003; McGuinness, 2006). 5 While over-education research is concerned with the realised returns to investments in education, the literature has, to date, examined only one such return: wages. Based on estimated wage penalties, the literature then infers all instances of over-education represent labour market failures. However, given a broader assessment of the returns to education, this may not be true. This paper recognises that many job attributes, not just the wage, can affect individuals utility levels (e.g., work hours, job security and required effort) and so some over-educated individuals may have actually obtained jobs that maximise their (expected) utility and, therefore, achieved their preferred outcome. Such individuals would incur an over-education wage penalty because they trade wages for non-pecuniary benefits of employment. 6 Thus, the paper considers the question: are some individuals voluntarily over-educated? To examine this issue, it is first necessary to develop a method to empirically identify voluntary over-education. This paper proposes a distinction 4 Such situations may arise if, for example, there is: asymmetric information in labour markets regarding human capital (i.e., firms are unable to perfectly observe individuals human capital, while individuals are unable to perfectly observe human capital requirements of jobs); significant adjustment costs in labour markets (i.e., costs, such as the purchasing of new production technologies, that prevent firms adjusting jobs and costs associated with job search and spatial relocation that constrain individuals choices); or labour market institutions that prevent the alteration or specialisation of jobs (e.g., worker trade unions and collective bargaining agreements). 5 Evidence of a significant wage penalty is critical for the study of over-education as, assuming wages reflect individuals labour productivity in the workplace, it provides empirical evidence in support of the assertion that over-educated individuals have human capital (from education) that is under-utilised in their current job. If there is no such wage penalty, then over-education may be merely a statistical artefact (McGuinness, 2006). While measurement error and unobserved individual heterogeneity potentially bias many of the existing estimates, this paper uses recent results in Lindley and McIntosh (2008), Verhaest and Omey (2009), Korpi and Tåhlin (2009) and Black (2012) to assume this key assertion holds. 6 To be more precise, these individuals trade wages for non-pecuniary job attributes reflecting their preferences and, because jobs paying lower wages likely require less human capital to be performed, they subsequently accept jobs for which they are over-educated. This is consistent with the theory of compensating wage differentials in labour markets (see, for example, Brown (1980), Biddle and Zarkin (1988), Hwang, Reed and Hubbard (1992) and Bender (1998)). - Page 2 -

4 based on individuals job satisfaction levels and desire for a new job (or intentions to quit): overeducated individuals who are highly satisfied with their job and highly unlikely to quit are deemed voluntarily over-educated, while the remainder are considered involuntarily over-educated. Resultant estimates must then be validated. If this method is valid, there should be evidence of trade-offs among individuals identified as voluntarily over-educated. Specifically, compared to the well-matched, they should experience wage penalties and improvements in other job attributes, while the involuntarily over-educated should incur wage penalties without such improvements. To perform such an empirical test, a series of econometric models are used to model the relationship between voluntary over-education and job attributes. Since job attributes are the dependent variables in the analyses and they vary between being measured as continuous, ordinal and binary, three different econometric models and estimators are used: the linear fixed effects, fixed effects ordered logit and random effects probit estimators. Each uses a regression framework and panel data to control for (observable and unobservable) individual heterogeneity, which ensures estimates are not confounded by other factors and reflect only the differences in job attributes that arise from being voluntarily (and involuntarily) over-educated rather than well-matched. Few studies have acknowledged the role individuals preferences may play in the incidence of over-education and, as a result, there is little empirical evidence on the existence of voluntary overeducation. This paper, therefore, makes two significant contributions to the over-education literature. Specifically, it is the first study to attempt to estimate the incidence of voluntary over-education, and it is the first to examine the relationship between such voluntary over-education and job attributes. In doing so, it also provides the first evidence on the particular job attributes for which voluntarily overeducated individuals trade wages (i.e., their reasons for accepting over-educated jobs). This investigation has important implications for the general perception of over-education. Currently, instances of over-education are generally regarded as labour market failures. Evidence of individuals voluntarily over-educated, however, would change this. Such instances of over-education would not be the result of labour market failures, but