The Supply Effect of Educated Labour on the Returns to Schooling
in the Netherlands: an Analysis
Bachelor Thesis by Vishand Lachmansingh
Name: Vishand A. Lachmansingh Student Number: 10244530
Supervisor: Sabina Albrecht Field: Labour Economics
1 Statement of Originality
This document is written by Student Vishand Amrish
Lachmansingh who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
2
Abstract
With a steady increase in educated labour, a decrease in returns to schooling is to be expected. However, dominant literature rarely reports a decrease in returns and instead argues that the demand for educated labour rises faster than the supply. To check if similar findings could be found for the Netherlands, this research has estimated the causal returns to schooling with the use of instrumental variable regression. This regression technique was used in order to tackle the endogeneity of the schooling variable which would lead to biased least squares estimates. The research provides evidence of a decreasing returns to schooling for the period 2000-2012. Furthermore, it attempts to determine the size of the relation between the fractions of highly educated labour and the returns to schooling. Lastly, information provided by other authors on the Dutch labour demand is then used to discuss the plausibility of the results.
JEL Classification: I26; J2; J24
1. Introduction
Education can be regarded as an investment in an individual’s human capital. Education is characterised by initial costs and opportunity costs, and eventually yields returns in the form of a higher wage. The most conventional tool that takes education and wages into account is the Mincer’s equation. It models the effect of human capital, such as education and experience on an individual’s earnings. Estimating this model simplifies the private benefits of education into the form of the returns to schooling. The returns to schooling is defined as the additional amount an individual would earn with one extra year of schooling. This return on investment enables an individual to improve their economical position in society.
During the period 1996-2014, the Dutch government was lenient with its funding of education and student loans. The value of a student loan was determined on the basis of parental income levels and living situation of the student (Ritzen, 1996). These student loans are to be granted if students obtain a degree within 10 years. These regulations imply some sense of leniency whereas most academic and vocational school programs take up to three or four years to complete. The Dutch government has also invested heavily in education by subsidising educational institutions throughout the country. Jongbloed (2005) estimated that the government has funded 81% of costs for educational institutions. Whether the lenient stance of the government on the financing of education has contributed to a well-educated society is still uncertain. Based on population data provided by the Central Bureau of Statistics the fraction of highly educated labour has increased by 26% in the time period of 2000-2012. The Central Bureau of Statistics (2012) found that a large portion of the increase in educated labour is the result of greater
3 participation rates of females and immigrants in higher education.
As stated before, education is an investment of an individual into his or her human capital. According to the basic principles of economics, an increase in the supply of highly educated labour would influence the price paid for educated labour. The “Returns to Schooling” would decrease when confronted by supply increases of educated labour. This introduces the research question of this thesis: can a supply effect be found of high educated labour on the returns to schooling? This question will be answered with the use of the IV regression technique to estimate returns to schooling. With the returns it will be possible to check for a correlation between the fraction of highly educated labour and the returns which will indicate the existence of a possible supply effect. The analysis of the results will then be discussed with qualitative information on the demand factors for educated labour in the Netherlands.
The thesis is divided into seven sections. The second section will discuss relevant information on the human capital model by Mincer. Furthermore, it will review studies from other researchers in the field of labour economics and discusses the endogeneity of the schooling variable and other potential sources of bias. Section three will contain a description of the data used for this research. Section four will explain the methodology required to estimate the causal returns to schooling which can then be used to find a correlation that could indicate the existence of a supply effect. Section five will report the results of the estimations after which section six will contain a discussion on these results. Section seven will include some concluding remarks. Furthermore, an appendix will be added with additional information that is referred to throughout the thesis.
2. Literature review
In what follows, several theories and proposed methodologies will be assessed as part of a literature review. First, the conventional Mincer’s earnings function will be introduced in terms of its overall concept. Second, the rationale behind the chosen methodology is discussed with relation to the endogeneity of schooling and other sources of bias. The third subsection will discuss the results and methodologies of other authors as is discussed in Card’s paper “Earnings, Schooling and Ability Revisited” (1995). Finally, the methodologies of Kalwij (2000) and Levin & Plug (1999) will be reviewed and will provide the basis of the methodology used for this research.
4 2.1 Mincer’s Earning Function
The returns to schooling have been estimated by various academic authors throughout the years. The most frequently applied model to estimate the effects of human capital on earnings is provided by Jacob Mincer (1974A):
ln 𝑤𝑎𝑔𝑒 =∝ + 𝑟 𝑆 + 𝛽 𝑋 + 𝛽 𝑋
In this model the logarithm of earnings is used as a dependant variable. The independent variables are S, the years of schooling, and X, work experience. ∝ Is a constant factor also viewed as the intercept of the model. Earnings is defined as a logarithm instead of absolute terms for two reasons: firstly, the use of logarithms is convenient for analysing results in contrast to earnings in absolute terms. Secondly, education is viewed as an investment made by individuals, they will keep investing in education until the rate of return equals their discount rate (Lemieux, 2006). Based on this notion, it is rational to present the rates of return to education rather than the absolute returns.
The Mincer earnings function also takes the age earnings profile of individuals into account. This profile implies that the age of individuals is correlated with their earnings through the acquisition of experience which benefits their productivity. Work experience is viewed as a factor that influences human capital after the schooling period of individuals (Borjas, 2013).
The proxy for work experience takes three alternative forms in the literature; “Potential work experience”, “Age” and “Actual Work Experience”. “Potential Work Experience” is defined as 𝑋 = 𝐴𝑔𝑒 − 𝑆 − 𝑏, where 𝑆 is the years of schooling and 𝑏 is the age when an individual starts school (Mincer, 1974A). The use of “Potential work experience” is preferred over “Age” since it yields better estimations for post school investment in human capital (Miller, 1993). “Age” could lead to less precise estimations if individuals within the sample leave school earlier and gain relatively more work experience. The use of “Potential Work Experience” as a proxy for the experience variable, assumes that work experience is identical for individuals with the same amount of schooling and age. This assumption is not valid when labour participation is discontinuous for individuals after they finished their education. With discontinuous labour participation the estimator for work experience will be biased by other factors such as the negative aging effect on earnings. With discontinuous labour participation and the use of “Potential Work Experience” the estimations for the return on experience will be imprecise when compared with the use of “Age” or “Actual Work Experience” (Miller, 1993). Mincer (1974B, p.192) therefore argued that data on real work experience is preferable.
