Job search duration of higher education graduates in the Netherlands

(1)

Master’s Thesis

Job Search Duration of

Higher Education Graduates in The Netherlands

Jana van Leuven

Student number: 10053050

Date of final version: February 13, 2016 Master’s programme: Econometrics

Specialisation: Free Track

Supervisor: dr. J. C. M. van Ophem Second reader: dr. N. P. A. van Giersbergen

(2)

List of Tables

4.1 Socio-economic characteristics . . . 19

4.2 Characteristrics of educational career . . . 20

4.3 Searching behaviour . . . 21

4.4 Number of complete and incomplete unemployment spells according to Definition 1 23 4.5 Time series: Some characteristics . . . 28

4.6 Results ADF-test . . . 29

4.7 Correlation coefficients before and after HP-filtering and corresponding cyclical behaviour . . . 29

4.8 Business sectors and educational clusters . . . 30

5.1 Likelihood Ratio test results . . . 35

5.2 Description of variables . . . 37

5.3 Estimated hazard ratios from PWC and Cox Model for HVE and university students . . . 51

5.4 Proportion of study programs in each area (top, sub top, sub bottom, bottom) for each cluster . . . 58

A.1 Educational sectors and their corresponding study programs . . . 65

A.2 Estimated hazard ratios of the HVE study program dummy variables belonging to the PWC model and Cox model in Table 5.3 . . . 67

A.3 Estimated hazard ratios of university study program dummy variables belonging to the PWC model and Cox model in Table 5.3 . . . 69

A.4 Estimated hazard ratios of the timepieces belonging to the PWC models in Table 5.3 . . . 71

(5)

List of Figures

4.1 Secondary and higher education system in the Netherlands . . . 17

4.2 Number of different searching methods used . . . 22

4.3 Distribution of searching moment around graduation . . . 22

4.4 Histogram of the job search duration according to Definition 1 . . . 23

4.5 Kaplan-Meier estimates of survival functions for HVE students and University students . . . 24

4.6 Kaplan-Meier estimates of survival functions for the different educational sectors (HVE and university together) . . . 24

4.7 Unemployment duration: Nelson-Aalen estimates of cumulative hazard functions for HVE students and University students . . . 25

4.8 Unemployment duration: Nelson-Aalen estimates of cumulative hazard functions for the different educational sectors (HVE and university together) . . . 25

4.9 The development of the unemployed labour force, the number of vacancies, the number of vacancies per unemployed and the GDP growth rate . . . 27

4.10 Plots of time series and their estimated business-cycle component . . . 31

5.1 Estimated baseline hazard rates for HVE and university students . . . 41

5.2 Relative effects of the study programs on the hazard rate . . . 52

5.3 Relative effects of the study programs on the hazard rate . . . 53

5.4 Average search duration of HVE study programs . . . 54

5.5 Average search duration of university study programs . . . 56

5.6 Distribution of the study programs over the different areas (top, sub top, sub bottom and bottom) per educational sector (Alpha, Beta, Gamma and Medical) 58 5.7 Average search duration of HVE and university students, together with their 95% confidence bands, and the national GDP growth rate . . . 58

5.8 Average search durations of the four educational clusters, together with their 95% confidence bands (HVE) . . . 59

5.9 Average search durations of the four education clusters, together with their 95% confidence bands (University) . . . 59

(6)

LIST OF FIGURES v 5.10 Average search durations of alpha students (HVE and university), together with

their 95% confidence bands, and the GDP growth rates of the Education and

Business Services sectors . . . 60

5.11 Average search durations of beta students (HVE and university), together with their 95% confidence bands, and the GDP growth rates of the Manufacturing

and Education sectors . . . 60

5.12 Average search durations of gamma students (HVE and university), together with their 95% confidence bands, and the GDP growth rates of the sectors Business

Services and Public Administration . . . 61

5.13 Average search durations of medical students (HVE and university), together with their 95% confidence bands, and the GDP growth rate of the Health and

Social Work Activities sector . . . 61

B.1 Goodness-of-fit of PWC and Cox model for HVE students . . . 72

(7)

Chapter 1

Introduction

In today’s competitive labour market, higher education graduates in most western economies have found themselves in a challenging situation. Due to growing participation rates in higher education and the right of free movement for workers in the European Union, students entering

the labour market face high competition. Moreover, students often have to compete with

individuals who have already established a position on the labour market, making competition even tougher (European Council, 2012; Wolbers, 2007). In this thesis, the labour market position of higher education graduates in the Netherlands is investigated by examining the job search durations of graduates from different fields of study.

Determining these job search durations is of particular interest from a policy perspective. First, information on the job search durations of graduates from different fields of study will help policymakers understand the nature of graduates’ chances on the labour market. Furthermore, this understanding will guide them in assessing their capacity to influence a smooth transition from school to work. By investing wisely in educational systems and redesigning programs and curricula, policymakers have the opportunity to bridge school and work and enhance opportuni-ties for newcomers in the labour market. For example, the level of stratification of educational systems may have an impact on the smoothness of the school-to-work transition. Furthermore, designing curricula in such a way that it allows for swichting later on in the study program increases flexibilty for students and helps students react to employers’ needs in the labour mar-ket (Pavlin et al., 2014). Lastly, it may become clear in which direction unemployed graduates who are not able to find a job quickly can be retrained (Berkhout, 2004). Thus, understanding the nature of the labour market relevance of graduates contributes to a smooth transition from school to work since it enables policymakers to align curricula with labour market needs.

Information on the labour market relevance of study programs is also useful for individuals. When entering higher education, students face the difficult decision of choosing a suitable study program that aligns with their interests and perspectives. This choice may have a large influence on their future life since it may affect their starting position on the labour market. Young adults are therefore in need of information that provides them with knowledge about the labour market perspective of the different study programs available to them (Berkhout, 2004).

(8)

CHAPTER 1. INTRODUCTION 2 The goal of this thesis is to determine which study programs offer higher education graduates in the Netherlands the best starting position in the labour market. The influence of the business cycle is of particular interest in the analysis. Specifically, it is investigated which sectors of the economy in the Netherlands are sensitive to fluctuations in the business cycle and how this influences the job search duration of higher education graduates from different fields of education.

The analysis in this thesis is carried out with use of two duration models in which the hazard rate is formulated as the probability of finding a job. This type of models has been chosen to conduct the research on the labour market relevance since the job search duration reflects the relationship between demand and supply of higher education graduates well. If the supply of graduates exceeds the demand, then the job search duration will typically be longer and the labour market relevance will typically be lower and vice versa. Hence, assessing the job search duration helps determine the labour market relevance of higher education graduates. The first duration model used in this thesis is the piecewise constant hazard rate model. This is an exponential hazard rate model in which the constant hazard rate is allowed to vary within per-defined time intervals. The second model is called the Cox proportional hazard rate model. This model does not require the formulation of a distribution function of the time component of the model. The analysis of the cyclical behaviour of the various business sectors in the Netherlands is executed with use of a correlation analysis.

The analysis is conducted using two data sources. Firstly, macroeconomic data are used to execute the cyclicality analysis. Specifically, quarterly data from 1996 to 2015 on the Gross Domestic Product (GDP) growth rate are used. Both sector-level and national GDP growth rates are taken into account. In the analysis, the GDP growth rates are used as the cyclical indicator. Other data reflecting the economic climate in the business sectors that are used in this thesis are data for the unemployed labour force and the number of job vacancies.