rather individuals preferences regarding various job attributes. Thus, from the perspective of individuals, no government policy interventions are needed to prevent or resolve them, as doing so would result in welfare losses for individuals. This paper uses Australian panel data the Household, Income and Labour Dynamics in Australia (HILDA) Survey data for the period 2001 to 2008, where roughly 19 per cent of males and 23 per cent of females are identified as over-educated. Results suggest the incidence of voluntary over-education among employed individuals is between 3 and 7 per cent for males, and 5 and 10 per cent for females. This means that between 17 and 35 per cent of over-educated males and between 21 and 42 per cent of over-educated females are actually voluntarily over-educated. It is also found that being voluntarily over-educated leads to a (minor) wage reduction (between 1 and 3 per cent) - Page 3 -

5 and a host of improvements in other job attributes, such as: greater job security; working the preferred number of hours; greater job flexibility (e.g., flexible start and finish times, flexibility to decide when to work and a job closer to home); and, greater satisfaction with work performed (e.g., having greater autonomy and a job that is less stressful). Overall, seemingly as a culmination of these improvements, the voluntarily over-educated are significantly more satisfied with their achieved balance between work and life than well-matched (and involuntarily over-educated) individuals. This paper proceeds as follows. Section 2 reviews existing evidence relevant to the possibility of voluntary over-education. Section 3 outlines the data, sample restrictions and method used to identify over-education. Section 4 discusses the identification of voluntary over-education and presents estimates for its incidence. Section 5 assesses the validity of these estimates by examining the relationship between voluntary over-education and job attributes. Section 6 concludes the paper. 2. Existing evidence Few studies have acknowledged the possibility of voluntary over-education. But, relevant evidence can be found in those that have considered over-education more generally. In particular, given the proposed method for identifying voluntary over-education, evidence on the effect being overeducated has on individuals job satisfaction levels and intentions to quit is relevant. Based on the many studies examining the link between over-education and job satisfaction, the prevailing finding is that, compared to the well-matched, over-educated individuals are significantly less satisfied with their jobs (Tsang, Rumberger and Levin, 1991; Battu, Belfield and Sloane, 2000; Bender and Heywood, 2006; Verhaest and Omey, 2006b; Allen and de Weert, 2007; Amador, Nicolás and Vila, 2008; Fleming and Kler, 2008; Korpi and Tåhlin, 2009; Green and Zhu, 2010). In addition, several studies have found the over-educated are more likely to want to quit, be actively searching for a new job and actually experience voluntary job separations (Tsang et al., 1991; Hersch, 1995; Robst, 1995; Sloane, Battu and Seaman, 1999; Groot and Maassen van den Brink, 2003; Wald, 2005; Di Pietro and Urwin, 2006; Allen and de Weert, 2007; McGuinness and Wooden, 2009). Such evidence suggests over-education is an involuntary state for most individuals, but does not preclude the possibility of it being voluntary for some. Only two studies have empirically examined the issue of voluntary over-education, both of which are very recent: Mavromaras, McGuinness, O Leary, Sloane and Wei (2010) and McGuinness and Sloane (2011). 7 Both studies considered the potential for voluntary over-education by examining the relationship between over-education and job satisfaction levels. Mavromaras et al. (2010) found 7 Important caveats for Mavromaras et al. (2010) and McGuinness and Sloane (2011) are that they only considered overeducation among university graduates and they concurrently examined the effects of over-skilling. - Page 4 -

6 being over-educated had little effect on individuals job satisfaction levels and, as a result, concluded some over-education may be voluntary. In contrast, McGuinness and Sloane (2011) found overeducation had a negative, statistically significant effect on job satisfaction, but since it was smaller in magnitude than the effect of being over-skilled they argued it was evidence of over-education containing a voluntary element. 8 As in this paper, McGuinness and Sloane (2011) also sought evidence of over-education representing trade-offs between wages and other job attributes (e.g., job security, work autonomy and new challenges in work). In support of voluntary over-education, they found a positive, statistically significant relationship between over-education and job security. 9 Their results, however, are subject to important limitations. First, the ordinal measures of job attributes were dichotomised into binary variables and univariate probit models estimated, rather than econometric models and estimators specifically designed for ordinal-valued measures. Second, given the use of cross-sectional data and univariate probit models, the evidence of trade-offs may be biased due to unobserved individual heterogeneity. Finally, and most importantly, the relationship between over-education and job attributes was examined without distinguishing between voluntarily and involuntarily over-educated individuals (i.e., used a single over-educated identifier). Given existing evidence suggests overeducation is involuntary for most individuals, such analyses are unlikely to produce accurate evidence of voluntary over-education because its effects will likely be offset by the effects of involuntary overeducation, where the involuntarily over-educated are unlikely to experience improvements in job attributes. Each of these limitations is addressed in this paper. 