Rather enrolled than unemployed : an empirical analysis of the impact of the Great Recession on enrolment in post-compulsory education in The Netherlands

Hele tekst

(1)Rather enrolled than unemployed An empirical analysis of the impact of the Great Recession on enrolment in post-compulsory education in The Netherlands. By Reineke Davidsz Supervised by Aaron Kamm University of Amsterdam Bachelor in Economics and Business, Specialization in Economics Student number: 10272763 1 July 2014. Abstract This paper investigates the effect of the economic crisis in 2008 on enrolment in higher education in The Netherlands. I use time-series data from 2000-2012 and include a time dummy as a measure of the crisis. Unfortunately the estimated coefficients are insignificant. Previous research and public discussions point to an increase in enrolment rates due to the crisis, however the results in this empirical research suggests a negative impact of the Great Recession on overall enrolment rates. Moreover, the effect of the crisis differs among fields of study, leading to lower enrolment for studies with low returns to schooling and higher enrolment rates for studies with high returns to schooling..

(2) 1 Introduction 1.1 Background Deciding whether to enrol for higher education is a pressing question for the youth. Moreover, deciding which field of study to choose might be even more challenging. Making this choice in times of a bad economy is likely to influence a young adult’s decision. As labour market opportunities become scarce, the youth tends to enrol in post-compulsory education due to lower opportunity costs and invest in skills to decrease expected future unemployment. De Volkskrant (Bouma and Hosselet, 2014), a Dutch newspaper, confirms this theory and writes that enrolment rates increased as an effect of the economic crisis. Using time-series data I estimate a negative effect of the crisis on enrolment rates. But previous literature and public discussions imply a positive relationship. Moreover, results show that effects differ across fields of study indicating that the recession has different effects on enrolment rates among studies. None of the estimated coefficients, unfortunately, are significantly different from zero due to a small sample size. On the one hand, it is often argued that the opportunity cost of education is reduced when recessions lead to high unemployment rates (see McVicar and Rice, 2001; Barr and Turner, 2012; Petrongolo and San Segundo, 2002; Tumino and Taylor, 2013). Clark estimates positive, statistically significant and large in magnitude youth unemployment effects on enrolment in post-compulsory education in England (2011, p. 539). Also Barr and Turner state that their analysis of the college enrolment response to the Great Recession shows an unambiguous and substantial link between adverse local economic conditions, as measured by the unemployment rate, and college enrolment (2012, p. 26). Also, Dutch news articles conclude an increase in enrolment rates due to the crisis confirming a positive relationship (Middelburg, 2009; van der Hulst, 2014; Bouma and Hosselet, 2014). On the other hand, higher unemployment rates, implied by the crisis, leads to expected future unemployment, which reduces the returns to education and hence the demand for schooling (Micklewright, Pearson and Smith, 1990, p. 163). Moreover, higher unemployment rates may imply lower budgets, which discourages post-compulsory education consumption. These two effects describe a negative impact of the recession on enrolment. Furthermore, the impact of the crisis is different among fields of study. According to De Volkskrant (Bouma and Hosselet, 2014), students more often choose studies with a high job security in the future. Arcidiacono R. Davidsz / University of Amsterdam / 2014. 2.

(3) (2004), Jacobs (2002) and Hamermesh and Donald (2008) show large differences in average earnings for different fields of study. Moreover, unemployment rates across fields of study likely differ as well, however this data is unfortunately unavailable. This, and other, previously performed research motivated me to examine in detail how the economic climate affects young adults and their decisions early in their career. Specifically, I look at the effects of the economic crisis in 2008 on enrolment in postcompulsory studies in The Netherlands. The two possible effects that I focus on are the overall enrolment rates in bachelor studies and secondly, the enrolment rates amongst studies in the Netherlands. In the following subsection an explanation of the Dutch educational system is given. Section 2 includes an overview of related literature. The data and empirical model are described in section 3. Section 4 shows the results of empirical research and section 5 concludes with a summary and discussion. 1.2 Dutch educational system Compulsory education in the Netherlands is from five to sixteen years old. Students who do not have a secondary school diploma at the age of sixteen must continue until they are eighteen years old. Most children enter the school system at the age of four and are enrolled for eight years. At the end of these eight years all children take a state exam, which determines their level of education in secondary school. Secondary school entails three different levels of education: lower vocational education, higher general secondary education and pre-university education. These different levels take four, five and six years respectively. Students specialize in their senior years in one out of four fields of study, namely Culture and Society, Economics and Society, Nature and Health and Nature and Technology. After completion students can enrol for post-compulsory education, which is split into intermediate vocational training, upper vocational training and academic training. The latter two are referred to as higher education or bachelor programmes. In my research I focus on enrolment for these two programmes and also differentiate between these two levels to see if there is more of an effect of the crisis for one of them. Upon completion of a bachelor, a student can choose to enrol for a master, however this is not the topic of my research.. R. Davidsz / University of Amsterdam / 2014. 3.

(4) 2 Related literature ‘Other things equal, education may be more attractive when the local youth labour market is weak’ (Clark, 2011, p. 523). As Clark states, there may be a relationship between recessions and enrolment. Increased unemployment is a result of the economic crisis and is, therefore, often used as a measure of the crisis itself in previous literature. However, in this empirical research a time dummy is added to fully capture the effect of the crisis on enrolment rates. This section gives an analysis of research on effects of unemployment on wellbeing and its effect on overall enrolment in bachelors. Moreover, it discusses literature on other variables affecting enrolment rates and also a student’s choice amongst studies. The recession of 2008 had a major impact on the youth around the world. Bell and Blanchflower (2011, p. 245) state that unemployment increased more rapidly among the young during the Great Recession in most OECD countries. High unemployment has large effects on people. As discussed by Clark and Oswald (1994), unemployed people in Great Britain in 1991 had much lower levels of mental wellbeing than those in work. Also Clark et al (2001) consider the psychological impact of past unemployment using German panel data. They find that current unemployment is associated with sharply lower levels of subjective wellbeing (p. 237). Moreover, they find that past unemployment reduces the wellbeing of those who are currently in work, in other words past unemployment scars. Knabe and Rätzel (2009, p. 292) elaborate on this ‘scarring’ effect. Their results show that the effect of past unemployment can be explained best through its effect on people’s fear of future unemployment. They conclude that past unemployment ‘scars’ because it ‘scares’. Since youth unemployment causes negative effects on wellbeing it is expected that the youth avoid entering a bad labour market. The lack of employment opportunities for them in the short run leads to relatively small opportunity costs of continuing post-compulsory education, thereby increasing enrolment in bachelors (see McVicar and Rice, 2001; Barr and Turner, 2012; Petrongolo and San Segundo, 2002; Tumino and Taylor, 2013). Consistently Clark (2011) estimates positive, statistically significant and large in magnitude youth unemployment effects on enrolment in post-compulsory education in England using regional panel data. Likewise, McVigar and Rice (2001) find that unemployment has small, but significant, positive effect on rates of participation based on time-series evidence from England and Wales. Rice (1998) finds consistently with time-series evidence that participation rates in post-compulsory education are positively related to the unemployment R. Davidsz / University of Amsterdam / 2014. 4.