2.2 Endogeneity of the Schooling variable and other forms of bias
To estimate the Mincers earnings function it is possible to use the conventional Ordinary Least Squares regression technique (OLS). This method has been widely used throughout economic literature to
5 estimate models. However, the use of OLS to estimate the Mincers earnings function could lead to biased estimates on the returns to schooling. There are several reasons found in the literature for this estimation problem. In 1975 Becker published his theorem on the human capital model that challenged one of the main assumptions of the Mincer Earnings Function. Mincer’s model assumed that the payoff of human capital is the same for every individual and that all individuals have access to the same human capital investment opportunities (Blackburn & Neumark, 1995). Becker changed this assumption to something more realistic; people with a higher innate ability will enjoy higher returns to human capital (Blackburn & Neumark, 1995). In other words, ability is a personal trait that would affect both the years of schooling and a person’s earnings (Kalwij, 2000). If ability is not accounted for in the model, the parameter for earnings will be biased. Blackburn & Neumark (1996) used IQ scores as a proxy for ability and found that including IQ scores in their model leads to an IV estimate of the returns to schooling that is close to the OLS estimate of the returns to schooling. With omitted variables, a potential estimator could pick up other effects it’s correlated with, which could lead to a less efficient estimator.
Besides the lack of a proxy for ability there are also other factors that lead to biased estimates when using the OLS technique. According to the human capital theory an individual will choose their years of schooling with certain expectations of their future earnings. This leads to a simultaneous decision process for both schooling and earnings which influences the standard errors of conventional human capital model (Plug, 2001).
This endogeneity problem can be solved with the use of instrumental variable regression. This estimation technique requires an “instrument” to estimate a parameter that is endogenous with the dependant variable. There are two conditions that apply when considering preferred instruments: the first condition is the relevance condition where the instrument 𝑍 should be related to endogenous variable, in this case 𝐶𝑜𝑣(𝑍, 𝑆) ≠ 0 (Plug, 2015). The second condition is the exogeneity condition where the instrument Z should not be related to other factors that affect earnings, in this case𝐶𝑜𝑣(𝑍, 𝜀) = 0. The relevance condition can be tested, the exogeneity condition can only be tested if there are more instruments than endogenous variables (over identification). If the endogenous variable is exactly identified, the instrument should be argued to be exogenous with the use of literature. The IV technique will be further discussed in the Methodology section.
Another source of potential bias in the literature comes with the definition of the earnings logarithm. Earnings defined as monthly or annual earnings will lead to biased estimations due to the fact that higher educated people tend to work more hours (Card, 1999; Gunderson & Oreapoulos, 2010). An estimate of the returns to schooling could therefore reflect longer working hours instead of a causal return to schooling. This problem can be solved by converting monthly or annual earnings to hourly earnings. Card (1999) found that the return to schooling parameters estimated in various American studies based on annual earnings were only for two thirds accountable to the causal effect of education on earnings. The other 33% of the parameter could be explained by the longer hours effect which these studies failed to take into account.
6 The last potential bias is measurement error. Many sets of data rely on self-reported values for the schooling variable. Card (1999) found that the reliability of such values are around 90%. This is because people can misreport their years of schooling and/or their earnings. The OSA Panel survey used in this research reports the highest obtained degree for individuals, this is converted into years of schooling. The problem with this approach is that the OSA panel survey does not take into account the possible extra years that individuals took to complete their schooling. The effects that could be measured would therefore be biased due to miscalculation in the years of schooling of the individuals.
2.3 Solutions for the endogeneity of Schooling by other Authors
The third section of this literature review will examine instruments that have been used in the past to solve the endogeneity problem and the validity of their exogeneity. Card (1995) reviewed the methodology of Angrist and Krueger (1991), Butcher and Case (1994) and Card (1993) in his paper “Earnings, Schooling, Ability Revisited”; Angrist and Krueger (1991) used time of birth as an instrument to estimate the effects on schooling. They linked the time of birth to the compulsory schooling laws in the United States. The idea behind the use of this instrument is that individuals born later in the year can drop out from high school at an earlier time than their classmates. This creates a variance in completed schooling for high school dropouts that were born in different quarters. This difference in completed schooling can then be used to estimate the effects on earnings. Their IV estimations are close to the OLS estimates which suggests little to no bias in conventional methods. Their IV estimate includes “Age” and “Age squared” as well as some personal characteristics such as “Sex” and “Race” which provides them with an estimate of 0.1007.
Butcher and Case (1994) used sex composition of siblings as an instrument to estimate schooling. They found that the sex composition of siblings only had effect on females and not on males. To be more specific, they found that females with sisters completed on average less years of schooling than those with brothers. This result is argued to be plausible based on the notion of classroom activity; females with brothers will be better able to compete for class resources and teacher’s attention than females that grow up with sisters. Estimating the returns to schooling using the conventional OLS estimates gave a parameter of 0.091. This estimation is close to the estimate of Angrist and Krueger. However, their IV estimate was 0.184, which is much higher than the OLS estimate and the IV estimate of Angrist and Krueger. They argue that the returns to schooling for females could have been seriously underestimated.
Card (1993) used geographic variation as an instrument to estimate the returns to schooling. He found that males who live closer to a college earn significantly more and have a higher education compared to other males. He argues that this effect is mostly noticeable for males with low educated parents, who would stop their investment in schooling based on distance. He estimated the effect of a
7 nearby college on earnings and found that this effect was not significant. This was to strengthen the argument that the proximity of a college only affects earnings through educational attainment and could therefore be viewed as a valid instrument for IV regression. Card’s (1993) OLS estimate of 0.073 is also close to other estimates found by Butcher, Case, Angrist and Krueger. His IV estimate using the proximity of a nearby college as an instrument is 0.117, which is significantly higher than the OLS estimate but smaller in comparison to the IV estimate of Butcher and Case.
2.4 Dutch studies on the Returns to Schooling
Kalwij (2000) tried to estimate the returns to schooling in the Netherlands using a panel data model. Instead of using the IV approach he made use of the fact that older people in his sample obtained less schooling than younger people. His method enabled him to control for birth cohort effects and found that controlling for these effects resulted in a higher return on “Work Experience”. He also found that failing to account for the endogeneity of schooling would result in a lower estimation of returns to schooling which is also suggested by other studies.