3. Data and identification of over-education To consider the existence of voluntary over-education, this paper uses individual-level survey data from Australia. In particular, the Household, Income and Labour Dynamics in Australia (HILDA) Survey data for the period 2001 to 2008 are used. The sample examined is employed individuals who are between 15 and 64 years of age, not a full-time student and not self-employed. Meanwhile, overeducation is identified using the so-called job analyst (JA) method (Hartog, 2000; McGuinness, 2006): the ABS (2006) Australian and New Zealand Standard Classification of Occupations (ANZSCO) provides the skill (or education) levels required to perform the jobs in Australian labour markets. Details on the data, sample restrictions and identification of over-education are presented below. 8 These contrasting findings likely result from differences in the data and econometric techniques used. Specifically, Mavromaras et al. (2010) used panel data and random effects probit models (with Mundlak-Chamberlain controls) to account for any effects of unobserved individual heterogeneity, whereas McGuinness and Sloane (2011) used crosssectional data and univariate probit models that do not account for such heterogeneity; this use of univariate probit models likely leads to upwardly biased estimates for the effect of over-education (Fleming and Kler, 2008). 9 Further significant relationships were found when models were estimated separately by gender. - Page 5 -

7 3.1 The HILDA Survey data The HILDA Survey is a longitudinal survey designed to contain a sample of individuals that is representative of the Australian population residing in private dwellings. Commencing its first wave in 2001, the HILDA Survey consisted of a large national probability sample of Australian households and personal interviews with all members aged 15 years and older in the selected households. The members of the responding households in Wave 1 are the foundation of the panel tracked in future waves, with any extensions to the sample the result of changes to the compositions of these original households. In Wave 1, interviews were conducted at 7,682 households, from which 13,969 individuals were successfully interviewed. Interviews occur between August and March each year on a roughly annual basis for each individual, and they are typically administered on a face-to-face nature. For further details on the design and progress of the HILDA Survey see Wooden and Watson (2007). The HILDA Survey data was chosen because it is a longitudinal survey that contains a sample representative of the Australian population. The longitudinal nature of the data enables the estimation of econometric models that control for unobserved individual heterogeneity, while its representativeness enables inferences to be made regarding over-education throughout Australian labour markets (i.e., among all employed individuals). In addition, the HILDA Survey data contain a vast array of information on the past and current circumstances of each individual in the sample, particularly regarding their current job. This means a comprehensive set of measures can be used to capture the attributes of individuals jobs. It also means the econometric models can control for a detailed set of individual characteristics, which likely improves the accuracy of parameter estimates. 3.2 Sample restrictions Throughout the analyses, the sample is restricted to employed individuals who are between 15 and 64 years of age (i.e., working-age in Australia), not a full-time student and not self-employed. The reasons for these restrictions are as follows. First, the sample is restricted to working-age, employed individuals because over-education is directly related to individuals employment outcomes (i.e., cannot occur for not employed individuals). Second, full-time students are omitted because their employment situation is likely quite different from other individuals. For full-time students, it is likely employment serves as merely a source of financial support and is subject to the varying time constraints associated with student schedules. In addition, this group includes individuals yet to enter labour markets on a full-time and permanent basis (e.g., secondary school students). Any such individuals who are employed may be regarded as investing in their human capital (i.e., acquiring skills from work experience and a knowledge of labour markets), or as merely seeking a source of supplementary income. Regardless, full-time students will typically seek flexible jobs rather than jobs - Page 6 -