Secondly, individual data collected through an online questionnaire are used. Each year since 1997, newly-graduated students from higher education institutions in the Netherlands are requested to complete a questionnaire consisting of over 60 questions about their educational career, job searching behaviour, current labour market position and personal characteristics. Each survey polls 20,000 graduates from more than one hundred study programs. The goal of the survey is to determine which study programs offer graduates the best opportunities on the labour market.

The remainder of this thesis is organized as follows. Chapter 2 contains a literature review. Chapter 3 describes the methodology used in the thesis. After that, a comprehensive description of the data is given in Chapter 4. The results of the analysis are presented in Chapter 5. Finally, Chapter 6 contains the conclusions of the thesis.

(9)

Chapter 2

Literature Review

In today’s competitive labour market, the employability of higher education graduates has become of growing importance for most western economies. Employability is defined as “the combination of factors which enable individuals to progress towards or enter employment, to stay in employment and to progress during their career” (European Council, 2012, p. 4). Due to growing participation rates in higher education and the right of free movement for workers in the European Union, students entering the labour market face high competition and are expected to have ‘world-class skills’ in order to be employable (Tholen, 2014; Leitch, 2006). Hence, the debate around graduate employability has become increasingly important in the political domain. In their education, youth, culture and sports meeting on the 10th and 11th of May 2012, the European Council stressed the importance of “strengthening links between higher education institutions, employers and labour market institutions in order to take greater account of labour market needs in study programmes, to improve the match between skills and jobs, and to develop active labour market policies aimed at promoting graduate employment” (European Council, 2012, p. 2). Moreover, the European Council adopted a target that by 2020 the share of employed graduates between the age of 20 and 34 years having left education no more than three years ago should be at least 82 percent (European Council, 2012; Pavlin et al., 2014).

This chapter discusses a number of existing theories in the school-to-work transition frame-work. The theories address why people invest in higher education and how this affects their employability. Consequently, empirical evidence on graduate employability and labour market relevance of graduates is discussed. What are important determinants of graduate employa-bility and why is enhancing employaemploya-bility among graduates important? Lastly, the role of the business cycle in the job-finding process of higher education students is pointed out.

2.1 Economic Theory

There is a number of well-known economic theories that relate to the school-to-work transition. One of the most important theories is the human capital theory. The human capital theory

(10)

CHAPTER 2. LITERATURE REVIEW 4 states that education is one of the most important investments in human capital and that education will help increase a person’s earnings, even after netting out direct and indirect costs of schooling and after adjusting for other variables such as family background and gender. Education is seen as an individual investment and the decision is based on a well-considered assessment of the costs (of education) and future benefits (a higher wage). By investing in their human capital, individuals create a favourable starting position on the labour market and enhance their chances of finding a job and attaining the desired wage (Becker, 2009).

Many economists have tried to estimate the return to years of schooling. Grubb (1993) estimated the returns to postsecondary education and showed that workers earn a substantial premium from obtaining a university degree. Furthermore, there has been attention to the extent to which the university premium differs across fields of education. Finnie and Frenette (2003), for example, investigated this for Canadian Bachelor’s level university graduates and showed a statistically significant difference in earnings across majors. According to Finnie and Frenette (2003), the highest earnings fields are Health, Engineering and Computer Science, Commerce and Mathematics/Physics. The low-earnings fields are, among others, Arts and Humanities and Agricultural Sciences. Kelly et al. (2010) investigated the economic returns to different fields of study in Ireland. Their results point out that there are high returns to Medicine and Veterinary, Education, Engineering and Architecture and Computer Sciences.

Although the human capital theory has been formulated mainly for wages, it is also helpful in explaining the labour market integration of young graduates. That is to say, another aspect of the human capital theory is the relationship between education and productivity. The theory states that investing in human capital is useful as long as it increases one’s productivity on the labour market. This productivity is then rewarded by employers offering the highest wages to those who have obtained the highest level of productivity. Thus, the human capital theory states that students who have acquired a degree in higher education have enhanced their employability (Wolbers, 2003).

On the contrary, the job-competition theory claims that wages are not influenced by an individual’s productivity but are primarily determined by job characteristics. Spence (1973) claims in his research that job applicants compete for a job of which the salary is fixed and are assessed based on their (educational) qualifications. In his model, the field of education of an applicant is a signal of his productivity. This productivity determines the costs involved to train this person on the job. The employer observes the signal and ranks the applicants accordingly. He searches for the best available candidate at the lowest training costs. The idea that the field of education does not increase a student’s productivity but is instead used as a criterion to screen and rank the candidates, is known as the signalling/screening-hypothesis (Spence, 1973). This hypothesis is relevant in the school-to-work transition framework because graduates have little work experience and hence a low or unknown labour productivity. When selecting future employees, employers are therefore forced to look at indicators of productivity such as field of education (Spence, 1973; Wolbers, 2003). The choice of study program in an early stage of a

(11)

CHAPTER 2. LITERATURE REVIEW 5 student’s academic career can therefore influence his/her employability when graduated.

A third theory regarding the school-to-work transition framework is the job-matching theory. This theory combines the human capital and job-competition theory and is based on the idea that the labour market consists of many different jobs that require different levels of skills, as well as workers of many different skill and experience levels. Jobs and applicants should be matched according to their skill level and one speaks of a mismatch if either the supply side of skilled workers or the demand side of skilled positions exceeds the other. The quality of a job match – the degree of fit between required and acquired skills – determines the level of productivity and the associated wage. An employee who is working in a mismatched job has a lower utility level, affecting his productivity and eventually resulting in a lower wage (Boudarbat and Chernoff, 2010; Wolbers, 2003). The allocation of workers is optimal if every worker is allocated to a job in which he performs best, relative to all the other workers. Job mismatching occurs when demand and supply differ and can be the result of incomplete information on the characteristics of the job offered and the abilities of the applicants. This is often the case for school-leavers. They possess little work experience and may have to compete with those who have already gained a position on the labour market. This often forces them to either accept a job that is below their educational qualifications or face unemployment (Wolbers, 2003). Hence, job mismatching is an important factor in the school-to-work transition framework.

These theories show the importance of investing in a higher education degree and the role this degree may play in the job searching process of graduates. It may increase someone’s productivity or serve as a signal to the employer. Either way, one can conclude that a degree in higher education increases one’s chances on the labour market. There may be differenes, however, among the various fields of study available in higher education.

2.2 Previous Literature

Now that a number of important theories related to the school-to-work transition is considered, previous literature on this topic is elaborated. According to the European Council’s definition introduced earlier, three phases of employability can be distinguished: preparing for employ-ment, transition from education to employment and staying in and progress during employment. This section examines why the second phase, i.e. the transition from school to work, is so im-portant in the life of a young adult.