Levin and Plug (1999) estimated the returns to schooling with several instruments provided by others to see if instruments used in other countries would yield the same results in the Netherlands. They analysed the following instruments: parental education and job level of parents, the quarter of birth, sibling rank and social status. Individuals coming from a family background with highly educated parents and high job level attain more schooling on average. The quarter of birth can be linked to compulsory schooling laws such as the method of Angrist and Krueger. Levin and Plug made use of sibling rank, under the assumption that first born children have more resources at their disposal, whereas later born children will be enjoy fewer resources due to the sharing of resources between siblings. The last instrument, social status, is measured by teacher’s evaluation in the survey. The assumption is made that people from a higher social status would attain more schooling which is a firm assumption. From the analysis they found that family background provided the most reliable results on returns to schooling which was in line with findings of other authors. A simple OLS estimate for the OSA labour survey of 1994 yields a returns to schooling of 0.036, while the IV estimation yields 0.05.
The dominant literature on the returns to schooling finds that OLS estimates of schooling on earnings functions are biased downward (Levin & Plug, 1999). This would make IV estimation a necessary method in estimating the returns to schooling when a proxy for ability is lacking. Furthermore the conventional mincer’s function is always estimated in combination with other exogenous personal characteristics such as “Sex”, “Race” or “Marital Status” to improve the parameter of their estimation. By including the exogenous variables the explanatory power of the model is increased resulting in a higher R-squared.
8
3. Data
The dataset that will be used for the estimations of returns to schooling is the OSA Panel survey. It was managed by the Organisation for Strategic Labour Research until the institute’s bankruptcy in 2009. The panel has been in existence since 1985 and is one of the longest running panel data surveys in the world. Due to the significance of the dataset the Social Cultural Planning Agency took over its management in 2009 to ensure its survival. The target population of the panel is the Dutch labour force aged 16 till 66. The dataset includes 4837 observations in 2012 with data on the following themes; background characteristics, current employment, characteristics of the unemployed, a special questionnaire for students, position in the labour market, labour history, education, trainee programs, income and social insurance laws, work and opinions regarding various statements (OSA, 2015). OSA aimed to make the dataset as representative of the Dutch population as possible. This makes it possible to draw conclusions from the sample that is related to the population.
4. Methodology
To analyse a possible supply effect of education on the returns to schooling it is necessary to first estimate the returns to schooling. IV regression has been acknowledged to lead to less biased estimates of the return to schooling than OLS regressions. To clarify this let’s suppose that the returns to schooling can be measured with a simple model:
𝐿𝑛 𝑤𝑎𝑔𝑒 = 𝛼 + 𝛽𝑆 + 𝜀
The OLS estimate would then be:
𝛽 = ( , )
( ) = 𝛽 + ( , )
( )
Where 𝛽 is viewed as the causal effect of schooling on earnings and 𝜀 represents other factors that influence earnings besides schooling. Performing an OLS regression will estimate 𝛽 . This parameter can only be interpreted as causal if the 𝐶𝑜𝑣(𝜀, 𝑆) is equal to zero (Kling, 2000). In case of the Mincer’s earnings function it is firmly believed by the academic world that 𝐶𝑜𝑣(𝜀, 𝑆) is not equal to zero. The instrument used for the schooling variable will be the education levels of an individual’s parents. Parental education levels are considered to be influential on the development of a child. When a child is raised by highly educated parents, they are usually subjected to an environment that promotes his educational career. The OSA Panel survey reports the highest obtained education for the mother and father of the individual. This will be divided into four different categories of education levels; low, intermediate, higher and unknown. The sample reports a small amount of individuals having a foreign
9 education. These observations have been deleted from the sample since they lead to highly inflated estimations and provide no significance. This instrument has been analysed by Levin and Plug (1999) in their paper “Instrumenting education and the returns to schooling in the Netherlands” alongside with another instrument ‘parental job level’ they found that these instruments led to the most reliable results.
The years of schooling of an individual was derived on basis of the highest obtained degree. Table I reports the conversion used for the degrees of individuals. For the instruments the following conversion applies: <12 years of schooling is considered as a low level of education, 14 years of schooling is seen as intermediate level of education and 17 years or greater is viewed as a higher level of education. Parental education was characterised by 30 different types of degrees. These options have been evaluated to determine if the parents obtained low, intermediate or higher education. Since the data-set is panel data it was possible to import missing values for parental education from other years. If parental education was still missing after this procedure it has been classified as “Parental Education level Unknown”. Table VII in the appendix reports the fractions of Parental Education levels for each of the estimated years.
The following OLS model will be estimated:
(1) ln 𝑤𝑎𝑔𝑒 = 𝜔 + 𝑟𝑆 + 𝛽 𝑋 + 𝛽 𝑋 + 𝛽 𝐹𝑒𝑚 + 𝛽 𝑀𝑎𝑟
Where wage is the dependant variable and years of schooling is the independent variable. As explained in the literature review, the parameter experience can be based on three different proxies: “Age”, “Potential Work Experience” and “Actual Work Experience”. As Mincer (1974B) suggested, “Actual Work Experience” was used as a proxy for experience since it accounts for possible discontinuity of an individual’s labour market participation.
In accordance with the IV method the following models were estimated: TABLE I
Conversion Table Years of Degrees to Years of Schooling
Primary education 8 years
LBO/MAVO/VBO 12 years
MBO/HAVO/VWO 14 years
HBO 17 years
10 (2) 𝑆 = 𝛿 + 𝛾 𝐿𝑜𝑤𝐹 + 𝛾 𝐼𝑛𝑡𝑒𝑟𝑚𝐹 + 𝛾 𝐻𝑖𝑔ℎ𝐹 + 𝛾 𝐿𝑜𝑤𝑀 + 𝛾 𝐼𝑛𝑡𝑒𝑟𝑚𝑀 + 𝛾 𝐻𝑖𝑔ℎ𝑀 + 𝛾 𝑍 + 𝜇
(3) ln 𝑤𝑎𝑔𝑒 = 𝛼 + 𝛽 𝑆 + 𝛽 𝑋 + 𝛽 𝑋 + 𝛽 𝐹𝑒𝑚 + 𝛽 𝑀𝑎𝑟 + 𝜀
The variable “Years of Schooling” (𝑆 ) was regressed on the instrument “Parental Education” which consists out of 8 dummy variables for both the mother and the father of the individual. 𝑍 represents personal characteristics such as marital status and sex. For the estimation, observations were dropped if individuals had 0 weekly working hours defined in their labour contract. Observations were also dropped if individuals were unemployed. These observations have been dropped as it was necessary to convert monthly earnings to hourly earnings.