(5) rate. Tumino and Turner (2013) use UK panel data and focus on school leaving decisions rather than enrolment decision. According to their estimates, an increase in youth unemployment decreases the probability of young people leaving school. Concerning the United States, recent research by Barr and Turner uses US panel data and conformingly concludes their research with the following: ‘Our analysis of the college enrollment response to the Great Recession shows an unambiguous and substantial link between adverse local economic conditions, as measured by the unemployment rate, and college enrollment’ (2012, p. 26). Bell and Blanchfower (2011) use US and UK case studies and they too find positive links between adverse local economic conditions and enrolment. Also older research by Betts and McFarland (1995) find a positive relationship between unemployment rates and enrolment using a US community college data set. Such a positive relationship is also confirmed by Corman’s US micro data (1993). There is research which sheds light on other countries too. For example, Albert (2000) finds a positive impact of unemployment on the demand for higher education in Spain. Also evidence suggests that high youth unemployment rates made some contribution to the rise in post-compulsory education in Spain, as concluded by Petrongolo and San Segundo (2002). The most robust finding of the study of four different countries, England and Wales, Germany, The Netherlands and Sweden, by McIntosh (2001) is the small, positive effect of youth unemployment levels on participation rates. However, there should be opposite effects of a recession on enrolment according to Micklewright, Pearson and Smith (1990, p. 163). They name two effects, namely the effects of parental unemployment and secondly, the expected future unemployment effect. Parents’ unemployment leads to lower household income and this may cause their children’s educational demand to fall for three reasons. First, the consumption demand for education will be lower, second a reduction in first period income reduces investment demand for education and third lower family incomes may restrict access to credit with further deleterious impact on education demand. Moreover, the expected future unemployment effect arises as unemployment rates increase in times of a bad economy and, therefore, the probability of future unemployment rises too. This reduces the rate of return to education and hence education demand. However, those who are higher educated may be less affected by increased unemployment rates and this may encourage further schooling. Unemployment rates among skilled workers are typically lower than among the unskilled, and therefore, schooling can help to escape future unemployment (Canton and de Jong, 2005). However, R. Davidsz / University of Amsterdam / 2014. 5.

(6) Canton and de Jong analyse the demand for higher education in the Netherlands and still find weak and insignificant effects of unemployment on student enrolment. Also Micklewright et al (1990) failed to confirm a positive relationship between unemployment and enrolment rates, as their UK data does not suggest that unemployment led to the decrease in leaving rates by 5 percentage points during 1978-’84. The failure to confirm a positive impact of unemployment on enrolment, leads to investigating other factors that might affect post-compulsory education decisions. One of these is financial support. Dolton and Lin’s research (2011) shows as a key result that less generous student financial support arrangements have had a significant negative impact on university enrolment based on UK evidence. Research from the Netherlands also shows a positive contribution of financial support to student enrolment (Canton and de Jong, 2005). In the US as well it is shown that aid can have a positive impact on college enrolment (Dynarski, 1999). Another factor that is discussed is the level of tuition fees. Canton and de Jong state that it is broadly supported in the international literature that students are rather insensitive to tuition fee changes, however it is difficult to compare price elasticities amongst different countries because of substantial variation in tuition fee levels (2005, p. 653). Household income may also influence post-compulsory education decisions due to credit constraints (see Shea, 2000). However, Micklewrite et al (1990) find a small effect and McVigar and Rice (2001) find no significant effect of household income on enrolment rates. Moreover, we see that enrolment rates may increase as skills are becoming more and more important in knowledge-based societies like The Netherlands (Canton and de Jong, 2005). It is therefore plausible that young adults enrol in higher education in order to remain competitive on the labour market. The discussion so far is based on enrolment rates for all studies, however the recession may have different effects on enrolment rates among different fields of study. Centraal Bureau voor de Statistiek (CBS), Statistics Netherlands, divides fields of study into eight main groups based on the International Standard Classification of Education (ISCED). In 1997 Unesco developed this classification in order to internationally compare education statistics and indicators. The eight groups consist of (1) Education, (2) Humanities and arts, (3) Social sciences, business and law, (4) Science, (5) Engineering, manufacturing and construction, (6) Agriculture, (7) Health and welfare, and (8) Services. Significant elasticities show responsiveness of male students in Science to changes in tuition fees, and the responsiveness of male and female Science and Health and welfare students and male R. Davidsz / University of Amsterdam / 2014. 6.

(7) Engineering, manufacturing and construction students to changes in financial support (Canton and de Jong, 2005, p. 660). This indicates that changes in variables have different effects among studies. Moreover, research shows that high average earnings studies are Science and Social sciences, business and law (Arcidiacono, 2004; Jacobs, 2002). Whereas Education is considered as a low average earnings study (Arcidiacono, 2004; Hamermesh and Donald, 2008). Besides the range in average earnings there may also be a difference in unemployment among studies (Canton and de Jong, 2005, p. 660). This might lead to different effects of the recession on enrolment rates among studies. Next to being an important topic in the academic world, enrolment rates in The Netherlands is discussed in the news. In 2009 Het Parool writes that enrolment increased with 25% as a result of the economic crisis (Middelburg, 2009). At the University of Amsterdam there was even an increase of 45% between 2008 and 2009. According to the VSNU, a union of Dutch universities, the economic crisis is the cause of the substantial increase in enrolment. The reason is that the youth wants to continue studying since the labour market opportunities are scarce. This is in line with empirical studies concluding an increase in enrolment rates during a crisis Moreover, recently De Volkskrant writes that students more often choose studies with a good chance to be employed (Bouma and Hosselet, 2014). These are studies in Engineering, manufacturing and construction, Health and welfare and Agriculture. Studies with a low chance of becoming employed such as Humanities and arts are less popular. This confirms that enrolment rates are differently affected by the crisis due to ranging returns to schooling. This empirical research is inspired by the public discussion on enrolment rates in the Netherlands after the economic crisis and builds on previous literature on this topic. Like other previous studies I use time-series data in order to assess the impact of the recession on enrolment in bachelors. Previous econometric models are used as examples for my empirical model (Clark, 2011; Canton and de Jong, 2005), which is discussed in the following section. 3 Data in context and empirical model 3.1 Data in context This research uses data provided by Centraal Bureau voor de Statistiek (CBS), Statistics Netherlands. This organization provides statistical information about much R. Davidsz / University of Amsterdam / 2014. 7.