8 that fully utilise their human capital; hence their omission from the sample. 10 Finally, self-employed individuals are omitted because, similar to full-time students, their employment situation is distinct from other individuals. In particular, their role as both employer and employee should mean greater flexibility in the job and no asymmetric information regarding the availability and use of human capital. As a result, self-employed individuals should be able to tailor their jobs to match their skills and, therefore, ensure the greatest possible utilisation of their human capital. 11 This is clearly different to the employment circumstances of individuals working as employees; hence self-employed individuals are removed from the sample. 3.3 Identifying over-education in Australian labour markets As with all empirical research into over-education, this paper assumes the existence of productivity ceilings in jobs means that jobs can be regarded as requiring a certain level of human capital to be competently performed. Further, it is assumed levels of education (or qualifications) can be used to quantify such required human capital levels, which leads to the definition of required education levels for jobs. Instances of over-education are then identified by comparing each individual s highest education level with the required education level of their job. 12 The over-education literature has developed three methods for estimating such required education levels: the worker self-assessed (WA), job analyst (JA) and realised matches (RM) methods. This paper uses the JA method as it has been established as the preferred method (Hartog, 2000; Verhaest and Omey, 2006a). Also, the HILDA Survey data does not contain the information necessary to use the WA method, while the RM method is considered a poor measure, particularly for analyses using longitudinal data (Green et al., 1999; Hartog, 2000; Jensen, 2003; Leuven and Oosterbeek, 2011). 13 The JA method relies on occupational classifications developed by professional job analysts to estimate required education levels (Hartog, 2000; McGuinness, 2006). Its main advantages are that it explicitly seeks to measure the required education levels of jobs based on the tasks of jobs and the skills needed to perform them, and it is an objective measure that is derived independent of workers and firms (Rumberger, 1987; Hartog, 2000; Verhaest and Omey, 2006a). Thus, it is unlikely to 10 Part-time students, however, are considered because they are likely to have entered labour markets on a full-time and permanent basis and be undertaking education to either facilitate a change in career or progress in their current career. Such transitions may be important means for entering and exiting over-education; hence part-time students remain in the sample. 11 This is predicated on the assumption that self-employed individuals, in their role as the employer, seek to maximise their profits (i.e., fully utilise the available human capital). 12 Based on this definition, it is implicitly assumed that all individuals who complete the same qualification will acquire the same level of human capital from it, regardless of their innate abilities, the quality of the educational institution attended and the year completed. This assumption is a key limitation of over-education research. 13 This is because the RM method considers the actual matches between individuals and jobs; it examines the distribution of education levels among individuals in similar jobs, and then defines individuals with an education level one standard deviation or more above the mean to be over-educated. Thus, it does not actually attempt to identify required education levels, and its relative nature means it will always identify over-education regardless of whether it actually exists. - Page 7 -

9 contain any bias and it is consistent across individuals employed in the same job. The JA method, however, is not without its limitations. Apart from first requiring a suitable occupational classification to exist, it must be assumed that the required education levels of jobs remain fixed throughout the time period examined. 14 Also, it is necessary to assume the homogeneity of jobs within the same occupational title of the classification. It is, therefore, optimal to use the occupational classification at highly detailed or disaggregated levels as this increases the likelihood that such an assumption will hold. 15 This study uses the ABS (2006) ANZSCO to derive estimates for required education levels. ANZSCO was developed by the ABS, Statistics New Zealand and the Australian Government Department of Employment and Workplace Relations to enable the consistent analysis of information and statistics on occupations across studies. 16 It is a skill-based classification that refers to all jobs in the Australian and New Zealand labour markets that are performed for a wage. The structure of ANZSCO consists of five hierarchical levels: Major groups; Sub-major groups; Minor groups; Unit groups; and, Occupations. The Occupations categories are the most detailed level of the classification, while the Major groups the broadest. Specifically, there are 998 Occupations categories and 8 Major groups. In creating these categories, ANZSCO groups jobs into occupations (i.e., sets of jobs with similar tasks), and then organises occupations into increasingly larger groups (i.e., the Unit, Minor, Sub-major and Major groups) based on the similarity of tasks, which is judged in terms of the level and specialisation of skill needed to complete the tasks. This aim of grouping together occupations with similar sets of tasks and required skill levels makes ANZSCO ideal for estimating the required education levels of jobs in Australian labour markets. At each