To check for the existence of a supply effect, the fraction of highly educated labour was computed from population data provided by the CBS (2015). The fractions within the OSA 2000-2012 sample has been checked against the fractions computed from the population data for the Netherlands. This is to verify the representativeness of the dataset. Stata was used to determine the correlation between the returns to schooling and the population fractions of highly educated labour within the Dutch labour force. The sign and strength of this correlation was analysed in combination with information on the demand-side of the Labour market provided by other authors.
5. Results
The results are divided into several subsections. First, the OLS and IV estimates will be discussed. As will the first stage regression results. The two different methods are then subjected to statistical tests to determine the validity of the IV approach. With the results obtained it is then possible to estimate the correlation between the fraction of highly educated labour and the returns to schooling.
5.1 Ordinary Least Squares Estimates
The OLS estimates for the returns to schooling are showed in table II. The estimated parameters for the independent variables are all highly significant throughout the years except for the dummy variable “Married”, which shows a weak significance in the years 2000, 2010 and 2012. The estimated returns vary between 0.0778 and 0.0628. This is close to the estimates of Card (1993), Butcher & Case (1994), Angrist and Krueger (1991) that were reviewed in section two.
The coefficient of female is negative throughout the years as is found in other studies. The estimated parameter however varies a lot throughout the years. The estimated parameters on work
11 experience are also in line with the literature, they show that “Actual Work Experience” has a positive impact on wages. The diminishing effect of work experience is captured by ‘Actual Work Experience Squared’ which has a small but significant impact on earnings throughout the years. The significance of the dummy variable “Married” varies throughout the years, but it shows significance at the 10% level for every year estimated. The R-Squared of the model is lowest in 2000, valued at 0.1974. In the years thereafter the model shows a high increase in the R-Squared, where it is ranging from 0.34-0.41.
TABLE II
OLS Estimates for the Returns to Schooling of survey participants 2000-2012: OSA Labour Supply Panel
Independent Variable (1) 2000 (2) 2002 (3) 2004 (4) 2006 (5) 2008 (6) 2010 (7) 2012 Schooling 0.0696 0.0739 0.0778 0.0707 0.0716 0.0690 0.0628 0.0035*** 0.0027*** .0023*** 0.0022*** 0.0022*** 0.0025*** 0.0025***
Actual Work Experience 0.0220 0.0275 0.0322 0.0322 0.0249 0.0219 0.0241 0.0033*** 0.0024*** .0019*** 0.0019*** 0.0018*** 0.0013*** 0.0013***
Actual Work Experience- -0.0003 -0.0004 -0.0005 -0.0005 -0.0034 -0.0003 -.0003 Squared .00008*** 0.00005*** 0.00004*** 0.00004*** 0.00004*** 0.00003*** 0.00003*** Female -0.1046 - 0.0519 -0.0398 -0.0508 -0.0358 -0.0624 -0.0629 0.0174*** 0.0134*** 0.0108*** 0.0107*** 0.1006*** 0.0114*** 0.0114*** Married 0.0509 0.0674 0.0627 0.0626 0.0334 0.0360 0.0345 0.0194* 0.0151*** 0.1244*** 0.01225*** 0.0115** 0.0137* 0.0132* Constant 1.0259 0.9867 0.8663 1.0208 1.1322 1.2507 1.3610 0.0594*** 0.0463*** 0.0378*** 0.0370*** 0.0381*** 0.0383*** 0.0396*** N 2573 3069 2813 3141 3025 2885 2848 F 127.57 243.77 402.82 382.17 309.35 370.99 352.65 Adjusted R^2 0.1974 0.2835 0.4167 0.3777 0.3377 0.3908 0.3818
12 5.2 Instrumental Variable Estimates
The results of the IV regressions can be found in table III. The returns to schooling appears to be higher with comparison to the OLS estimates. Another noteworthy aspect is that the married coefficient has a weaker significance in the 2000, 2010 and 2012 years when compared to the estimates under the OLS technique. The size of the returns to schooling is close to findings found by Card (1993) and Angrist & Krueger (1991). Card (1993) estimated the returns to schooling at 0.117 while Angrist and Krueger (1991) estimated the returns to schooling at 0.1007. The estimates do not come close to the estimate of 0.184 by Butcher and Case. The OLS estimates and the IV estimates are much higher when compared to the estimate of 0.05 from Levin and Plug in 1999.
The first stage regression results are reported in Table VI in the appendix. The “low level of education” dummy for the father has been insignificant at the 5% level throughout the whole sample. The “low level of education” dummy for the mother only becomes significant in the years 2010 and 2012. The “higher level of education” dummy for the father is significant at the 5% level for almost every year, with the exception of 2008 and 2010. The first stage regressions for the years 2004, 2006
TABLE III
IV Estimates for the Returns to Schooling of survey participants 2000-2012: OSA Labour Supply Panel
Independent Variable (1) 2000 (2) 2002 (3) 2004 (4) 2006 (5) 2008 (6) 2010 (7) 2012
Schooling 0.1022 0.0947 0.0975 0.0855 0.0919 0.0817 0.0852
0.0135*** 0.0100*** 0.0105*** 0.0111*** 0.0088*** 0.0090*** 0.0084** Actual Work Experience 0.0196 0.0265 0.0310 0.0309 0.0243 0.0203 0.0213
0.0035*** 0.0025*** 0.0021*** 0.0021*** 0.0019*** 0.0017*** 0.0016** Actual Work Experience- -0.0002 -0.0003 -0.0004 -0.0005 -0.0003 -0.0002 -0.0003 Squared 0.00008 0.00006*** 0.00005*** 0.00005*** 0.00004*** 0.00004*** 0.00004** Female -0.0990 -0.0486 -0.0377 -0.0476 -0.0348 -0.0599 -0.0571 0.0179*** .0136*** 0.0110*** 0.0111*** 0.0102*** 0.0116*** 0.0117** Married 0.0651 0.0733 0.0672 0.0671 0.0387 0.0412 0.0426 0.0206** 0.0154*** 0.0128*** 0.0128*** 0.0119*** 0.0142** 0.0137** Constant 0.5612 0.6820 0.5756 0.8063 0.8229 1.0675 1.0342 0.1958** 0.1474*** 0.1563*** 0.0162*** 0.1350*** 0.1310*** 0.1225** N 2574 3068 2813 3141 3025 2885 2848 F 57.25 115.34 178.37 192.89 111.73 228.70 245.72 Adjusted-R^2 0.1699 0.2700 0.4009 0.3691 0.3184 0.3852 0.3649 F-Value of Instruments 32.44 42.80 23.13 22.43 33.72 39.05 49.77
13 and 2008 provides less significant estimates than other years. This is because parental education was not available for these years and it had to be imported from the years 2000, 2002, 2010 and 2012. The result was that there were many observations that did not account for the parental level of education which then got labelled as “Unknown level of parental education”. This could hamper the efficiency of the first stage regressions for these years. The size of the coefficients are according to theory. Low levels of parental education adds less value to the amount of schooling than intermediate and higher levels.