(8) information related to the topic of this research, such as enrolment, youth unemployment and the labour force. Data from 2000-2012 is used since this includes multiple pre-crisis years as well as years during the crisis that started around 2008. Unfortunately there is no data available for a longer time-period. The number of observations for this analysis is thirteen. This is not a large number and this leads to low statistical power. It is therefore difficult to detect a significant effect even if there was an effect of the crisis on enrolment. Below follows a descriptive analysis of the data. This in combination with findings of previous literature will help with interpreting the results of my empirical research. It discusses how key variables are measured, what the expected relationship is to enrolment rates and how the results can be interpreted. A summary of the descriptive statistics of the data is presented in Table 1. Enrolment rates: The dependent variable is enrolment rates in higher education. This is calculated as the ratio of total enrolment in higher education to youth population. The numerator represents all first-year students in higher education enrolled for a bachelor programme in any field of study in each year. Only students that make the transition from secondary school to higher education are included because they made the decision to continue post-compulsory education in that specific year. The periods 2000/’01-2012/’13 are included. The denominator includes the population between 15 and 25 years old in each year. Youth unemployment rate: Over the 2000-2012 time period youth unemployment has always been well above any other age category’s unemployment rate. In 2004 youth unemployment was at its peak of this time-period. It decreased until to 2008, but never reached its lowest point from 2000. The economic crisis hit many economies and had a large impact on unemployment, especially youth unemployment. As seen from figure 1, this rapid increase of youth unemployment, represented by the 15-25 years old age category, was also present in The Netherlands. The increase in unemployment rate after the 2008’ crisis was highest for the youth, namely an increase from 8.4% in 2008 to 12.6% in 2012 whereas other age groups saw an increase of only around 3% to maximum 6.5%. The correlations between unemployment rates among different age groups are high; see Table A in Appendix A. This means that they largely move together, some more than others. Due to this high correlation, the model does not include a measure of adult unemployment. R. Davidsz / University of Amsterdam / 2014. 8.

(9) Figure 1: Unemployment per age group in the Netherlands 2000-2012. Source: CBS Unemployment rate (%). 14. 12. 10. 8. 6. 4. 2. 0. 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012. 15-25 years old. 25-35 years old. 45-55 years old. 55-65 years old. 35-45 years old. CBS provides annual youth unemployment rates. These are constructed as the ratio of the unemployed labour force to the total labour force amongst 15-25 years old. The labour force is defined as all persons between 15-65 years old who work at least twelve hours a week, accepted a job and are going to work for at least twelve hours a week or are willing to work at least twelve hours a week and are looking for such a job. On the one hand, the expected relationship is positive due to low opportunity costs and escaping future unemployment. On the other hand, the impact may be negative due to budget constraints and lower returns to schooling if students foresee unemployment in the future. A potential problem with this measurement of youth unemployment is that when more young adults choose to enrol in higher education, less young adults are in the labour force, which increases the youth unemployment rate. This makes it difficult to conclude based on the estimated coefficient of this independent variable as it may be biased. This has to be acknowledged as a caveat of the model but cannot be avoided since this is the way unemployment rates are measured in The Netherlands. Figure 2 shows a long-term picture of enrolment rates and youth unemployment rates from 1996-2012. The correlation between these two variables is 0.05, which means that they barely move together. Enrolment rates have mostly been increasing, apart from the decreasing rates between 2000 and 2002, and 2009 and 2010. On the other hand, youth unemployment has fluctuated over time, showing decreases as well as increases over time. In this empirical. R. Davidsz / University of Amsterdam / 2014. 9.

(10) model we use the data from 2000-2012 since data on other independent variables is limited to this time-period. Figure 2: Overall enrolment rates and the youth unemployment rate in the Netherlands 1996-. 3,50. 14. Enrolment Rate (%). 12. 10. 3,00. 8. 6. 2,50. 4. 2. 2,00. Youth Unempkloyment Rate (%). 2012. Source: CBS. 0. 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012. Enrolment Rate. Youth Unemployment Rate. Dummy and interaction term: In order to capture the effect of the Great Recession a dummy variable is added to the model. This binary variable equals one for the years 2008-2012 since the crisis was present in these years and zero for the other years included in the dataset. The coefficient shows how the economic crisis affects enrolment rates. This coefficient is expected to be positive based on the discussed literature and news articles above. There is also evidence that it may be negative due to budget constraints and lower returns to schooling. However, a positive relationship is mostly concluded. An interaction term between the dummy and youth unemployment is included in the model to test the effect of youth unemployment during the economic crisis. Household income: Preferably there would be data available on the household income of families with children who are about to leave secondary school. Unfortunately there is no data on these specific families; therefore, the best available measurement is used in this research. This is the average of gross household income among families with children, see figure 3. Data on households consisting of couples and single-parent families is used who have either only children younger than eighteen year old or at least one child older than eighteen. On the R. Davidsz / University of Amsterdam / 2014. 10.

(11) one hand it is arguable that including families with children only younger than eighteen biases the results, as many children younger than eighteen do not enrol in higher education. However, it is incorrect to exclude these since the number of children younger than eighteen who enter higher education is substantial, see figure 4. Including both types of households is, thus, reasoned to be the optimal measure. As discussed by McVigar and Rice the enrolment rate is expected to increase with average real household income because higher education is regarded as a normal good in consumption and secondly, where capital markets are imperfect, family income is an important source of finance for educational investments (2001, p. 51). Figure 3: Average gross household income of families with children. Source: CBS.. Average gross household income (x1000). 80. 70. 60. 50. 40. 30. 20. 10. 0. 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012. Figure 4: Ratio of seventeen year old or younger first-year student to total first-year students.. Seventeen year old or younger first-year students (%). Source: CBS. 19,5. 19. 18,5. 18. 17,5. 17. 16,5. 16. 15,5. 15. 14,5. 14. R. Davidsz / University of Amsterdam / 2014. 11.