level of the classification, ANZSCO defines a required skill level for each of the occupation categories. Specifically, ANZSCO defines five skill levels and assigns one (or more) of them to each occupation category. These skill levels are defined in terms of levels of formal education and training (i.e., qualification levels based on the Australian Qualifications Framework (AQF)), and years of previous experience in a related occupation (i.e., relevant experience) that may substitute for this formal education (see Table A1 in Appendix A for the definition of the ANZSCO skill levels). Thus, ANZSCO provides estimates for the required education levels of jobs and, therefore, enables the identification of over-education. This paper uses ANZSCO at the Unit group 14 Unless, of course, the occupational classification is updated over time. If not, required education levels may be underestimated, particularly for jobs that change rapidly due to technological advances, and, as a result, the incidence of overeducation would be over-estimated (Rumberger, 1987; Borghans and de Grip, 2000; Chevalier, 2003; McGuinness, 2006). 15 This is dependent, however, on the level of disaggregation reported in the data used. Also, for the given data, it must be assumed individuals are accurately coded to the occupational classification. Given its importance for identifying overeducation, this study has exerted considerable effort in ensuring the accuracy of the ANZSCO occupation (and educational attainment) information in the HILDA Survey data (details on this data cleaning are available from the author on request). 16 The following discussion of ANZSCO draws on the supporting documentation provided in the ABS (2006) publication. - Page 8 -

10 level as, according to ABS (2006), using the 358 occupation categories at the Unit group level is sufficient for an accurate analysis of occupations by skill level because virtually all occupations within the same Unit group have the same required skill level. It is also the most detailed level at which ANZSCO information is available in the HILDA Survey data. 17 Hence, the necessary assumption of homogeneity of jobs within the same occupational category is made with some confidence. Since ANZSCO defines a range of formal education levels for each skill level (i.e., defines a minimum and maximum required education level for each occupation), over-education is identified as follows. An individual is deemed over-educated if their highest education level exceeds the maximum required education level considered necessary to perform their current job. On the other hand, an individual is deemed under-educated if their highest education level is less than the minimum required education level for performing their job and their years of relevant work experience are less than those necessary to substitute for this formal education. This definition of under-education explicitly allows for the substitution of work experience for formal education because it, arguably, better reflects the reality of modern labour markets (i.e., the willingness of firms to recognise that certain amounts of work experience likely result in individuals obtaining the same human capital, or achieving the same labour productivity in the workplace, as if they had completed a formal qualification). 18 Thus, it aims to minimise the number of anomalous cases where individuals appear to be employed in jobs for which they have insufficient skills to perform. Finally, an individual is deemed well-matched if they are neither over-educated nor under-educated. Such individuals are considered to have human capital, at least in terms of formal education and work experience, which roughly matches that required to perform their current job. Hence, they are fully utilising their human capital derived from education. Table 1 presents the estimated incidences of each state. 17 The availability of such detailed ANZSCO information is another reason HILDA Survey data was chosen for this paper. 18 It is assumed that relevant work experience refers to the total number of years in which an individual has worked in their current occupation, and not necessarily on a continuous basis (i.e., their tenure in current occupation). To the author s knowledge, there are no previous studies that use such a definition of under-education. - Page 9 -

11 Table 1: Incidence of over-education by year and gender Employed individuals aged years, excluding full-time students and self-employed (%) Overall Males Over-educated Under-educated Well-matched Total Sample size (N) 3,141 3,006 3,034 2,952 3,060 3,127 3,112 3,165 24,597 Females Over-educated Under-educated Well-matched Total Sample size (N) 3,077 2,864 2,904 2,865 3,060 3,131 3,167 3,190 24,258 SOURCE: Author s calculations using HILDA Survey data (Release 8.0). NOTES: Figures are proportions that sum to for (sub-sections in) each column and are weighted using crosssectional population weights to make them representative of the Australian population of employed individuals who are between 15 and 64 years of age, not a full-time student and not self-employed. The results in Table 1 indicate over-education is more prevalent among females, with a range between roughly 21 and 24 per cent compared to 16 to 20 per cent for males. On average, over the eight-year period, approximately 23 per cent of females and 19 per cent of males are observed overeducated. In addition, roughly 70 per cent of males and females are well-matched and 10 per cent under-educated, though the estimated incidences of well-matched and under-educated individuals tend to be somewhat higher for males than females. For the eight-year period, Table 1 also suggests a small increase in the rate of over-education (and under-education) in Australia. 