5.3 Test results on Instruments, Endogeneity and Coefficients
It is necessary to perform certain tests to demonstrate the validity of the IV procedure and to verify if the IV results can be used to verify the existence of a supply effect. The first test performed is to check if there is a strong enough correlation between the endogenous variable “Schooling” and our chosen instrument of Parental Education. This has been tested by doing a joint test of coefficients on equation 2, which estimates schooling based on the instrument and the exogenous variables “Female”, “Actual Work Experience” and “Marital Status”. Literature states that the found F-value in case of a single endogenous regressor (Schooling) should be greater than ten. The found F-values are shown in table III and are much greater than the value required to prove the instruments are valid. Thus there is sufficient evidence to state that the instrument “Parental Education” has a strong correlation with the endogenous variable Schooling.
The evidence found on the strength of the instrument “Parental Education” allows to perform a test to check for the endogeneity of the variable Schooling. The Wu-Hausman test provides this by testing the efficiency of one estimator against that of another. The null hypothesis states that both estimators are consistent but the OLS estimate is more efficient. The Wu-Hausman test can be used to check for endogeneity with the knowledge that an endogenous variable provides a biased and inconsistent estimate. One of the key assumptions to perform the Wu-Hausman test is that errors should be homoscedastic, if this is not the case the validity of the standard errors of the OLS and IV estimate would be compromised. This assumption has been checked by plotting the residuals of every year
*,**,***, implies significance at the 10, 5 and 1% levels respectively. Parental Education has been used as an instrument.
TABLE IV
Results Statistical Tests: OSA Labour Supply Panel (1) 2000 (2) 2002 (3) 2004 (4) 2006 (5) 2008 (6) 2010 (7) 2012 Wu-Haussmann Test on Endogeneity Reported P-values 0.01115** 0.02798** 0.05181* 0.17106 0.01526** 0.14168 0.00420*** T-test on IV estimates of the Returns 23.92*** -10.47*** 42.72*** -25.07*** 44.05*** -15.22***
14 separately and verifying their homoscedasticity. The p-value of the Wu-Hausman test has been reported for all the years in table IV. The test shows that at a significance level of 5% the years 2000, 2002, 2008 and 2012 have found sufficient evidence on the endogeneity of schooling. The year 2004 is significant at the 10% level while the years 2006 and 2010 are not significant. The results for the majority of the years imply that the OLS estimate is inconsistent when compared to that of the IV counterpart. Therefore we will use the IV estimates of the returns to schooling when checking for a supply effect of education on the returns to schooling.
The next step is to subject the IV estimates to a two sample t-test to verify that the found estimates differ from each other. This will be checked to verify that the returns to schooling have not been constant throughout some years. The key assumption behind a t-test is that the data should be normally distributed. This assumption has been checked by plotting the residuals in Stata, therefore, the data fulfils this requirement. The found t-values are reported in table IV. The t-values are highly significant for each subsequent year which dismisses the notion that the estimates of the returns to schooling could have remained the same for a subsequent pair of years.
5.4 Correlation between the fraction of Highly Educated Labour and the Returns to Schooling The final step is to check if the changes in the returns to schooling are influenced by fluctuations in the fraction of highly educated labour within the total labour force. The OSA states that they have persevered a representative dataset for Dutch labour force. To verify this, the fraction of highly educated labour within the labour force has been calculated on the sample for each corresponding year. The fraction of highly educated labour within the labour force has also been calculated using data on the Dutch labour force provided by the Statline database of the Central Bureau of Statistics (2015). The values of these fractions are reported in table V. The sample data over represents highly educated labour in the years 2002, 2004 and 2008, while underrepresenting this group by a small margin in the remaining years. Over
A correlation of -0.8105 has been found between the population fraction of High Educated Labour within the Total Dutch Labour Force and the Returns to Schooling.
TABLE V
Fraction Highly Educated Labour within the Sample against the Returns to Schooling (IV Estimates) (1) 2000 (2) 2002 (3) 2004 (4) 2006 (5) 2008 (6) 2010 (7) 2012 Fraction Highly Educated
Labour
Population 0.2704 0.2638 0.2984 0.3080 0.3221 0.3403 0.3427
Sample 0.2706 0.3036 0.3119 0.2986 0.3353 0.3227 0.3415
15 the years 2000-2012, the population fractions show an increase in the supply of highly educated labour ranging from 0.2704 to 0.3427. These data points show an increase in the fraction of highly educated labour of 26.74% over the span of twelve years. One particular abnormality is the year 2002 where the supply of highly educated labour has dropped. However, the majority of the years reveal an increasing supply of highly educated labour. The returns to schooling estimated on the OSA sample range between 0.1022 and 0.0852. The estimated parameter shows an overall decrease in value over the years 20002012. Using the population fractions and the returns to schooling a correlation has been found of -0.8105. This shows that there is indeed a strong negative supply effect as according to economic theory of supply and demand.
6. Discussion
The methodology used in this research will be under revision and discussed on basis of internal validity. There are several sources of potential bias that need revision to determine the quality of the results. Second, the results will be compared with other studies that estimate the returns to schooling. In addition with information on demand factors of the labour market, the external validity of the decreasing returns to schooling will be discussed as will the implications this has for the found correlation between the fractions of highly educated labour and the returns to schooling.