(12) Financial Support: Dutch students are entitled to receive study finance from DUO (Dienst Uitvoering Onderwijs), a governmental organization. The purpose of study finance is to partly take away the financial burden of students in order to stimulate enrolment in post-compulsory education. Study finance is divided into four parts: a base grant, an additional grant, a public transportation card and a loan. The grant is a temporary loan, which turns into a gift if a student completes his or her studies within ten years. All students are entitled to a base grant and a public transportation card for four years, but only those whose parents cannot contribute to their child’s education are entitled to an additional grant and all students themselves can choose to take an additional loan. The amount of base grant depends on whether the student lives at home or not. The amount of additional loan depends on the parents’ income. Research shows that financial support has a positive impact on enrolment rates since the financial burden of studying is lowered (Dolton and Lin, 2011; Canton and de Jong, 2005; Dynarski, 1999). It is measured as total government spending on financial aid to bachelor students divided by the total number of enrolled students, giving the average financial support per student. As seen from figure 5, financial support has not been constant over time, as it fluctuated and overall increased since 2000. Unfortunately no distinction could be made between grants and loans. This is a shortcoming since grants decrease the net private contribution to educational investment, thereby raising the expected private return to schooling. Loans do not affect expected return to schooling, but may help to alleviate the credit market imperfection present in the higher education sector (Canton and de Jong, 2005, p.654). Tuition Fees: Prior to higher education no tuition fees are charged. The government finances compulsory education because it should be available to everyone. Higher education does charge tuition fees, which are centrally determined annually by the government. It is expected that tuition fees discourage youths to enrol for post-compulsory education, as it lowers the expected return to education (Canton and de Jong, 2005). Historic values of tuition fees over the years 2000-2012 are used, see figure 5.. R. Davidsz / University of Amsterdam / 2014. 12.

(13) 1900. 5000,00. 1800. 4500,00. 1700. 4000,00. 1600. 3500,00. 1500. 3000,00. 1400. 2500,00. 1300. 1200. 2000,00. 1100. 1500,00. 1000. 1000,00. Tuition Fee . Financial Support. Tuition Fee. Figure 5: Yearly average financial support and tuition fees. Source: CBS.. Avg Financial Support per Student. Education level labour force: This variable is represented by the ratio of labour force participants with a higher education diploma to the total labour force. Figure 6 shows the ratios of different education levels in the workforce to the total work force from 2000 till 2012. We see an increase in the ratio of those with a high education level and a decrease in the ratio of those with a low education level, confirming the idea of The Netherlands becoming a knowledge-based economy. Overall, those with a middle education level represent the largest part of the work force. But those with a high education level are rapidly increasing and may overtake the middle group in the near future. As the labour market has increasing numbers of skilled workers, young persons may feel pushed to enrol for a bachelor in order to be able to compete for jobs. This suggests that the expected sign of this variable is positive.. R. Davidsz / University of Amsterdam / 2014. 13.

(14) Figure 6: Ratio of labour market participants with low, medium or high education levels to total work force. Source: CBS.. Education Level Ratio (%). 50. 45. 40. 35. 30. 25. 20. 15. 10. 5. 0. 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012. Education Level: High. Education Level: Middle. Education Level: Low. Note: Those with a low education level at most obtained a primary school diploma, those with a middle education level at most obtained a secondary school diploma and those with a high education level at most obtained a higher education diploma.. Cohort size: A measure of cohort size is generated based on youth population data from the CBS. It includes the total number of young adults between 15-25 years old. Figure 7 plots how cohort size has been increasing over 2000-2012. This variable may be interpreted as a proxy for potential supply constraints, similar to Clark’s (2011) method. A higher number of youth in the population may lead to higher supply constraints, as higher education does not have an infinite number of spots available to students. This means that cohort size has a negative effect on enrolment rates. However, McVigar and Rice write that it is recognized that supply constraints may have been binding in certain areas, but the consensus is that for higher education as a whole it may be treated as perfectly elastic (2001, p. 50). A different interpretation is, therefore, used to describe the effect of cohort size. The relationship can be positive because a larger cohort may lead to more competition for jobs, encouraging students to enrol for higher education. On the other hand, it may be negative because a large cohort may signal a receptive labour market for young adults with or without skills; therefore students are less inclined to continue post-compulsory education as return to schooling decreases. Thus, the expected sign for cohort size is not clear. R. Davidsz / University of Amsterdam / 2014. 14.

(15) Figure 7: Cohort size. Source: CBS 2300000. Cohort size . 2250000. 2200000. 2150000. 2100000. 2050000. 2000000. 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012. Table 1 Descriptive statistics. Variables. Observations. Mean. Std. Dev.. Min. Max. ER. 13. 3.128. 0.261. 2.699. 3.472. YU. 13. 10.162. 2.061. 6.7. 13.2. INC. 13. 59526.92. 6355.021. 48675. 67750. FS. 13. 3592.59. 632.66. 2516.56. 4446.66. TF. 13. 1526.54. 141.84. 1304. 1771. EDUC. 13. 30.478. 2.965. 25.657. 34.267. COH. 13. 2163732. 55028.75. 2090850. 2256247. Note: Overall enrolment rate (ER), youth unemployment (YU), adult unemployment (AU), household income (INC), financial support (FS), tuition fees (TF), education level labour force (EDUC), cohort size (COH) and cohort population ratio (COHPOP).. R. Davidsz / University of Amsterdam / 2014. 15.

(16) 3.2 Enrolment in different fields of study The variables description above focuses on the effect of the crisis on aggregate enrolment rates. However, the impact of these variables on enrolment may be different among fields of study. Figure 8a and 8b plot the enrolment rates among studies from 2000-2012. The independent variable, enrolment rates per field of study, is calculated as the ratio of enrolment per field of study to the total enrolment in a year. The numerator includes first-year students in higher education enrolled for a bachelor programme in a certain field of study. The denominator is the total number of enrolments in higher education in each year. The figures show that Social sciences, business and law by far have the highest enrolment rates from 2000/’01 to 2012/’13. The next most popular field of study is Health and welfare and the field of study with the lowest enrolment rates is Agriculture. After the crisis in 2008 we see several changes in the enrolment rates per field of study. Education bachelors saw a decrease in enrolment rates from 10.4% to 9.3%. Hamermesh and Donald (2008), and Arcidiacono (2004) find that Education has the lowest average earnings, indicating lower returns to schooling. This may be a reason for decreased enrolment rates after the crisis. Humanities and arts, Social sciences, business and law, and Engineering, manufacturing and construction had minor decreases in enrolment rates. Agriculture and Services barely faced changes in enrolment rates after the recession. Science and Health and welfare are the two fields of study that saw increased enrolment rates since 2008. High average earnings for Science studies meaning high returns to schooling may lead to students particularly choosing this field of study in times of a bad economy (Hamermesh and Donald, 2008; Arcidiacono, 2004). It should be noted that the values of some variables, for example unemployment rates and the degree of competition, are different among studies. Ideally there would be time-series data about unemployment rates for different professions and cohort information of persons with the same educational background. However, this data is not available.. R. Davidsz / University of Amsterdam / 2014. 16.