Although the transition from school to work is only the initial step into the labour market, many studies have emphasized the importance of a smooth transition. Not only does a smooth transition minimize early unemployment, it also speeds up the process towards permanent employment. Difficulties in an early stage of a professional career can lead to unemployment, which in turn increases an individual’s future risk of being unemployed again later on in its career and may have a long-lasting effect on a person’s income. It may also lead to insecure employment, such as temporary or parttime contracts (Korpi et al., 2003; Schmelzer, 2011). As stated in the previous section, this early stage in the labour market is turbulent due to

(12)

CHAPTER 2. LITERATURE REVIEW 6 uncertainty on the part of both employers and employees. Graduates have little work experience, making it difficult for them to prove to future employers that they posses the skills necessary to execute the job. In addition, they sometimes have to compete with individuals who have already established a position in the labour market. The evidence graduates have is their educational degree and the employer uses this degree as a signal of productivity and trainability. It is often the case that the skills and knowledge required to do a particular job are extremely job-specific. So specific that almost all training takes place on the job. The extent to which an educational degree indicates the ease with which an individual is able to adopt these new skills is therefore extremely important for a graduate’s employability. It reflects the ease with which an individual can be trained and is thus an indication of the costs associated with this training (Korpi et al., 2003; Wolbers, 2007). Hence, the field of education plays an important role in youth employability because it enhances the employment opportunities of graduates as it serves as an indicator of the costs involved for the employer.

Another argument why a smooth transition is important is that a turbulent school-to-work transition may lead to education-to-job mismatches and overeducation (Levels et al., 2014; Verhaest and Van der Velden, 2013). One speaks of a horizontal mismatch if the job is not in the domain of the field of study. A vertical mismatch takes place if the job requires a lower level of education than acquired. Mismatches are usually the result of educational systems that are failing to transfer useful skills to students and labour markets that are unable to effectively allocate these students to the jobs available. Vertical mismatching comes at a cost for the economy as a whole. If overeducated workers’ skills and knowledge were fully utilized, the overall productivity and welfare of the country would be higher. In the same way, an undereducated workforce harms the productivity since these workers may be lacking the knowledge needed to perform well in their jobs (Levels et al., 2014). A considerable amount of studies have pointed out that a large part of the working population is overeducated. For example, Groot and Van den Brink (2000) conducted a meta-analysis by examining 25 studies on overeducation that represent several years of data collection in several countries. Groot and Van den Brink (2000) found an average overeducation of 21.5% in European countries.

Apart from productivity consequences, overeducation also decreases the level of job satisfac-tion for employees. According to Verhaest and Omey (2009), overeducated workers experience psychological costs and lower utility levels due to low job satisfaction. This low utility level is substantial and in some cases a wage increase is not enough to compensate for this low utility level (Verhaest and Omey, 2009). For employers, vertical job mismatching is far from ideal too. As emphasized in the signalling theory, employers aim to hire individuals who are expected to be the most productive and least costly. Since vertical mismatching affects the productivity of a person, the costs may increase.

Horizontal mismatching comes at a cost as well. Wolbers (2003) finds that horizontally mismatched workers achieve a lower professional status and more often search for another job. Moreover, they often participate in ongoing vocational training on the job, which again comes

(13)

CHAPTER 2. LITERATURE REVIEW 7 at a cost for the employer.

Lastly, a smooth labour market entry proccess has social implications for graduates. Eco-nomic independence is essential in the transition from youth to adulthood. Most young people do not start forming a family until they have attained a sufficient level of financial stability. Youth unemployment can affect the timing of this process (Schels, 2013). Moreover, it may constrain young people’s opportunities for social participation, which in turn may lead to psy-chological distress and a general lack of social integration (Korpi et al., 2003).

Now it is clear why a smooth transition from school to work is important, it is interesting to investigate which factors enhance a smooth transition. Previous research has identified three categories of characteristrics that contribute to the employability of individuals. First, there are personal characteristics, such as age, gender, marital status, household structure, region of living, level of education and parental level of education. A number of papers have pointed out that these individual facets have a significant effect on the probability of finding a job (see for example Chuang (1995); Longhi and Taylor (2014); Perez et al. (2010); Narendranathan and Stewart (1993)).

The second category consists of local market characteristics. These characteristics are not always included in the model directly. Instead, some researchers use regional characteristics, such as the unemployment rate (as a proxy for local demand conditions), or an interaction term between a regional dummy and a year dummy (as a proxy for unobserved labour market characteristics) (see Lynch (1989); Schels (2013); Pavlin et al. (2014); Gorter and Kalb (1996)). GRIP et al. (2004) take a wholly different approach in this and do not consider any personal characteristics. They approach employability from the perspective of the industry. They criti-cize the fact that in existing literature individual, “supply-side” characteristics are usually taken as the main units of the analysis on employability, disregarding the the labour market segments in which workers are employed. They consider employability to be merely a “demand-side” issue, focusing on inter-sectorial variation in the need for employability. In their paper, they construct a so-called Industry Employability Index (IEI) that relates individual employability to the need for employability in a particular sector. The index comprises three characteristics: current workforce employability, the need for employability and effectuation conditions. They use among others, job tenure, participation in trainings, training provisions in the sector, tech-nological developments in the sectors and demographic developments (an ageing workforce) in the sector (GRIP et al., 2004).

Lastly, the extent to which higher education study programs are designed to ensure a smooth transition from school to work is a determinant of the employability of graduates. The level of vocational orientation in the curriculum is a good example. Vocational training teaches skills that are strongly in demand by employers and hence give employers an incentive to choose these students because they may be more productive or more suitable for the job. This emphasis on vocational training is usually more present in the lower level of the Dutch higher education system, called Higher Vocational Education. The other level in higher education in

(14)

CHAPTER 2. LITERATURE REVIEW 8 the Netherlands is university. The characteristics of Higher Vocational Education and university are discussed in more detail in Section 4.1.

Another facet that better connects school to work is the stratification of the educational system. A stratified educational system captures the categorization of students with different ability and skill levels into separate tracks. This has two implications. First, employers are better informed about the ability level of the applicant in a stratified system. Second, the job description of required skills can be formulated more precisely in highly stratified educational systems. This makes it more clear to applicants which ability levels the various jobs available on the labour market require (Levels et al., 2014; DeLuca et al., 2015; Wolbers, 2007).

An important factor that is often disregarded in existing literature is the role of the business cycle in the job searching process of graduates. This thesis aims to fill this gap by taking into account the cyclical behaviour of business sectors in the analysis. Business cycle dynamics are likely to influence the job search duration since economic fluctuations alter the demand and supply context in the labour market. During periods of low labour demand, job competition is tough which can result in longer-than-expected search durations Longhi and Taylor (2014).

(15)

Chapter 3

Methodology

In this chapter, the methods used in this thesis are presented. First, it is discussed how the degree of cyclicality of the different business sectors in the Netherlands is determined. After that, the models for assessing the labour market relevance of higher education graduates are discussed.

3.1 Cyclicality Analysis

Before analyzing the effect of the economic situation in the business sectors on the job search duration of higher education graduates, it is interesting to investigate the cyclical behaviour of these sectors. How sensitive is each sector to movements in the business cycle? This section outlines the method used to address this question.

Firstly, it is required to choose a macroeconomic variable that serves as an indicator for the economic climate in each sector. There is a variety of available macro-fundamentals that can serve this purpose. In this thesis, the Gross Domestic Product (GDP) growth rate has been chosen as an indicator of economic prosperity. It would be interesting to assess the robustness of the results to the choice of cyclical indicator. However, few data are available on sectorial level. Hence, one has to rely on the sectorial GDP growth rates.