4. Estimating the incidence of voluntary over-education In this paper, individuals are considered voluntarily over-educated if their current job has a set of attributes that satisfies their preferences. If the job is truly their preferred outcome, it is reasonable to expect voluntarily over-educated individuals would report being highly satisfied with their job and not wanting to leave it, at least in the near future. Involuntarily over-educated individuals, on the other hand, would be unsatisfied with their current job and, most likely, seeking a new one. Thus, it appears information on individuals job satisfaction levels and their desire for a new job (or intentions to quit) can be used to empirically identify instances of voluntary over-education. This paper proposes, therefore, that over-educated individuals who are highly satisfied with their job and highly unlikely to quit (in the near future) can be considered voluntarily over-educated, while the remainder are involuntarily over-educated. Information in the HILDA Survey dataset on individuals overall job satisfaction levels and intentions to quit their current job (in the next 12 months) is used to perform this distinction. Job - Page 10 -

12 satisfaction levels are measured on a scale from 0 (totally dissatisfied) to 10 (totally satisfied), while intentions to quit are measured as probabilities from 0 per cent (no chance) to 100 per cent (absolute certainty). Given these measures, there is then a need to define thresholds for highly satisfied and highly unlikely to quit. Choosing the values for such thresholds is somewhat arbitrary (e.g., is a job satisfaction level of 8 sufficiently high for an individual to be classified highly satisfied, or should the threshold be a job satisfaction level of 9) and, of course, different values will produce different estimates of voluntary over-education. Defining such thresholds is therefore a limitation of the analysis that leads to uncertainty regarding voluntary over-education estimates. As a result, sensitivity analyses using various thresholds are conducted to examine the robustness of the results. To determine appropriate thresholds for highly satisfied and highly unlikely to quit, the distributions of overall job satisfaction levels and intentions to quit are examined for all employed individuals. 19 The data (pooled across all eight years) reveal around 70 per cent of employed individuals report a job satisfaction level in the range from 7 to 9; roughly 30 per cent report an 8 (the mode value), while the remaining 40 per cent are distributed evenly between responses 7 and 9. A further 12 per cent report a job satisfaction level of 10. Regarding intentions to quit, the data reveal around 55 per cent of employed individuals report a 0 per cent chance of quitting, while at the other extreme roughly 7 per cent report a 100 per cent chance. Also relevant to determining the appropriate thresholds is the possibility that individuals job satisfaction levels and intentions to quit may change with time in the same job. Thus, individuals may move between being voluntarily and involuntarily over-educated without changing job. The concern here is that contentment with jobs (i.e., individuals initially dissatisfied becoming contented due to a lowering of their expectations over time) may lead to individuals being wrongly identified as voluntarily over-educated. 20, 21 In an attempt to avoid such mismeasurement highly satisfied is defined as a job satisfaction level of 9 or 10; this threshold is higher than the modal job satisfaction level among employed individuals and is, arguably, sufficiently high to ensure over-educated individuals who become contented with their job are not deemed voluntarily over-educated. In addition, highly unlikely to quit is defined as a 5 per cent chance or less of quitting; this threshold is unlikely to have much effect on voluntary over-education estimates given most employed individuals report a 0 per cent chance of quitting. 19 See Figures B1 and B2 in Appendix B for these distributions. Further analyses (not presented, but available on request) show the distributions are similar for males and females; hence, separate thresholds appear unnecessary. 20 Of course, not all movement between voluntary and involuntary over-education that occurs without a change of job is necessarily measurement error; given job attributes may change over time, it is entirely plausible that individuals may move from involuntarily over-educated to voluntarily over-educated (and vice versa) without changing jobs. 21 This paper does not examine the dynamics of job satisfaction levels and intentions to quit among over-educated individuals or the dynamics of voluntary and involuntary over-education; these are suitable avenues for future research. - Page 11 -