(17) Figure 8a: Enrolment rates among three fields of study with highest enrolment rates. Source: CBS. 45. 40. 35. 30. 25. 20. 15. Education. Social sciences, business and law. Health and welfare. 10. 5. 0. Figure 8b: Enrolment rates among five fields of study with lowest enrolment rates. Source: CBS. 12. 10. Humanities and arts. 8. Science. 6. 4. 2. Engineering, manufacturing and construction. Agriculture. Services. 0. R. Davidsz / University of Amsterdam / 2014. 17.

(18) 3.3 Empirical model This research uses time-series data and this poses special challenges (Stock and Watson, 2012, p. 558). Using time-series data provides the opportunity to estimate the time path of the effect on enrolment rates of a recession and changes in other possible influences. In other words, it allows estimation of the dynamic causal effect on enrolment rates. Dynamic effects necessarily occur over time, therefore, the econometric model used to estimate dynamic causal effects needs to incorporate lags (p.630). Such a model is called a distributed lag model. In this research a simplified version is used including differences and excluding lags of the variables. Including differences of the variables is necessary to ensure the variables are stationary. However, including lags of the limited data at hand leads to few observations and low degrees of freedom. All variables included in the distributed lag model must be exogenous, which holds if the conditional mean of the error ut in the model of enrolment rates on past and current values of the independent variables does not depend on past and current values of the independent variables. For this research youth unemployment, household income, financial support, tuition fees, educational level labour force and cohort size are assumed to be exogenous based on a combination of economic theory and previous literature. The OLS estimators of the coefficients in such a regression are, therefore, consistent estimators of the causal effects. The independent variables are strictly exogenous if, in addition, the conditional mean of ut does not depend on future values of the independent variables. It would then be preferable to use OLS estimation of an ADL model or by GLS. However, this is a very strong assumption, which I cannot make. A problem that arises when using time-series data is that the error term ut can be serially correlated (Stock and Watson, 2012, p. 643). This results in consistent OLS coefficient estimators but the OLS standard errors are not, leading to false hypothesis tests and confidence intervals. The solution is to calculate the standard errors by using a heteroskedasticity- and autocorrelation-consistent (HAC) estimator of the variance, also called the Newey-West variance estimator, or by using robust standard errors for the regression. Therefore, regressions with Newey-West and robust standard errors are performed. These lead to similar results as is shown in the next section. Another challenge with time-series data is the likelihood of stochastic trends present in the data. As Canton and de Jong mention, there might be problems concerning stochastic R. Davidsz / University of Amsterdam / 2014. 18.

(19) trends present in for example financial support (2005, p. 655). Most of the time, using first differences eliminates random walk trends in a time series (Stock and Watson, 2012, p. 597). Dickey-Fuller tests are performed to determine if there are stochastic trends in the data. The null-hypothesis of the Dickey-Fuller test is that a variable has a unit-root. Continuously performing Dickey-Fuller tests until the null hypothesis is rejected gives the correct differences for each variable ensuring stationarity. This results in taking first differences for enrolment rates, financial support and education level, second differences for household income, tuition fees and cohort size, and a fourth difference for youth unemployment rates. Beyond second differences, however, it becomes hard to interpret the estimated coefficients. This means that the model is statistically correct but cannot be interpreted. Therefore, the main model of this research discusses up to second differences of the independent variables in order to be able to interpret the results. Furthermore, a potential hazard is multicollinearity, which means that two or more regressors are highly correlated (Stock and Watson, 2012, p. 244). If the regressors are imperfectly multicollinear, then the coefficients on at least one independent variable will be imprecisely estimated. The correlations between the independent variables show that the second difference of cohort size is highly correlated with the second difference of income (-0.5484) and the second difference of tuition fees (0.8050), see Table B in Appendix A. Moreover, the regressor tuition fees is highly correlated with financial support (-0.5616). Cohort size and tuition fees are, therefore, excluded from this empirical analysis to avoid multicollinearity. Given all the information above, the main empirical model for this research looks as follows D.ln ERt = β1 + β2 Crisis + β3 Crisis#D2.ln YUt + β4 [X’t] + ut, X’t = [D2.ln YU; D2.ln INCt; D.ln FSt; D.ln EDUCt], where D.ln ERt represents the first difference of the natural logarithm of the enrolment rate in year t. X’ is the vector of the first or second differences of the natural logarithms of the independent variables household income (INC), financial support (FS) and education level labour force (EDUC) in year t. In order to capture the effect of the Great Recession, a dummy variable is added. Moreover, an interaction term of the dummy with youth unemployment is R. Davidsz / University of Amsterdam / 2014. 19.

(20) included. Crisis is the dummy variable and Crisis#D2.ln YU is the interaction term. The dummy equals one for the years 2008-2012 as these are years during which the crisis was present. β2 presents the difference in enrolment rates when there is a crisis and when there is not, holding everything else constant, β3 shows the difference in the effect of youth unemployment rates during a crisis and in periods of no crisis and β4 represents the effects of the independent control variables. A log-log model is used because all variables are most naturally discussed in percentage terms. Coefficients can, therefore, be interpreted as elasticities. β2 is considered to be most important for this research as it describes the immediate effect of the crisis on enrolment. Moreover, increased youth unemployment is one of the effects of the crisis and, therefore, indirectly represents the effect of the crisis on enrolment. The interaction term can indicate whether the effect of youth unemployment is different in times of a bad economy. 4 Results Time-series regressions are run in Stata 12 using aggregated and disaggregated enrolment rates data. The results of the main empirical model are presented in Table 2. In Eq. (1) I perform a regression with Newey-West errors and a maximum of two lags1. In Eq. (2) I perform the same regression but with robust standard errors. This results in the same estimated coefficients, however, slightly larger standard errors. On the positive side, Eq. (2) reports the R-squared and this gives insight in the fraction of the time-series variation of the enrolment rate that is explained by the regressors. About 49% of the variation of the enrolment rate is explained by the model. This high R-squared may indicate multicollinearity, however, testing different model specifications shows that the coefficients do not change drastically (Table 3), thus, the models are likely not misspecified.. 1 A guideline for choosing the number of lags in practice is to use the formula m = 0.75T1/3 rounded to an integer (Stock and Watson, 2012), where m is the number of lags and T is the time-period that is considered in the model. Trying more lags results in similar results, estimating the same coefficients and slightly smaller standard errors. R. Davidsz / University of Amsterdam / 2014. 20.