When analyzing macroeconomic fluctuations, one can distinguish between two different ap-proaches. The classical approach involves an analysis of the total fluctuations in a time series over a given period of time, independent of the underlying nature of the change. The modern approach, on the other hand, focuses on the cyclical fluctuations in time series around their long run trends. These short-term cyclical fluctuations are identified through de-trending pro-cedures. An advantage of de-trending is that outcomes may be more robust when the underlying trend in the economy is separated out. A disadvantage of de-trending is that useful information might be lost (Rand and Tarp, 2002).

In recent studies, de-trending often occurs using a filter in order to obtain a nonstationary trend component and a stationary cyclical component of a time series. This de-trending en-sures that the trend component can be extracted from the time series and hence the cyclical

(16)

CHAPTER 3. METHODOLOGY 10 component can be observed separately. One filter that is commonly used in past literature is the Hodrick-Prescott (HP)-filter. For example, Fujita and Ramey (2009) use the HP-filter in their investigation of the cyclical behaviour of separation and job finding rates. Rand and Tarp (2002) examined the business cycle movements of different developing countries and applied the HP-filter prior to assessing the cyclical behaviour of each country.

The HP-filter decomposes a time series yt into a nonstationary trend gt and a stationary

residual component ct as follows:

yt= gt+ ct for t = 1, ..., T. (3.1)

Since gt and ct are unobservable, the problem boils down to how to extract gt from yt. The

HP-filter addresses this problem by estimating the unobservable time trend gt in the following

way: min {gt}T_t=−1 T X t=1 (yt− gt)2+ λ T X t=1 [(gt− gt−1) − (gt−1− gt−2)]2 ! . (3.2)

The first term is the sum of squared deviations which penalizes variety in the cyclical component

ct= yt− gt. The second term allocates a weight λ to gt that penalizes variablility in the linear

trend component (lack of smoothness). The larger the value of λ, the higher is the penalty and hence the smoother is the resulting time series, since more weight is allocated to the linear trend. Generally, high frequency data are more noisy relative to low frequency data. Thus, high frequency data require a higher value for the smoothing parameter λ. Hodrick and Prescott (1997) claim that for quarterly data, a choice of λ = 1600 is appropriate. For more details on the HP-filter, see Hodrick and Prescott (1997).

It may be useful to decide if filtering is actually needed, i.e. if the time series considered in the analysis have a linear trend component. Nonstationary time series are candidates for the HP-filter and nonstationarity can, for example, be tested using the Augmented-Dickey Fuller (ADF) test. The ADF testing procedure requires the estimation of the following equation:

φ(L)yt= α + t, with φ(L) = 1 − φ1L − ... − φpLp and Lkyt= yt−k, the so-called lag operator.

The nonstationairy testing problem is then formulated as: H0 : φ(1) = 0 (yt has a unit root)

against H1 : φ(1) > 0 (yt is stationary). For more details on how to use this estimation

procedure, see Dickey and Fuller (1979).

After HP-filtering the time series, the cyclical behaviour of the sectors can be examined. It is investigated how the sectors comove with the business cycle by examining the relationship between sector-level GDP growth rates and the national GDP growth rate (as a proxy for the business cycle). This relationship is measured in terms of the contemporaneous correlation coef-ficient. The sector is said to be procyclical, acyclical or countercyclical if the contemporaneous correlation coefficient between the sectorial GDP growh rate and the national GDP growth rate is positive, zero or negative respectively. Moreover, the sectorial GDP growth rate and the

national GDP growth rate are strongly correlated if 0.26 ≤ |ρ(yt, xt)| ≤ 1, weakly correlated if

(17)

CHAPTER 3. METHODOLOGY 11 national and sectorial GDP growth rates and ρ is their correlation coefficient. These boundaries are chosen taking the research of Rand and Tarp (2002); Stavarek (2013) as an example.

The cyclicality analysis will give an idea of the cyclical behaviour of the business sectors and how this may influence the job search duration of graduates when looking for a job in these respective sectors. The sectorial GDP growth rates are used as an indicator of the cyclicality of the sectors. Which business sectors are considered and what the results of the cyclicality analysis are can be found in Section 4.3.6.

3.2 Single-spell Duration Models

This section points out the two models used to investigate the labour market relevance of higher education graduates. Both models are so-called duration models. Duration models have been chosen to examine the labour market relevance of higher education graduates because the job search duration reflects the relationship between demand and supply of higher education graduates well. For instance, if the supply of higher education graduates exceeds the demand within a certain branch of business, it is on average more difficult to find a suitable job within that branch. In this case, the job search duration will typically be longer and the labour market relevance will typically be lower. If the demand for higher graduates exceeds the supply, the job search duration will typically be shorter and the labour market relevance will typically be higher. Hence, the labour market relevance of graduates can be determined by means of the job search duration.

There is a plethora of available duration models to choose from. One can, for example, use a parametric regression model in which one chooses a certain distribution function for the duration variable. The exponential distribution and the Weibull distribution are popular choices in this class of regression models. These models are relatively simple to estimate but impose a strong restriction on the data by assuming a certain distribution function. This will result in inconsistent parameter estimates if the distribution is misspecified. Hence, these models are not always preferable. To avoid this problem, one can choose a parametric functional form that is more flexible and thus provides some protection against misspecification. In this thesis, the piecewise constant hazard rate model has been used. This is an exponential hazard rate model where the constant hazard rate is allowed to vary within pre-defined time intervals. This approach requires a less than complete distributional specification. The second duration model that is used in the analysis is the Cox proportional hazard rate model. This model does not require the formulation of a distribution function of the time component of the model. The downside of this method is that, as a consequence, no time effects are estimated.

3.2.1 Piecewise Constant Hazard Rate Model

First, this section provides the basic definitions of a duration analysis. Consequently, it is described how the piecewise constant hazard rate model is estimated.

(18)

CHAPTER 3. METHODOLOGY 12 The variable of interest is the unemployment duration or job search duration, denoted T . The event of a failure, in this case, is the event that a person finds a job. The cumulative distribution function of T is F (t) and the density function is f (t) = dF (t)/dt. The probability that the spell length is less or equal to t, i.e. the probability that a person finds a job before or on time t, is then defined as

F (t) = P [T ≤ t] (3.3)

=

Z t

0

f (s)ds.

Another key concept of a duration analysis is the hazard function, which is the probability of leaving a state conditional on surviving to time t. In our case, this is the probability of finding a job at time t conditional on being unemployed until time t:

λ(t) = lim ∆t→0 P [t ≤ T < t + ∆t|T ≥ t] ∆t (3.4) = f (t) 1 − F (T ).

Another important function is the survival function S(t). It is formulated as the probability that the duration is greater than t:

S(t) = P [T > t] (3.5)

= 1 − F (t). Moreover, when integrating λ(t) it can be shown that

S(t) = exp − Z t 0 λ(s)ds . (3.6)

A final related function is the cumulative hazard function or intergrated hazard function: Λ(t) =

Z t

0

λ(s)ds (3.7)

= −lnS(t).