13 Table 2: Incidence of voluntary and involuntary over-education by year and gender Employed individuals aged years, excluding full-time students and self-employed (%) Overall A. Preferred estimates ( highly satisfied if job satisfaction 9; highly unlikely to quit if chance quit 5%) Males Over-educated Voluntary Involuntary Sample size (N) 3,141 3,006 3,034 2,952 3,060 3,127 3,112 3,165 24,597 Females Over-educated Voluntary Involuntary Sample size (N) 3,077 2,864 2,904 2,865 3,060 3,131 3,167 3,190 24,258 B. Alternative estimates of voluntary over-education More restrictive definition (i.e., lower bound estimates) ( highly satisfied if job satisfaction = 10; highly unlikely to quit if chance quit = 0%) Males Females Less restrictive definition (i.e., upper bound estimates) ( highly satisfied if job satisfaction 8; highly unlikely to quit if chance quit 10%) Males Females SOURCE: Author s calculations using HILDA Survey data (Release 8.0). NOTES: Figures are proportions that are weighted using cross-sectional population weights to make them representative of the Australian population of employed individuals who are between 15 and 64 years of age, not a full-time student and not self-employed. Italicised figures are the estimated incidences of over-education reported in Table 1. Based on the above thresholds, panel A in Table 2 presents estimates of the incidence of voluntary over-education among males and females. Specifically, previous estimates of overeducation (as presented in Table 1) are disaggregated by whether individuals are voluntarily or involuntarily over-educated. As proportions of all employed persons, estimated incidences of voluntary over-education are quite small; over the eight-year period roughly 3 per cent of employed males are voluntarily over-educated, while the rate is higher for females at around 5 per cent. However, with regards to all instances of over-education, they are certainly important: around 17 per cent of over-educated males and 21 per cent of over-educated females are actually voluntarily overeducated. Sensitivity analyses using alternative thresholds examine the robustness of these estimates. In particular, job satisfaction levels of 8, 9 and 10 are used as thresholds for highly satisfied, and intentions to quit of 0 per cent, 5 per cent and 10 per cent are used as thresholds for highly unlikely to quit. The results estimates of voluntary over-education based on the various thresholds are presented in Table C1 in Appendix C. As expected, altering the threshold for highly unlikely to quit has little effect on voluntary over-education estimates. The threshold for highly satisfied, however, - Page 12 -

14 greatly affects the estimates: using a job satisfaction level of 8 (rather than 9) roughly doubles the estimates, while a job satisfaction level of 10 more than halves them. To outline the variability (or uncertainty) associated with voluntary over-education estimates, panel B in Table 2 presents what are, arguably, the lower bound and upper bound estimates. 22 These estimates suggest the rate of voluntary over-education among employed males is approximately between 1 and 7 per cent, and for employed females between 2 and 10 per cent. The above results may actually overstate the imprecision of the voluntary over-education estimates. Specifically, using a job satisfaction level of 10 as the threshold for highly satisfied seems unnecessarily restrictive, and it produces lower bound estimates that are most likely too low. If the results in panel A of Table 2 are instead used as lower bound estimates, then the estimated rates of voluntary over-education are roughly between 3 and 7 per cent for males and 5 and 10 per cent for females. 23 These correspond to between 17 and 35 per cent of over-educated males and 21 and 42 per cent of over-educated females being voluntarily over-educated. Nevertheless, despite some uncertainty surrounding the estimates, the results clearly indicate that some individuals are indeed voluntarily over-educated. 5. Voluntary over-education and job attributes To assess the validity of the above estimates of voluntary over-education, this section examines the relationship between such voluntary over-education and job attributes. If the method developed in Section 4 truly identifies voluntary over-education (i.e., individuals who accept over-educated jobs because they have a set of attributes that satisfy their preferences), then there should be evidence of the individuals identified as voluntarily over-educated having made trade-offs between wages and other job attributes. In particular, it should be found that, compared to well-matched individuals, the voluntarily over-educated experience wage penalties and improvements in other job attributes. The experiences of the voluntarily over-educated and involuntarily over-educated should also be different; most importantly, evidence should be consistent with the assumption that involuntary overeducation is not the result of trade-offs made by individuals (i.e., is an undesirable state). Hence, the involuntarily over-educated should be found to also incur wage penalties but without improvements in other job attributes. To examine the relationship between voluntary over-education and job attributes, econometric models are developed with a series of individuals job attributes as dependent variables 22 The most restrictive thresholds for highly satisfied and highly unlikely to quit are used to derive the lower bound estimates, while less restrictive thresholds are used to derive what are, arguably, upper bound estimates. 23 These may still overstate the imprecision of voluntary over-education estimates: to the extent a job satisfaction level of 8 is too low a threshold for highly satisfied and leads to contented individuals (as discussed above) being mistakenly identified as voluntarily over-educated, then these upper bound estimates will be too high. - Page 13 -