6.1 Bias of results
The first source of bias is the endogeneity of the schooling variable in the earnings equation. With the use of ‘Parental Education” as an instrument, the problem of endogeneity of the schooling variable has been solved. The instruments are significant as reported in table III and the results of the Haussmann test reported in table IV provides evidence for inconsistency in estimations when estimating the returns with OLS.
The second source of bias can be found in the definition of earnings when calculating the returns to schooling. The number of hours employees work is not considered when using monthly wages as dependant variable. This could lead to inflated parameters as highly educated employees tend to work more hours than the lower educated ones (Card, 1999, p.1809). Transmuting the dependant variable from monthly to hourly wages can account for this effect.
The third source of potential bias is found in the measurement error. This is difficult to eliminate as self-reported values such as earnings can be misreported by participants of the OSA labour supply panel. Furthermore, the size of this bias is unknown and could therefore lead to inconsistent estimates. Since the schooling variable is measured with the proxy “Highest Obtained Degree”, it is more difficult
16 to misreport this variable. Using the actual years of schooling as a survey question could result in misreporting the years of schooling by a few years. But misreporting the highest obtained degree is less likely since an individual would not be confused by the number of years. A flaw of the current approach in the methodology is that the returns to schooling does not represent the causal effect of a year of schooling but rather the causal effect of the returns to a degree. The proxy used does not recognise the possibility that individuals went to school longer than the regular time frame of their degree. It could be the case that some individuals have enjoyed extra years of schooling, or an extra degree without this being captured by the proxy “Highest Obtained Degree”. With the current use of “Highest Degree Obtained”, individuals with double degrees or extra years of schooling are valued at the same rate as their counterpart who obtained a single degree without extra years of schooling. However the methodology used to estimate the returns does not account for this phenomenon.
Another issue with the current methodology is that the jobs of survey participants have not been categorised in occupational sectors. This could be an issue because the public sector of the labour market could reflect wages that are not the same as the market wages in the private sector (Psacharopoulus & Patrinos, 2010). However, based on the assumption that individuals are rational we can derive the argument that an individual would not opt to work for the public sector if it would make him or her worse off. Moreover, the returns to schooling also depends on the labour demand of a certain sector (Gunderson & Oreopoulos, 2010). For example, graduates from technical studies are in shortage when compared to graduates from other sectors. This would probably result in wage differentials between individuals with the same amount of schooling but employed in different sectors. Based on the limitations discussed, the results found could reflect the average causal returns of a degree instead of the causal returns to schooling.
Another point worth mentioning is that information on parental education has not been available for all observations. These unaccounted observations were labelled as having unknown parental education levels. The complete structure on parental education is reported in table VII of the appendix. The years 2000-2008 lacked more data than recent years. For example, the year 2000 is missing data on 43.87% of the total participants with regard to parental education levels, while this percentage was just 9.46% in 2012. Although the amount of unknown parental education levels are high in earlier years, the F value of instruments is still greater than 10. However, the IV method takes missing data into account and incorporates this in the standard errors of the parameters. If the missing data would be available it would probably only lead to lower standard errors and a stronger significance of the estimated parameters.
A final suggestion for the current methodology is to introduce an interaction variable between “Female” and “Marital Status”. The current method assumes that if one is married the potential difference in earnings will be the same for males and females. However, this parameter could be different for females and males and the introduction of an interaction term could result in a higher significance of the estimated parameter “Marital Status”.
17 6.2 Other studies on the Returns to Schooling over time and the Dutch Labour Market
There are a few studies on the returns to schooling that could have an implication on the findings in this thesis. One study to be discussed is the study of Card and Lemieux (2001) that researched the combined effect of age cohorts and the returns to schooling over time for young male college graduates. Card & Lemieux (2001) report the returns to schooling different time periods for the United States (1959-1999), the United Kingdom (1975-1995) and Canada (1980-1995) and found that in all three countries the returns to schooling have been increasing over time in spite of the increases in average education levels of the labour force. These findings suggest that the demand for educated labour has been rising faster than the supply of educated labour. Another reason for this increase in returns is provided by Gunderson & Oreapoulos (2010) who argue that the skilled-unskilled wage differential has widened over time. If these factors are also in play for the Dutch labour market the difference in results of this study could be due to a different period of time. The increase in demand could have slowed down at the beginning of the 21st century which could explain the difference in results. The current results show an increase in supply and a decline in the returns to schooling which suggest that the supply increase outstrips any potential fluctuations in demand.
Another study from Psacharopoulus & Patrinos (2010) combines various studies and estimates on the returns to schooling for males carried out across 49 countries. For each country they report three to four estimated parameters from different studies in the timeframe of 1962-1996. Before they adopt an estimation in their evaluation of global returns to schooling Psacharopoulus & Patrinos (2010) evaluate the quality of the methodology of each eligible estimate. One of the criteria’s insisted upon by the authors is that the estimated parameter should always be based on a representative dataset for a country’s population. They found that average global returns to schooling for males have decreased by 0.6%, while the average education obtained by males has increased. They concluded that the supply of education did indeed have an effect on the returns to schooling (Psacharopoulus & Patrinos, 2010). However, the problem is that the crude pooling of parameters together does not take into account the different structures of labour markets and educational institutions throughout the world. However it is one of the few studies that argues that the supply effect can be observed throughout the period 1962-1996 in spite of an increasing demand for educated labour and a widening of the skilled-unskilled wage differential.
Information on the demand side of the Dutch labour market is scarce. However in a policy letter to the government, economist Ter Weel (2012) evaluates wage disparity in the Netherlands and the supply and demand side of the labour market. He argues that the demand for educated labour has risen in recent years and finds that technological investment is positively correlated with the wages of highly educated labour. However, it is negatively correlated with wages of intermediate educated labour, which implies that technological advancement has led to an outsourcing of intermediate level jobs, whilst also increasing the demand for highly educated jobs. Ter Weel (2012) further strengthens his argument by
18 stating that highly educated jobs and technological investments are compliments. As an example, computers have taken a place in current work environment which demands an analytical capacity of employees (Ter Weel, 2012).