The piecewise constant hazard rate model is a special case of the proportional hazard model, with the conditional hazard rate λ(t|x, β, α) factored into separate functions:

λ(t|x, β, α) = λ0(t|α)φ(x, β), (3.8)

with φ(x, β) = exp(x0β) and λ0(t|α), the baseline hazard, a step function with M different

segments so that

λ0(t|α) = eαm, cm−1 ≤ t < cm, m = 1, ..., M, (3.9)

where c0 = 0, cM = ∞ and the other breakpoints c1, ..., cM −1are pre-specified. The parameters

(19)

CHAPTER 3. METHODOLOGY 13 in survival analysis because it permits coefficients to be easily interpretable and ensures that φ(x, β) > 0. Therefore, this specific form of φ(x, β) has been chosen in this research. Further explanation on why it permits easy interpretation can be found in Section 5.1.

The integrated baseline hazard is a step-function of the form

I(t|α) =

m−1

X

i=1

(ci− ci−1)eαi+ (t − cm−1)eαm. (3.10)

With use of (3.6) and (3.8), this leads to a survival function of the form

S(t|x, β, α) = exp −exp(x0β)I(t|α) , (3.11)

with I(t|α) as in (3.10) (Cameron and Trivedi, 2005; Murphy, 1996). Since in this research it is known in which month, but not the exact day on which the respondent has found a job, the most detailed specification of the job search duration is in months. Therefore, the length of the timepieces in this thesis is one month.

An important aspect of modeling unemployment spells is that the data are right-censored. This means that some unemployment spells in the sample are incomplete, indicating that these respondents have not found a job at the moment of observation. All that is known is that these spells will end some time in the interval (c, ∞) where c is the censoring time. For uncen-cored/completed spells, the full unemployment period is observed and the contribution of these observations to the likelihood is f (t|x, α, β). For the right-censored observations, it is known that the unemployment duration exceeds t. Therefore, the contribution of censored observations is P [T > t] = Z ∞ t f (u|x, θ)du (3.12) = 1 − F (t|x, θ) = S(t|x, θ),

with θ = (α0, β0)0. The contribution of the ith observation to the likelihood function can be

written as

f (ti|xi, θ)δiS(ti|xi, θ)1−δi, (3.13)

where δi is the right-censoring indicator with δi= 1 (no censoring) and δi= 0 (right-censoring).

With use of all these ingredients, the log-likelihood can be formulated as

ln L(θ) =

N

X

i=1

[δi ln f (ti|xi, θ) + (1 − δi) ln S(ti|xi, θ)], (3.14)

where N is the total number of observations. The maximum likelihood estimator (MLE) is the

solution ˆθ that maximizes (3.14).

3.2.2 Cox Proportional Hazard Rate Model

The second duration model used in this thesis is called the Cox proportional hazard rate model. Again, the starting point is the assumption of a proportional hazard rate with the conditional

(20)

CHAPTER 3. METHODOLOGY 14 hazard rate factored into two separate functions:

λ(t|x, β) = λ0(t)φ(x, β). (3.15)

As earlier, the function λ0(t) is called the baseline hazard rate and is a function of time. As

in the piecewise constant hazard rate model, it is assumed that φ(x, β) = exp(x0β). The

difference between the piecewise constant hazard rate model and the Cox proportional hazard rate model is that the Cox method does not specify the functional form of the baseline hazard

λ0(t). The Cox method estimates β but does not require simultaneous estimation of λ0(t). The

Cox proportional hazard rate model is useful as a robustness check of β because it allows one to compare the estimates of β of both models. The parameters α of the baseline hazard rate in the piecewise constant hazard rate model cannot be compared.

The setup of the Cox estimation method is as follows. Suppose that there are k ordered

discrete failure times t1 < t2 < ... < tj < ... < tk. The observations are categorized into those

that fail or are at risk at each failure time. Define the risk set R(tj) as the set of individuals who

are at risk of failing just before the jth ordered failure time tj, D(tj) as the set of individuals

that fail at time tj (i.e. the set of unemployment spells completed at time tj)) and dj the

number of individuals that fail at time tj (i.e. the number of spells completed at time tj). The

probability that the at-risk unemployment spell j actually ends at time tj is then equal to the

conditional probability of failure for spell j divided by the conditional probability that a spell

of any observation in the risk set R(tj) fails:

One can see that the baseline hazard rate λ0(tj) drops out of the probability and hence is not

identified.

Because the month, but not the day, in which the unemployment duration ends is observed, it is likely that the data include ties (i.e. more than one unemployment spell ends at a given

failure time). In this case, dj > 1 and the above probability (3.16) needs to be adjusted. There

are several methods for handling tied failures. Since the exact contributions to the likelihood become quite complicated with many tied values, a number of approximations is developed. The approximation for the likelihood contribution of spell j that is used in this thesis is the one developed by Breslow and Peto (see Cox and Oakes (1984)). This method states that because the order of failures is unknown, the largest set of spells at risk for each tied failure event is used. (3.16) then becomes:

P [Tj = tj|j ∈ R(tj)] ' Q m∈D(tj)φ(xm, β) h P l∈R(tj)φ(xl, β) idj. (3.17)

(21)

CHAPTER 3. METHODOLOGY 15 The partial likelihood function is then defined as the joint product of (3.17) over the k ordered failure times: Lp(β) = k Y j=1 Q m∈D(tj)φ(xm, β) h P l∈R(tj)φ(xl, β) idj. (3.18)

The estimated β is the result of minimizing the log partial likelihood function ln Lp:

ln Lp= k X j=1   X m∈D(tj) ln φ(xm, β) − dj ln   X l∈R(tj) φ(xl, β)    . (3.19)

Censored (incomplete) unemployment spells appear only in the last term of (3.19) because they

are not part of D(tj), the set of completed spells at time tj. For more comprehensive comments

on computing the partial likelihood function in a Cox PH model, see Cameron and Trivedi (2005) and Cleves et al. (2008).

(22)

Chapter 4

Data

This chapter explains how the data that are used in this thesis are collected. Subsequently, it provides a comprehensive description of the data. Before doing so, the higher education system in the Netherlands is explained briefly to provide the reader with some insight on the different Dutch higher education levels and their characteristics.

4.1 Higher Education System in the Netherlands

Before one can explain the higher education system in the Netherlands, the secondary education system should be discussed. The Dutch secondary education system consists of three levels. The lowest level is called Lower Vocational Education (LVE), a four-year program. The middle level is called Higher General Secondary Education (HGSE) and takes five years. Pre-university Education is the highest level in secondary education and takes six years. After obtaining an LVE degree, it is possible to proceed to HGSE and obtain an HGSE diploma after one or two additional years of education. The same holds true for the transition from HGSE to Pre-university Education (see Figure 4.1).

Higher education in the Netherlands consists of two levels. The lower level is called Higher Vocational Education (HVE) and the higher level is called University. There is a number of differences between the two levels. First, the required secondary education degree with which one is allowed to enroll in HVE and University is different. The required secondary education degree for HVE is at least HGSE, while the required degree for University is Pre-university Education. It is not possible to enter higher education with an LVE diploma. If one has an LVE diploma, one can proceed with Primary Vocational Education. It is, however, possible for LVE students to enter higher education through Primary Vocational Education (see Figure 4.1).