15 and identifiers for being voluntarily and involuntarily over-educated included as explanatory variables. The HILDA Survey dataset contains an array of information describing the characteristics of each individual s current job. This analysis seeks to consider the key job attributes. 24 In particular, the following measures are used to capture the attributes of individuals jobs: real hourly wage; chance of job loss in the next 12 months (i.e., perceived job security); hours worked (per week); travel time to and from work (hours per week); satisfaction with pay; satisfaction with job security; satisfaction with hours worked; satisfaction with work-life balance achieved; satisfaction with the work itself; flexibility to decide when do work; autonomy to decide how do work; input about what happens at work; stressfulness of job; complexity and difficulty of job; required learning of new skills in job; extent job uses individuals skills; indicator for being employed on a permanent/ongoing contract; indicator for individual working their preferred number of hours per week; and, 25, 26 indicator for job providing flexible start and finish times. Since job attributes are the dependent variables in this analysis and they vary between being measured as continuous (e.g., real hourly wage), ordinal (e.g., satisfaction with pay, flexibility to decide when do work) and binary (e.g., indicator for working preferred number of hours), three different econometric models and estimators are used. Each uses a regression framework and panel data with the aim of controlling for (observable and unobservable) individual heterogeneity. This 24 In some cases, however, proxy measures are used rather than measures for the job attributes themselves. For example, individuals satisfaction with their level of job security is examined; this measures an outcome (individuals satisfaction level) associated with the job attribute, rather than the job attribute itself (the level of job security). Nevertheless, these proxy measures should provide valid information on the underlying job attributes, and provide valuable evidence regarding the states of voluntary and involuntary over-education. 25 For further details on these job attributes measures see Appendix D: Table D1 contains descriptions and details on variable derivations; and, Table D2 contains descriptive statistics (means and standard deviations) by over-education status. 26 The following are the proxy measures for job attributes: satisfaction with pay; satisfaction with job security; satisfaction with hours worked; satisfaction with work-life balance achieved; and, satisfaction with the work itself. - Page 14 -

16 ensures the resultant estimates are not confounded by other factors and reflect only the differences in job attributes that arise from being voluntarily over-educated rather than well-matched Econometric models and estimators The econometric models and estimators used to examine the continuous-, ordinal- and binary-valued job attributes are outlined below. Continuous-valued job attributes For continuous-valued job attributes, the underlying model for each attribute is defined as follows:, [1] where are job attributes for individual i at time t, is a vector of controls for individual characteristics at t, is a vector of coefficients associated with, and are indicators for being voluntarily over-educated and involuntarily over-educated at t, and are the estimated parameters of interest, is a time-invariant random variable representing unobserved effects for individual i, and are idiosyncratic errors that vary across i and t (the usual regression disturbances). 28 Since may contain factors correlated with both and (e.g., certain personality traits being correlated with an individual s education level and their number of hours worked per week), OLS estimation of [1] may produce biased and inconsistent estimates of and. To overcome any such unobserved heterogeneity bias (i.e., eliminate from [1]), the models 29, 30 for continuous-valued job attributes are estimated using the (linear) fixed effects estimator. 27 Thus, statistically significant evidence also identifies the particular job attributes for which voluntarily over-educated individuals trade wages. 28 The continuous-valued job attributes examined are: real hourly wage; chance of job loss; hours worked; and, travel time (hours per week). Note, however, that treating chance of job loss as continuous may not be entirely appropriate given it is bounded on the scale from 0 to 100. As a result, sensitivity analyses are performed where it is treated as ordinal-valued. 29 The figures in Table E1 in Appendix E indicate the voluntarily over-educated and involuntarily over-educated identifiers contain sufficient within variation to enable the fixed effects estimator to produce valid estimates. For further details on the fixed effects estimator see, for example, Wooldridge (2003) and Cameron and Trivedi (2005). 30 Since the errors for each individual in [1] are almost certainly correlated over time, robust (panel-corrected) standard errors must be calculated to ensure valid statistical inference (Cameron and Trivedi, 2005). - Page 15 -