(21) Table 2 Higher education enrolment rates affected by Great Recession (Dependent variable: D.lnER) Using different standard errors Eq. (1). Eq. (2). -0.008 (0.018). -0.008 (0.025). 0.086 (0.080). 0.086 (0.107). 0.011 (0.108). 0.011 (0.131). D2. lnINC. 0.128 (0.531). 0.128 (0.627). D. lnFS. 0.173 (0.150). 0.173 (0.124). D. lnEDUC. 0.071 (0.391). 0.071 (0.422). N. 11. 11. Crisis D2. lnYU Crisis#D2.lnYU. R2. 0.4949. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. In Eq. (1) heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. In Eq. (2) robust standard errors are used. Both models include a constant term.. Table 3 Higher education enrolment rates affected by Great Recession (Dependent variable: D.ln ER) Using different model specifications. Crisis D4. lnYU Crisis#D4.lnYU. Eq. (3). Eq. (4). -0.009 (0.015). -0.010 (0.014). 0.086 (0.074). 0.091 (0.056). 0.005 (0.090). 0.008 (0.094). D2. lnINC. 0.123 (0.489). D. lnFS. 0.167 (0.117). 0.176 (0.134). D. lnEDUC. 0.064 (0.381). N. 11. 11. R2*. 0.4890. 0.4928. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. Both models include a constant term. * R2 is calculated using the same model specifications but with robust standard errors instead of a Newey-West regression.. R. Davidsz / University of Amsterdam / 2014. 21.

(22) Figure 9 shows the actual and predicted values of the first difference of the natural logarithm of enrolment rates. Because the main model includes second differences two years of observations are lost, 2000 and 2001. Overall, the plot shows that the predicted values are close to the actual values, which indicates that the model is suitable for estimating enrolment rates in the Netherlands. Figure 9: Actual and predicted values of D.lnER 0,08. 0,06. 0,04. 0,02. 0. -0,02. 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012. -0,04. -0,06. -0,08. -0,1. Actual Values. Predicted Values. The following analysis focuses on the regression with Newey-West errors since the standard errors are smaller in this specification. Unfortunately none of the estimated coefficients2 of the model in Table 2 are statistically different from zero. This is largely attributed to the small sample size of thirteen observations. However, it is insightful to discuss the results of the regression and consider possible effects of the variables on enrolment rates. The coefficient of the dummy variable is counter intuitively negative meaning that enrolment rates are lower in times of a bad economy holding everything else constant. The coefficient of youth unemployment is positive which means that youth unemployment positively affects enrolment rates when there is no crisis. The positive coefficient of the interaction term tells us that youth unemployment increases enrolment more in times of a crisis confirming the arguments of lower opportunity costs and escaping future unemployment. The coefficients of 2 Estimated coefficient of a variable refers to the coefficient of the specified difference of the natural logarithm of that variable. R. Davidsz / University of Amsterdam / 2014 . 22.

(23) household income, financial support and education level have the expected sign in the regression model. A negative impact of the crisis on enrolment rates is counterintuitive given the results from previous research and news articles. However, the reason for this result can be derived from Figure 2. It shows a steady increase in enrolment rates over the years 2002 till 2009. From 2010 till 2012, however, enrolment rates are relatively flat and slightly lower than in 2009. This correctly results in a negative OLS estimate of the dummy variable. Again the limited observations may have led to this unexpected sign. It would have been preferable if there were data available from after the crisis in order to see what happens to enrolment rates afterwards. Supported by other research, however, a positive relationship between the Great Recession and enrolment rates is still most likely. In Appendix B results of several other model specifications are shown. Table A shows the results from the correct model, including the fourth difference of youth unemployment. Even though this is statistically speaking the correct model, including a fourth difference makes the model hard to interpret. Moreover, it results in only nine observations, which makes the data set even more limited. In Table B, Table C and Table D first, second and zero difference models are presented respectively. These models are likely incorrect due to multicollinearity since the R2 is high in each model. Table E shows a distributed lag model, including lags of youth unemployment, a dummy variable and an interaction term. It is not possible to analyse a distributed lag model of multiple independent variables due to the limited data set meaning a low number of observations. Including lags is sensible since past values can be important in a time series model, however the main model is preferred because including control variables gives more reliable estimates than a model with only one variable and its lags. Future values might also affect a students’ choice to enrol in higher education. A lead model is, therefore, tested which includes future values of the variables, see Table F. It is however unrealistic to assume that students can perfectly predict future values of the independent variables. Expected values would have been optimal to use, but this data is unfortunately not available. Ultimately, the main model of this research gives the most reliable and best interpretable results. Lastly, Table G shows the effects on upper vocational training and academic training separately to see if there are differences. The difference in the effect of the crisis on the two programmes is very small. De Volkskrant (Bouma and Hosselet, 2014) writes that enrolment in academic training increased with 5% between 2013 and 2014, whereas enrolment in upper vocational training stayed the same. In order to conclude about R. Davidsz / University of Amsterdam / 2014 . 23.

(24) the different effects, an adjusted model is preferable since there are other variables that need to be added, such as level of secondary education. However, this is not in the scope of this research and therefore it is left to future research. So far the discussed results were based on the effect of the variables on overall enrolment for all studies. However, the effects of the variables are different among fields of study. The model specification for the following regressions is the same as the main model apart from the dependent variable, which is now described as the enrolment rates specific to eight different fields of study. In Table 4 the estimated coefficients are presented with respect to the dummy variable, the youth unemployment rate and the interaction term. Unfortunately, the R2’s of most of the regressions are high indicating multicollinearity. Nonetheless some coefficients confirm the findings from previous empirical research. It appears that the crisis has the largest negative impact on enrolment for Education and Services studies, whereas it the largest positive impact on enrolment for Science and Agriculture studies. In other words, fewer students choose Education and Services studies and more students opt for Science and Agriculture studies in times of a bad economy. This is in line with findings that Education and Science studies have the lowest and highest average earnings translating in low and high returns to schooling respectively (Arcidiacono, 2004; Hamermesh and Donald, 2008). On the other hand, there are coefficients that are not in line with the discussed theories. For example, Social sciences, business and law studies have high average earning, but the estimated effect of the crisis is negative. Moreover, De Volkskrant (Bouma and Hosselet, 2014) writes that enrolment for Humanities and arts studies decreased due to the crisis, but the estimated coefficient is positive. All in all, the results in Table 4 have to be looked at critically, but in combination with previous research and public discussion the conclusion is that studies with low and high returns to schooling are respectively negatively and positively affected by the crisis.. R. Davidsz / University of Amsterdam / 2014 . 24.