The second difference between HVE and University lies in the structure. Although both levels have a bachelor-master structure, it is less common for HVE students to proceed with an HVE master’s program after obtaining a bachelor’s degree than it is for university stu-dents. HVE students often enter the labour market after finishing their bachelor’s degree, while

(23)

CHAPTER 4. DATA 17 university bachelor students often decide to continue studying and to obtain a master’s degree. HVE students have the option to do a university master’s program after finishing their HVE bachelor’s program. Before proceding with a university master’s program, the HVE student has to complete a so-called pre-master. This is a linkup between HVE and University and usually takes one year (see Figure 4.1).

HVE students also have the option of entering University earlier in their educational career, namely after completing the first year of HVE. After succesfully completing all first-year courses, the HVE student has obtained a so-called foundation course, with which it is possible to proceed to a university bachelor program. In practice, however, university bachelor programs often have additional requirements which makes switching from HVE to University after the first year difficult. Moreover, not all university bachelor programs allow for this act.

Figure 4.1 provides a graphical description of what is discussed above. Note that it is not meant to provide a complete description of the secondary and higher education system in the Netherlands. Rather, this section is designed to point out the main characteristics of HVE and University.

Figure 4.1: Secondary and higher education system in the Netherlands

4.2 Data Collecting

The data used in this thesis are collected through a survey. Before 2005, the survey was

handed out in paper form. Since 2005, it is conducted online. The survey is distributed among newly-graduated higher education students in the Netherlands by SEO Economic Research. SEO Economic Research is a research agency affiliated with the University of Amsterdam that, among other things, examines the position of higher education graduates on the labour market each year since 1997. Their annual survey polls 20,000 graduates from Dutch higher education institutions who are requested to fill in the questionnaire. Students from more than one hundred higher education study programs are approached and they are surveyed approximately two years after graduating. The questionnaire is updated each year, meaning that some questions are removed and some new questions are added. Nevertheless, most of the questions stay the same over the years. The survey consists of approximately 60 questions which can be categorized into four blocks: educational career (secondary and higher education), search behaviour, current

(24)

CHAPTER 4. DATA 18 position on the labour market and personal characteristics. Most of the questions are multiple choice questions. The goal of the survey is to determine which study programs offer graduates the best opportunities on the labour market. Elsevier magazine publishes an annual theme edition which presents the most important results from the survey. At the same time, SEO Economic Research publishes a report that presents the scientific background for the survey.

Other data used in this thesis are macroeconomic data. The data are collected from the

Statitics Netherlands online database1. Specifically, time series of sectorial GDP growth rates,

the national GDP growth rate, the number of vacancies (sector-level and nationally) and the unemployed labour force are used (see Section 4.3.5 and Section 4.3.6).

4.3 Characteristics of the Data

In this section, a comprehensive description of the data is given. The section is divided into six subsections, each discussing one part of the data. The first four subsections describe the data from the survey and the last two subsections describe the macroeconomic data.

4.3.1 Socio-economic Characteristics

In Table 4.1, a number of socio-economic characteristics of the higher education graduates in the sample are shown. The following becomes apparent. Almost half of the respondents are male (44.4%). A small part of the sample (12.4%) has a non-Dutch background. A non-Dutch background signifies that either the graduate himself or at least one parent was born outside the Netherlands. The average age at which students begin their study program is 19.4 years. On average, higher education students graduate when they are 24.2 years old. The educational level of the parents shows that mothers are still lower educated than fathers. More than 40% of mothers have primary education or primary vocational education, against 25.9% of fathers. On higher education level, this difference is also visible: about 30% of mothers have a higher education degree. For fathers this is almost 40%. Moreover, only 7.1% of mothers have a university degree against 20.2% of fathers.

The graduates are not evenly distributed across the country. More than half of the graduates live in the western part of the Netherlands. The western part consists of the provinces Noord-Holland, Zuid-Holland and Utrecht. 19.7% settle down in the south of the country (consisting of the provinces Noord-Brabant, Limburg and Zeeland), followed by 19.0% in the east (Flevoland, Gelderland and Overijssel). Only 8.6% of the graduates live in the nothern provinces (Drenthe, Friesland and Groningen). The living situation (at the moment the survey was held) shows three large groups: together with a partner and no children (44.1%), alone without children (30.5%) and living with their parents (15.5%). Only a small part of the graduates has children (3.3%) and 6.6% have another living situation than previously specified.

(25)

CHAPTER 4. DATA 19 Table 4.1: Socio-economic characteristics

percentage man 44.4%

percentage non-Dutch background* 12.4%

age at the start of the study program 19.4

age when graduating 24.2

educational level of: mother father

primary eductation 4.8% 4.9%

primary vocational education 36.4% 21.0%

secondary education 28.5% 25.4%

higher vocational education 23.2% 28.4%

university 7.1% 20.2% region of living West NL 52.7% South NL 19.7% East NL 19.0% North NL 8.6% living situation

with his/her parents 15.5%

alone (no children) 30.5%

alone (with children) 0.1%

together with partner (no children) 44.1%

together with partner (with children) 3.2%

other 6.6%

*The student himself or at least one parent was born outside the Netherlands.

4.3.2 Educational Career

Table 4.2 shows the characteristics of the educational career of the graduates in the sample. More than half of the graduates have a university degree (53.6%). The other part of the sample has a degree in Higher Vocational Education (HVE). The remainder of the table shows statistics

for HVE and university students separately. Firstly, consider the educational sectors2. The

distribution of students among the four sectors is more or less the same for HVE and university students. Most of the students have finished a gamma study program (31.1% of HVE students and 32.4% of university students). Medical students make up the smallest part of the sample (15.9% of HVE students and 14.5% of university students).

University students finish their degree with an average grade of 7.3. This is slightly higher than that of HVE students, who graduate with an average of 7.1. On average, HVE students complete their degree after 4.0 years of studying. University students take much longer to obtain

2_{A list with all the study programs and to which educational sector they are entitled can be found in Table}

(26)

CHAPTER 4. DATA 20

a degree, namely 5.5 years3.

Lastly, extracurricular activities that a student can undertake during his educational career are shown. Some large differences are visible between HVE and university students. In Higher Vocational Education, it is much less common to be a member of a study association than when enrolled in university: 10.6% of HVE students are a member of a study association against 71.4% of university students. 13.9% of HVE students and 33.0% of university students are a member of a students’ association. A students’ association is the equivalent of a fraternity or sorority (not study related). Students often have the opportunity to apply for a position on the board of, for example, a study association or sports club. 51.1% of university students take a position on a board, against 26.7% of HVE students. During their studies, approximately one-third of students (HVE and university) works in a field related to their study program. Lastly, 13.5% of HVE students and 24.9% of university students have studied abroad as part of their higher education degree.

Table 4.2: Characteristrics of educational career

percentage HVE graduates 46.4%

percentage University graduates 53.6%

HVE University educational sector Alpha 26.7% 22.7% Beta 26.3% 30.4% Gamma 31.1% 32.4% Medical 15.9% 14.5% study performance

average grade in higher education 7.1 7.3

study duration (years) 4.0 5.5

extracurricular activities

member of study association 10.6% 71.4%

member of students’ association 13.8% 33.0%

experience as a board member 26.7% 51.2%

work experience in own study field 32.5% 32.6%

studied abroad 13.5% 24.9%

4.3.3 Searching Behaviour

In this section, the searching behaviour of higher education students is examined. Specifically, the methods students use when searching for a job and when they actively start searching are analyzed.