The influence of technology on the labour market is further explored by Goos, Manning and Salomons (2009) who found that employment has been increasing in the low and high skilled sector of the labour market while the opposite is true for the middle skilled sector. Goos et al. (2009) offer a different yet similar hypothesis as Ter Weel (2012) for the cause of this skill biased technological change: the more routine tasks an occupation has, the more it is subjected to outsourcing and thus a lower demand on the labour market. Furthermore, they discuss the shares of the total hours worked for low, middle and high-paying occupations for several European countries, including the Netherlands, over the period of 1993-2006. Goos et al. found a clear decrease for each country in the shares of total hours worked for middle paying occupations. They also found that the shares of total hours worked has increased for low and high-paying jobs throughout Europe. However, for the Netherlands the share increase for high-paying occupations over the time period 1993-2006 has been lower than the total European increase.
The current results imply that the returns to schooling have decreased over the years 2000-2012. This decrease is probably the result of an increase in supply of highly educated labour in the Netherlands, which can be based on the found correlation of -0.8105 between the fraction of highly educated labour and the estimated returns to schooling. However the information provided by Ter Weel (2012) implies that the demand for highly educated labour has risen over the years. With an increase in demand and supply it is possible to argue that the found correlation of -0.8105 is weaker than the actual correlation between these variables. That is, the supply effect is stronger and outperforms the demand effect in the labour market. Furthermore the data reported by Goos et al. (2009) implies that the shares of total hours worked of high-paying occupations in the Netherlands has risen for the period 1993-2006. Under the assumption that earnings are strongly correlated with education, the argument is possible that demand for highly educated labour has indeed increased in the Netherlands. However, this increase in demand will be lower than the overall European increase in demand for highly educated labour. This implies that it is certainly possible that the supply effect in the Netherlands has outperformed the demand effect for highly educated labour in the Dutch labour market.
7. Conclusion
In this thesis an effort has been made to verify if there has been a supply effect of the increase in highly educated labour on the returns to schooling in the Netherlands. To verify the existence of this effect it was necessary to estimate the returns to schooling. Instrumental Variable regression has been used to
19 estimate the returns to schooling due to the possible endogeneity of the variable “Years of Schooling”. The instrument “Parental Education Level” has been chosen as an instrument as it is believed to be a valid instrument.
The found IV estimates of the returns to schooling show an overall decline from 0.1022 to 0.0852 for the time period of 2000-2012. Using population data to compute the fraction of highly educated labour within the labour force show an increase from 0.2704 to 0.3427. A correlation between these two parameters show a correlation of -0.8105. The correlation found is in line with economic theory which states that an increase in supply would have a negative effect on the price of a good. The joint F-test on coefficients reports values over the critique value of 10 (Table III). These results that the instrument “Parental Education” has a strong correlation with the endogenous variable “Years of Schooling”. Furthermore the Wu-Hausman test results are in favour of the IV method implying that the majority of the years have an inconsistent OLS estimator. However, this initial result only based on the supply side of the labour market.
Literature on the demand side of the labour market argue that the demand for highly educated labour in the Netherlands has been rising in the period 2000-2012. However, the Dutch demand increase is believed to be lower than the European demand increase for educated labour. This could be the key factor why the estimated returns to schooling for the Netherlands are decreasing while similar studies for other countries show an increase in returns. There are certain limitations to the results, there is the issue that a proxy for “Years of Schooling” has been used which could mean that the found returns are not the returns to schooling but rather the returns to a degree. Also, the instrument “Parental Education” was not available for a significant fraction of the sample, this made the estimations less precise. Furthermore, jobs of survey participants have not been categorised. This makes it impossible to account for different sectors in the labour market, such as the public sector of the technical sector who are believed to have different demand and supply curves. This limitations imply that the returns estimated are the average returns to a degree, rather than specific private returns to schooling.
Based on the results it can be argued that the returns to schooling have been decreasing in the Netherlands as a consequence of the increasing supply of educated labour. In 2015 the Dutch government will implement new regulations which abolishes the study grant system that was in place during the time period 1996-2014. The Dutch government will be introducing a student loan system which will increase the costs of studying for the Dutch citizens (CPB, 2013). This new system could lead to an exogenous shock that could decrease supply and therefore affect the returns to schooling. Future research could estimate the effects of this exogenous shock which would provide more insight on the supply effect of educated labour on the returns to schooling.
20
8. Reference list
Angrist, J., & Krueger, A. (1991). Does Compulsory School Attendance Affect Schooling and Earnings? The Quarterly Journal of Economics, 106, 979-1014.
Blackburn, M., & Neumark, D. (1995). Are OLS Estimates of the Return to Schooling Biased Downward? Another Look. The Review of Economics and Statistics, 77, 217-229. Borjas, G. (2013). Human Capital. In Labor economics (6th ed.). New York: McGraw-Hill.
Butcher, K., & Case, A. (1994). The Effect of Sibling Sex Composition on Women's Education and Earnings. The Quarterly Journal of Economics, 6, 531-563.
CPB. (2013, June 6). Aflossing Studieschuld bij Sociaal leenstelsel. Retrieved June 24, 2015, from http://www.cpb.nl/publicatie/aflossing-studieschuld-bij-sociaal-leenstelsel
Centraal Bureau voor de Statistiek. (2012). Jaarboek Onderwijs In Cijfers. Retrieved June 28, 2015, from
http://www.cbs.nl/NR/rdonlyres/3036B4E1-A671-4C9E-95BF-90C0493B4CD9/0/2012f162pub.pdf
Centraal Bureau voor de Statistiek(2015). Retrieved June 25, 2015, from
http://statline.cbs.nl/Statweb/publication/?DM=SLNL&PA=71882ned&D1=1&D2=0&D3=0,
10-15&D4=200,203,205,207,209,211,213&HDR=G3&STB=G1,G2,T&VW=T Card, D. (1993). Using Geographic Variation in College Proximity to Estimate the Return to
Schooling. (Working Paper 4483).
Card, D. (1995). Earnings, Schooling, and Ability Revisited. Research in Labor Economics,14, 23-48. Card, D. (1999). The Causal Effect of Education on Earnings (O. Ashenfelter & D. Card, Eds.).
Handbook of Labor Economics, 3(A), 1801-1863.
Card, D., & Lemieux, T. (2001). Can Falling Supply Explain the Rising Return to College for
Younger Men? A Cohort-Based Analysis. The Quarterly Journal of Economics, 116, 705-746.
Goos, M., Manning, A., & Salomons, A. (2009). Job Polarization in Europe. American Economic Review, 99(2), 58-63.