(25) Table 4 Higher education enrolment rates per field of study affected by Great Recession (Dependent variable: D.ln ERfield of study) Field of study. Crisis. D2.ln YU. Crisis#D2.ln YU. R2. Education. -0.038 (0.026). 0.009 (0.132). -0.113 (0.159). 0.5558. Humanities and. 0.018 (0.015). -0.326 (0.066)*. 0.376 (0.082)*. 0.8682. -0.010 (0.014). 0.022 (0.047). -0.034 (0.068). 0.5366. Science. 0.122 (0.019)*. -0.385 (0.072)*. 0.630 (0.082)*. 0.9529. Engineering,. 0.000 (0.015). 0.018 (0.074). 0.020 (0.120). 0.1479. 0.139 (0.027)*. -0.010 (0.109). 0.199 (0.083). 0.8956. Health and welfare -0.009 (0.029). 0.073 (0.108). -0.163 (0.146). 0.4212. Services. 0.357 (0.108)**. -0.283 (0.073)**. 0.8182. arts Social sciences, business and law. manufacturing and construction Agriculture. -0.038 (0.025). Note: Sample period is 2000-2012. Number of observations is 11 for all models. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. All models include a constant term. * means significant at the 1% level and ** means significant at 5% level.. R. Davidsz / University of Amsterdam / 2014 . 25.

(26) 5 Conclusion This paper empirically analyses the impact of the Great Recession on enrolment rates in higher education in The Netherlands. It builds on previous literature from the UK, US and other countries based on panel and time-series data, and the public discussion in The Netherlands on enrolment rates. A disappointing aspect of my analysis is the small sample size leading to insignificant coefficients. According to my results the crisis is negatively related to enrolment rates, meaning that less young adults choose to continue their educational path during recession times. However, previous findings of amongst others Clark (2011) and McVigar and Rice (2001) show significant, positive effects and the general public conception is that enrolment rates increased due to the economic crisis. The positive effect is explained by students having lower opportunity costs during a recession since employment opportunities are diminished and students escaping future unemployment since skilled workers are less likely to end up being unemployed. Running regressions for enrolment rates among fields of study also gives interesting results. It shows a negative impact of the crisis on enrolment in Education studies and a positive impact on enrolment in Science studies. These studies have the lowest and highest average earnings, which indicates low and high returns to schooling respectively. The estimated coefficients indicate that the recession affects enrolment rates among studies differently due to returns to schooling, leading to less students opting for the low earnings study and more students choosing the high earnings study. It should be noted that variables such as youth unemployment might differ among studies, however I did not have this data. Obtaining such information could substantially improve the results of this research. Another way to improve the results is by obtaining individual specific information. It is therefore recommended to perform a national survey among students in different fields of study and young adults who choose not to enrol in higher education. Information on for example personal background, family background, employment expectations, earnings expectations and interests should be gathered. This can give insights into the youth’s individual decision-making process concerning enrolment in higher education. Answering my research question is an exciting research area that has implications for both education and labour market policy. A positive relationship implies that there is more R. Davidsz / University of Amsterdam / 2014 . 26.

(27) demand for higher education during a crisis meaning that, for example, schools need to be able to facilitate the increased number of students. Negative relationships for certain studies, for example, imply that less young adults will graduate in these fields which may lead to a shortage of labour market participants with these skills. Ultimately, the limited data available hindered this empirical analysis from conformingly concluding a positive relationship between the crisis and enrolment rates. Therefore, future research should expand the data set either by performing a national survey or including more years after the crisis once this is available. Possibly this empirical model will then conclude an increase in enrolment rates in higher education in The Netherlands due to the economic crisis.. R. Davidsz / University of Amsterdam / 2014 . 27.

(28) Appendix A: Correlations Table A: Correlations between unemployment rates among different age groups. Age Group. 15-25. 25-35. 35-45. 45-55. 15-25. 1.0000. 25-35. 0.5987. 1.0000. 35-45. 0.4527. 0.9524. 1.0000. 45-55. 0.6023. 0.9335. 0.9649. 1.0000. 55-65. 0.8099. 0.8143. 0.8128. 0.9091. 55-65. 1.0000. Table B: Correlations between dependent and independent variables D.ln. D2.ln. ER. YU. Crisis. D2.ln. D.ln FS. INC. D2.ln. D2.ln. D.ln. TF. COH. EDUC. D.ln ER. 1.000. D2.ln YU. 0.3254. 1.000. Crisis. -0.1269. 0.2565. 1.000. D2.ln INC -0.1567. -0.3427. -0.1974. 1.000. D.ln FS. 0.3390. -0.5153. -0.1875. -0.0151. 1.000. D2.ln TF. -0.4197. 0.4708. 0.2152. -0.4457. -0.5616. 1.000. D2.ln. -0.2476. 0.2222. 0.0331. -0.5484. -0.1110. 0.8050. 1.000. 0.3011. 0.1590. -0.3299. -0.0712. 0.1339. -0.2563. 0.0685. COH D.ln. 1.000. EDUC R. Davidsz / University of Amsterdam / 2014 . 28.

(29) Appendix B: Different empirical models Table A: Statistically correct model (Dependent variable: D.lnER) Eq. (A) Crisis D4. lnYU Crisis#D4.lnYU. -0.015 (0.021) 0.018 (0.053) 0.008 (0.045). D2. lnINC. -0.483 (0.710). D. lnFS. 0.007 (0.121). D. lnEDUC. 0.445 (0.452). N. 9. R2*. 0.4283. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression. Both models include a constant term.. Table B: First differences model (Dependent variable: D.lnER) Eq. (B) Crisis D. lnYU Crisis#D.lnYU. 0.007 (0.009) -0.062 (0.065) 0.115 (0.131). D. lnINC. 0.100 (0.561). D. lnFS. 0.125 (0.181). D. lnEDUC. 0.777 (0.400). N. 12. R2*. 0.5747. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression. Both models include a constant term.. R. Davidsz / University of Amsterdam / 2014 . 29.