3

(27)

CHAPTER 4. DATA 21 First, consider the statistics shown in Table 4.3. In the sample, 19.5% of the graduates were asked for their current job. The average number of submitted job applications until a job was found, or, in case of censored data until the moment the survey was held, is 12.5. Also shown in the table are the different searching methods used by the graduates. The most popular searching method is looking for existing vacancies and writing application letters (70.3%). Other favoured methods are looking for vacancies on the internet and writing open application letters (54.0% and 53.0%, respectively). Approaching an intermediary (for example an employment agency) is done by 41.5% of the students and 41.1% use their own informal network of family and friends to find a job. Applying as an intern first is done by 27.9%. Other methods are used by 5.7%.

It is of course possible to use more than one searching method when looking for a job. Figure 4.2 displays the number of job searching methods used by the students in the sample. The average number of used searching methods is 3.8. 11.6% say that they did not use any searching methods. These students are most likely asked for the job. More than half of the students use 1, 2, 3 or 4 searching methods. About 10% use more than seven searching methods. Figure 4.3 shows when students start searching for a job. A negative number means that the student started searching before graduating, a positive number means that the student started searching after he graduated. The graduates were asked in which year and month they started searching for a job. The graduation dates have been made available by the Ministery of Education, Culture and Science. On average, students start searching for a job a little before their graduation date (0.21 months before). Furthermore, 47% start searching before graduation, while 53% start searching in or after the graduation month.

Table 4.3: Searching behaviour

percentage that was asked for current job 19.5%

number of job applications* 12.5

searching methods

looking for vacancies 70.3%

looking for vacancies online 54.0%

having open interviews 53.0%

using an intermediary 41.5%

using own informal network 41.4%

doing an internship 27.9%

other 5.7%

*Until a job was found or, in case of incomplete searching spells, until the moment the survey was held.

(28)

CHAPTER 4. DATA 22

Figure 4.2: Number of different searching methods

used

Figure 4.3: Distribution of searching moment around graduation

4.3.4 Unemployment Duration

In this thesis, the job search duration or unemployment duration is defined as the time spent searching for a job that matches the graduate’s educational level. The job searching period starts at the moment the individual begins searching for a job in an active manner. This can happen before or after the graduation date (see Figure 4.3). As stated earlier, the job searching moment (month and year) is known because it was asked in one of the survey questions. When a person skipped this question, the start of the job searching spell has been put equal to the graduation date minus the average number of months before graduation a person actively starts searching for a job (0.21 months). The graduation date is known for all individuals in the sample. The unemployment spell is completed if the individual is employed at a job at his/her educational level. Hence, for university graduates this is a job that requires an academic background, for HVE graduates this is a job that requires at least HVE. This is known since respondents were asked what level of education their employer required for their job. Since it is known in which month the individual started searching and in which month the individual has found a job, the unemployment duration is measured in monthly intervals.

Definition 1 Unemployment Spell The unemployment duration spans the period between the moment a person actively starts looking for a job and the moment a person is employed at a job that matches his/her level of education.

Two different groups can be distinguished in the data: individuals who were able to find a fitting job and individuals who did not find a fitting job at the moment of observation (the date the survey was held). People from the first group have a completed unemployment spell, while people from the second group have an uncompleted spell. The number of completed spells in the sample is 79.553 and the number of incompleted spells is 9.244 (see Table 4.4). A histogram of the job search duration is presented in Figure 4.4. It is apparent that the spell lengths range from 1 month to 36 months. The average search duration in the sample is 7.06 months and 18.85% find a fitting job in the first month. Moreover, a substantial amount of people (64.7%)

(29)

CHAPTER 4. DATA 23 have a job search duration of 6 months or less. Only a small part of the sample has a job searching spell longer than 30 months (less than 1%).

Table 4.4: Number of complete and incomplete unem-ployment spells according to Definition 1

Spell Number in the sample

Completed spell 79.553

Incompleted spell 9.244

Figure 4.4: Histogram of the job search duration according to Definition 1

The duration data can be shown graphically as an empirical survival function. Note that

“surviving” until time tj in this context means still being unemployed at time tj. The

Kaplan-Meier survival function estimate, which is the discrete sample analogue of S(t) in (3.5) is defined as follows: ˆ S(t) = Y j|tj≤t (1 − ˆλj) (4.1) = Y j|tj≤t rj− dj rj ,

where rj equals the number of spells at risk just before time tj and djequals the number of spells

ending at time tj. The confidence intervals of ˆS(t) are based on the variance of ln(−ln( ˆS(t)))

rather than on the variance of ˆS(t) itself. The reason for this is that this transformation ensures

that the confidence bands lie in the range of the survivor function (between zero and one). The

100(1 − α)% confidence interval of ˆS(t) is equal to

ˆ_S(t)exp(−zα/2σ(t))ˆ _{, ˆ}_S(t)exp(zα/2ˆσ(t))

, (4.2)

where zα/2is the (1 − α/2) quantile of the standard normal distribution and ˆσ(t) is the standard

deviation of ln(−ln( ˆS(t))). The latter is computed using the estimated variance of ln(−ln( ˆS(t))):

ˆ σ2(t) = P j|tj≤t dj rj(rj−dj) h P j|tj≤tln _r j−dj dj i2. (4.3)

(30)

CHAPTER 4. DATA 24 Moreover, α = 0.05 is chosen. The resulting empirical survival functions, together with their 95% confidence bands, are shown in Figure 4.5. The vertical axis shows the proportion of unemployment spells that are still in progress after a stated number of months. The blue line describes the survival function of HVE students, the red line illustrates the survival function of university students. As one can see, the functions are decreasing step functions with jumps at

each discrete failure time tj. As expected, the functions start at one and monotonically decrease

towards zero. However, since a number of the spells in the sample is incomplete (censored), the survival probability is (small but) positive in the last months. Moreover, the functions decline rapidly at first and then slowly towards the end. From Figure 4.1 it can be seen that the survival function for university students lies above the one for HVE students. Overall, it appears that university students face more difficulty in finding a fitting job than HVE students.

The same analysis has been carried out for the different educational sectors. The resulting survival functions are shown in Figure 4.6. Note that HVE and university students are bundled in this figure. The survival functions again start at one and monotonically decrease towards zero, but will never reach zero due to censoring. The functions for the educational sectors Alpha and Gamma are more or less the same in early stages of unemployment (up to the fifth month). After that it becomes clear that alpha students have the highest survival probability/highest probability of remaining unemployed, followed by gamma students. The survival curves of beta and medical students decline more rapidly, indicating that they end their unemployment spell more quickly than the other educational sectors. This is also in line with what is expected from theory. Alpha students usually have the most difficulty finding a job. Beta students have an appealing educational background and their technical skills are in high demand. Medical students also find a job quickly because their spot on the market is more or less predetermined by policymakers in their branche.