Gunderson, M., & Oreopoulos, P. (2010). Returns to Education in Developed Countries. In D. Brewer & P. McEwan (Eds.), Economics of education (1st ed., pp. 37-43). Oxford, UK: Academic Press.
Jongbloed, B. (2004). Tuition Fees in Europe and Australasia: Theory, Trends and Policies. In Higher education: Handbook of theory and research. Vol. 19, 241-310, Dordrecht: Kluwer Academic. Kalwij, A. (2000). Estimating the Economic Return to Schooling on the Basis of Panel Data.
Applied Economics, 32(1), 61-71.
Kling, J. (2000). Interpreting Instrumental Variables Estimates of the Returns to Schooling. Journal of Business & Economic Statistics, 358-364.
21 Lemieux, T. (2006). The “Mincer Equation” Thirty Years After Schooling, Experience, and
Earnings. In S. Shechtman (Ed.), Jacob Mincer a Pioneer of Modern Labor Economics (pp. 127-145). New York, NY: Springer Science Business Media.
Levin, J., & Plug, E. (1999). Instrumenting Education and the Returns to Schooling in the Netherlands. Labour Economics, 6, 521-534.
Miller, C. (1993). Actual experience, Potential Experience or Age, and Labor Force Participation by Married Women. Atlantic Economic Journal, 21(4), 60-66.
Mincer, J. (1974A). The Human Capital Earnings Function. In Schooling, experience, and earnings (pp. 83-96). New York: National Bureau of Economic Research; distributed by Columbia University Press.
Mincer, J. (1974B). Summary and Findings. In Schooling, Experience, and Earnings (pp. 127- 145). New York: National Bureau of Economic Research; distributed by Columbia University Press. OSA. (n.d.). Arbeidsaanbodpanel 1985 t/m 2012 - EASY. Retrieved June 25, 2015, from
https://easy.dans.knaw.nl/ui/datasets/id/easy-dataset:60607
Plug, E.J.S. (2001). Season of birth, Schooling and Earnings. Journal of Economic Psychology, 22, 641-660.
Plug, E. (Director) (2015, January 1). Human Capital (and the Education Production Function). Lecture conducted from University of Amsterdam, Amsterdam.
Psacharopoulos, G., & Patrinos, H. (2010). Returns to Investment in Education: A Further Update. Education Economics, 12:2, 111-134. doi:10.1080/0964529042000239140
Ritzen, J. (1996, March 28). Law Change With Regard to Student Grants 1996. Retrieved June 24, 2015, from http://wetten.overheid.nl/BWBR0007972/geldigheidsdatum_24-06-2015
Ter Weel, B. (2012). Loonongelijkheid in Nederland stijgt. Retrieved June 24, 2015, from http://www.cpb.nl/publicatie/loonongelijkheid-nederland-stijgt
22
9. Appendix
TABLE VI
First Stage Regression of Parental Education on Schooling, of survey participants 2000-2012: OSA Labour Supply Panel Independent Variable (1) 2000 (2) 2002 (3) 2004 (4) 2006 (5) 2008 (6) 2010 (7) 2012 Education Level Low -0.3355 -0.2799 -0.0220 0.0695 -0.1879 -3.8090 0.2637
Father 0.02745 0.2009 0.3525 0.4857 0.6570 2.1540* 0.1812
Education Level 0.5972 0.5424 0.7450 0 .7701 0.3729 -3.1167 0.6447 Intermediate Father 0.3001** 0.2178** 0.3661** 0.4930 0.6635 2.1558 0.1925***
Education Level Higher 1.1974 1.0422 1.2989 1.3478 1.2036 -2.5065 1.4038 Father 0.3198*** 0.2323*** 0.3749*** 0.4948*** 0.6658* 2.1570 0.1954***
Education Level Low 0.3004 .2903 -0.3112 -0.3468 -0.2757 4.5615 0.9303 Mother 0.2719 .2017 0.3540 0.4859 0.6571 2.2412** 0.1918***
Education Level 1.0121 1.0195 0.1170 -0.0707 0.2682 4.9366 1.3291 Intermediate Mother 0.3216*** .2364*** 0.3743 0.4996 0.6688 2.2439** 0.2117***
Education Level Higher 1.4433 1.4217 0.3827 -0.0567 0.3159 5.2704 1.6191 Mother 0.4012*** .2917*** 0.4194 0.5230 0.6809 2.2464** 0.2349***
N 2574 3068 2813 3141 3025 2885 2848
F 29.10 37.38 24.47 23.39 33.24 49.87 57.97
Adjusted-R^2 0.0985 0.1061 0.0770 0.0666 0.0963 0.1449 0.1667
Standard Errors in Italics, *, ** and *** implies significance at the 10%, 5% and 1% levels, respectively. Parental Education Levels are used as an instrument. The Table excludes the exogenous regressors Fem, Actwex, Actwex2, Mar and the Intercept.
23 TABLE VII
Frequency of Parental Education Levels: OSA Labour Supply Panel 2000-2012
Independent Variable (1) 2000 (2) 2002 (3) 2004 (4) 2006 (5) 2008 (6) 2010 (7) 2012 Education Level Father
Low 975 1510 1251 1436 1583 1750 1454 % 37.85% 48.95% 44.33% 45.54% 49.22% 60.53% 50.96% Intermediate 271 448 376 432 544 631 557 % 10.52% 14.52% 13.32% 13.7% 16.92% 21.83% 19.52% Higher 200 352 275 384 430 493 572 % 7.76% 11.41% 9.74% 12.18% 13.37% 17.05% 20.05% Unknown 1130 774 920 901 659 17 270 % 43.87% 25.12% 32.6% 28.58% 20.49% 0.59% 9.46%
Education Level Mother
Low 1197 1895 1541 1825 1932 2120 1881 % 46.47% 61.43 54.61% 57.88% 60.07% 73.33% 65.93% Intermediate 185 316 285 301 418 493 481 % 7.18% 10.24 10.10% 9.55% 13% 17.05% 16.86% Higher 70 129 101 136 209 262 263 % 2.72% 4.18 3.58% 4.31% 6.5% 9.06% 9.22% Unknown 1124 745 895 891 657 16 228 % 43.63% 24.15 31.72% 28.26% 20.43% 0.55% 7.99% N 2577 3085 2822 3153 3216 2891 2853