(30) Table C: Second difference model (Dependent variable: D2.lnER) Eq. (C) Crisis D2. lnYU Crisis#D2.lnYU. -0.057 (0.022)*** 0.403 (0.136)** -0.343 (0.082)**. D2. lnINC. -0.994 (0.579). D2. lnFS. 0.075 (0.070). D2. lnEDUC. -0.709 (0.576). N. 11. R2*. 0.7409. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression. Both models include a constant term. ** and *** mean significant at the 5% and 10% level respectively.. Table D: No differences model (Dependent variable: lnER) Eq. (D) Crisis. -0.150 (0.130). lnYU. -0.086 (0.043)***. Crisis#D4.lnYU. 0.082 (0.060). lnINC. -0.271 (0.127). lnFS. 0.075 (0.171). lnEDUC. 0.882 (0.353). N. 13. R2*. 0.9695. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression. *** means significant at the 10% level.. R. Davidsz / University of Amsterdam / 2014 . 30.

(31) Table E: Distributed lag model (Dependent variable: D.lnER) Eq. (E) D2.lnYU. 0.058 (0.043). L.D2.lnYU. 0.035 (0.039). L2.D2.lnYU. 0.021 (0.085). Crisis. -0.022 (0.021). Crisis#D2.lnYU. -0.044 (0.073). R2*. 0.3838. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression. Model includes a constant term.. Table F: Lead model (Dependent variable: D.lnER) Eq. (F) Crisis. 0.033 (0.031). F.D2.lnYU. -0.191 (0.123). Crisis#F.D2.lnYU. 0.246 (0.147). F.D2.lnINC. 0.953 (0.758). F.D.lnFS. 0.045 (0.124). F.D.lnEDUC. 1.075 (0.468). R2*. 0.5635. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression. Model includes a constant term.. R. Davidsz / University of Amsterdam / 2014 . 31.

(32) Table G: Enrolment rates in upper vocational and academic training affected by Great Recession (Dependent variable: D.lnER) Upper Vocational Training. Academic Training. -0.009 (0.011). -0.007 (0.025). 0.094 (0.042). -0.021 (0.108). -0.033 (0.059). 0.144 (0.147). D2. lnINC. -0.059 (0.325). 0.113 (0.630). D. lnFS. 0.095 (0.091). 0.189 (0.144). D. lnEDUC. 0.045 (0.140). 0.135 (0.555). N. 11. 11. R2*. 0.6170. 0.4305. Crisis D2. lnYU Crisis#D2.lnYU. Note: Sample period is 2000-2012. Standard errors are in parenthesis. The regression method is OLS. Heteroskedasticity- and autocorrelation-consistent, also known as Newey-West standard errors are used and the maximum lag is 2. Both models include a constant term. * R2 is calculated using the same model specification but with robust standard errors instead of a Newey-West regression.. R. Davidsz / University of Amsterdam / 2014 . 32.

(33) References Albert Verdu, C. (2000). Higher education demand in Spain: the influence of labour market signals and family background. Higher Education, 40 (2), 147–162. Altonji, J. G., Blom, E. & Meghir, C. (2012). Heterogeneity in human capital investments: high school curriculum, college major, and careers. National Bureau of Economic Research Working Paper, No. 17985. Arcidiacono, P. (2004). Ability sorting and the returns to college major. Journal of Econometrics, 121, 343-375. Bouma, K. and Hosselet, L. (2014). Studenten kiezen steeds vaker studie met goede kans op baan. De Volkskrant. Bell, D. N. F. & Blanchflower, D. G. (2011). Young people and the Great Recession. Oxford Review of Economic Policy, 27(2), 241-267. Betts, J. R. & McFarland, L. L. (1995). Safe port in a storm: The impact of labour market conditions on community college enrollements. The Journal of Human Resources, 30(4), 741-765. Canton, E. & Jong, F. de. (2005). The demand for higher education in The Netherlands, 19501999. Economics of Education Review, 24, 651-663. CBS Statline. Retrieved from http://statline.cbs.nl/StatWeb/?LA=nl Clark, D. (2011). Do recessions keep students in school? The impact of youth unemployment on enrolment in post-compulsory education in England. Economica, 78(311), 523– 545. Clark, A. E., Georgellis, Y. & Sanfey, P. (2001). Scarring: The psychological impact of past unemployment. Economica, 68, 221-241. Clark, A. E. & Oswald, A. J. (1994). Unhappiness and unemployment. The Economic Journal, 104(424), 648-659. Corman, H. (1983). Postsecondary education enrolment responses by recent high school graduates and older adults. The Journal of Human Resources, 18(2), 247-267. Dolton, P., & Lin, L. (2011). From grants to loans and fees: the demand for post-compulsory education in England and Wales from 1955 to 2008 (No. 0127). Centre for the Economics of Education, LSE. Dynarski, S. M. (1999). Does aid matter? Measuring the effect of student aid on college attendance and completion. National Bureau of Economic Research, No. 7422. Hamermesh, D. S., & Donald, S. G. (2008). The effect of college curriculum on earnings: An affinity identifier for non-ignorable non-response bias. Journal of Econometrics, 144, 479-491. van der Hulst, A. (2014). Studenten kiezen voor een studie met uitzicht op een baan. NRC. Jacobs, B. (2002). An investigation of education finance reform; graduate taxes and income contingent loans in the Netherlands (No. 9). CPB Netherlands Bureau for Economic Policy Analysis. Knabe, A., & Rätzel, S. (2011). Scarring or Scaring? The psychological impact of past unemployment and future unemployment risk. Economica, 78(310), 283-293. McIntosh, S. (2001). The demand for post-compulsory education in four European countries. Education Economics, 9(1), 69-90. McVicar, D., & Rice, P. (2001). Participation in further education in England and Wales: an analysis of post-war trends. Oxford Economic Papers, 53(1), 47-66. R. Davidsz / University of Amsterdam / 2014 . 33.

(34) Micklewright, J., Pearson, M., & Smith, S. (1990). Unemployment and early school leaving. The Economic Journal, 100(400), 163-169. Middelburg, B. (2009). Crisis jaagt studenten naar UvA. Het Parool. Petrongolo, B., & San Segundo, M. J. (2002). Staying-on at school at 16: the impact of labor market conditions in Spain. Economics of Education Review, 21(4), 353-365. Rice, P. (1999). The impact of local labour markets on investment in further education: evidence from the England and Wales youth cohort studies. Journal of Population Economics, 12(2), 287-312. Shea, J. (2000). Does parents’ money matter?. Journal of public Economics, 77(2), 155-184. Spijkerman, C. (2013). Door crisis flinke toename aanmeldingen voor opleiding in zorg. NRC. Tumino, A., & Taylor, M. (2013). The impact of local labour market conditions on school leaving decisions. In Population Association of America Annual Meeting, New Orleans, LA.. R. Davidsz / University of Amsterdam / 2014 . 34.

(35)

No results found