0 .25 .5 .75 1 Survival Probability 0 10 20 30 40

Unemployment duration in months

95% CI HVE HVE 95% CI University University

Kaplan−Meier Survival Function Estimate

Figure 4.5: Kaplan-Meier estimates of survival func-tions for HVE students and University students

0 .25 .5 .75 1 Survival Probability 0 10 20 30 40

95% CI Alpha Alpha 95% CI Beta Beta 95% CI Gamma Gamma 95% CI Medical Medical

Kaplan−Meier Survival Function Estimate

Figure 4.6: Kaplan-Meier estimates of survival func-tions for the different educational sectors (HVE and uni-versity together)

(31)

CHAPTER 4. DATA 25 Another empirical funtion is the Nelson-Aalen cumulative hazard function. It is the sample analogue of Λ(t) in (3.7) and defined as

ˆ Λ(t) = X j|tj≤t ˆ λj = X j|tj≤t dj rj (4.4)

where rj equals the number of spells at risk just before time tj and dj equals the number of

spells ending at time tj. Pointwise confidence intervals of ˆΛ(t) are based on the variance ˆξ2(t)

of ln ˆΛ(t):

ˆ

ξ2(t) = V [ ˆˆ Λ(t)]

ˆ

Λ(t)2 . (4.5)

The 100(1 − α)% confidence interval for the Nelson-Aalen estimator ˆΛ(t) is then equal to

ˆ_{Λ(t)exp(−z}

α/2ξ(t)), ˆˆ Λ(t)exp(zα/2ξ(t))ˆ

, (4.6)

where α is chosen to be 0.05. The results are shown in Figure 4.7 and 4.8. The cumulative hazard functions exhibit the expected pattern: the cumulative hazard rate of HVE students increases at a higher speed compared to university students. The hazard rate for medical students in-creases most rapidly, followed by beta students, then gamma students and lastly alpha students.

0 1 2 3 4 5 Cumulative Hazard 0 10 20 30 40

95% CI HVE HVE 95% CI University University

Nelson−Aalen Cumulative Hazard Function Estimate

Figure 4.7: Unemployment duration: Nelson-Aalen es-timates of cumulative hazard functions for HVE students and University students

0 2 4 6 Cumulative Hazard 0 10 20 30 40

95% CI Alpha Alpha 95% CI Beta Beta 95% CI Gamma Gamma 95% CI Medical Medical

Nelson−Aalen Cumulative Hazard Function Estimate

Figure 4.8: Unemployment duration: Nelson-Aalen es-timates of cumulative hazard functions for the different educational sectors (HVE and university together)

4.3.5 Economic Setting: 1996–2013

The analysis of the thesis spans from 1996 to 2013: the first group of students actively starts searching for a job in 1996, the last group of students starts searching in 2013. It is therefore interesting to briefly sketch the economic setting in this period. Figure 4.9 shows data for several macroeconomic variables from 1996:Q1 to 2015:Q3. Unfortunately, data for the number of vacancies is not available before 1997 (see Figure 4.9b). Moreover, quarterly data for the unemployed labour force are only available from 2003 onwards. In the period from 1996 to 2002, the data are available at annual intervals and have been intrapolated linearly to obtain quarterly data (see Figure 4.9a). Figure 4.9c shows the number of vacancies per unemployed

(32)

CHAPTER 4. DATA 26 person. This is a common measurement for shortage or tension on the labour market. Figure

4.9d displays the national GDP growth rate4.

In the period from 1996 to 2001, the unemployed labour force decreased from 501,000 in March 1996 to 256,000 in December 2000 (see Figure 4.9a). During this period, the number of vacancies increased from 80,000 in March 1997 to 201,000 in June 2001. Hence, the number of vacancies more than doubled. After 2000, the economy has passed its pinnacle: the unemployed labour force expands, while the number of vacancies decreases. This may be a response to the losses on the stock market at the end of 2000/begin 2001. The decreasing GDP growth rate in this period is primarily due to decreasing consumer confidence (CBS-Persdienst, 2001a). Later that year, the business cycle is largely influenced by the terrorist attack in the Unites States of America on September 11. Stock prices decreased heavily and the consumer confidence in many countries declined because of the attack (CBS-Persdienst, 2001b). The financial crisis is also clearly visible in the data. In 2008, a large drop in the number of vacancies, the number of vacancies per unemployed person and the GDP growth rate is visible. The unemployed labour force expanded in 2008.

4.3.6 Cyclicality of the Business Sectors

The effect of the economic climate on job search duration in the various business sectors is an important topic of this thesis. It is therefore interesting to investigate the cyclical behaviour of the sectors, i.e. to which extent they comove with the business cycle. Some sectors may be strongly procyclical, which can have negative consequences for the job search duration when the overall economy is in a recession. Other sectors may be countercyclical or are not influenced by the overall state of the economy (acyclical). This can be investigated with the use of a correlation analysis on de-trended sectorial and national GDP growth rates.

The dataset for the cyclicality analysis consists of quarterly data for the GDP growth rates of several business sectors and of the national GDP growth rate over the period 1996:Q1 -2015:Q2. For the sectors Education and Public administration, data are available from 2005:Q1 to 2015:Q2. The time series are collected from the online database of Statistics Netherlands (also known as CBS) and formulated as the percentage change in the value of the GDP in comparison to one year earlier. In Table 4.5, one can find the time series considered in the cyclicality analysis. The Utilities sector is a combination of two other sectors, namely Energy and Gas Supply and Water Supply and Waste Management. The GDP growth rate of Utilities is the average of the two GDP growth rates of these sectors. It has been chosen to define this sector as such because it coincides with the sector Utilities as formulated in the survey.

Stationarity of the time series is tested using the ADF-test. The lag-length p in the test-ing equation (see Section 3.1) is chosen based on the Schwartz Information Criterion (SIC). Table 4.6 displays the results. The number of lags is shown in parentheses. Based on a

signifi-4

The GDP growth rate is defined as the percentage change in the value of the GDP compared to the value one year ago

(33)

CHAPTER 4. DATA 27 Figure 4.9: The development of the unemployed labour force, the number of vacancies, the number of vacancies per unemployed and the GDP growth rate

(a)Unemployed labour force (x1000) (b)Number of vacancies (x1000)

(c)Vacancies per unemployed (v/u) (d)GDP growth rate (%)

Source: CBS Statline. The data shown are national data.

cance level of 5%, the GDP growth rates of Agriculture, Manufacturing, Financial Institutions, Transportation and Storage and Wholesale and Retail Trade are stationary time series, while the GDP growth rates of Accomodation and Food Services, Construction, Utilities, Culture, Recreation and Other Services, Business Services, Health and Social Work Activities, Infor-mation and Communication, Education and Public Administration are nonstationary, together with the national GDP growth rate. These results suggest that only the latter should be HP-filtered, while the former can stay as they are. Figure 4.10 shows the time series before and after HP-filtering. Indeed, for the stationary time series, filtering has almost no effect.

Contemporaneous correlations between the sectorial GDP growth rates and the national GDP growth rate are calculated both before and after HP-filtering the data. Table 4.7 shows the results. For six business sectors, the cyclicality conclusion differs before and after filtering. Holding on to the definitions of cyclicality given in Section 3.1, Culture, Recreation, Other Services is weakly procyclical before filtering and acyclical after filtering. The construction industry is strongly procyclical before filtering and weakly procyclical after filtering. Public Administration is acyclical if the time series is not filtered and weakly countercyclical after filtering. Utilities is weakly countercyclical before filtering but strongly countercyclical after

Job search duration of higher education graduates in the Netherlands

Master’s Thesis