The effect of class size on children’s academic
performance
Master’s Thesis
University of Groningen
Faculty of Economics and Business
Abstract
This thesis evaluates the effects of the class sizes on the academic performance of children. Through the use house prices in the Netherlands, an exogenous instrument, this thesis attempts to solve the endogeneity issues associated with the relationship of class size and academic performance; better schools attract more students so they have to make bigger classes. Through the use of a panel dataset of all Dutch schools and their characteristics, this thesis found a negative effect of class size on the academic performance, but the results are not significant. Nevertheless, this thesis suggest some concrete possibilities for further research.
Table of Contents Table of Contents ... 2 1. Introduction ... 3 2. Literature Review ... 5 3. Institutional Setting ... 6 4. Data ... 9 4.1. Data Source ... 10 4.2. Main Variables ... 10 4.3. Sample restrictions ... 12 4.4. Descriptive Statistics ... 12 5. Methodology ... 13 5.1. Non-linear models ... 15 5.2. Instrument validity ... 16 5.3. Robustness tests ... 17 6. Results ... 18 6.1. Non-linear model ... 22 6.2. Robustness tests ... 26 6.3. Interpretation of results... 29 7. Conclusion ... 32 References ... 33
Appendix A; District redrawing in Amsterdam ... 36
1. Introduction
In this thesis, the effect of class sizes on the academic performance of children is investigated. Intuitively, many people feel that smaller classes result in a better development of children since there is more individual attention. However, schools that perform better, attract more students resulting in bigger classes, whereas ill performing schools have difficulty attracting students, so class sizes decrease. Furthermore, there is a severe degree of sorting in the choice of schools in the US and the UK through the existence of school districts; schools located in richer districts have students from richer parents with more financial capacity to encourage development of their children through extracurricular activities. These problems pose endogeneity issues when testing this relationship. Many studies have attempted to estimate the impact of class size on school performance while accounting for these endogeneity issues. They find a significant positive effect of smaller classes in the short run (Angrist & Lavy, 1999), and in the long run (Fredriksson, Öckert, & Oosterbeek, 2013; Chetty, et al., 2011; Krueger & Whitmore, 2001).
The STAR project is often used in studies due to its experimental design; it is a randomised control trial where students are randomly allocated to different class sizes. Finn and Achilles (1999) and Krueger and Whitmore (2001) are examples of papers that use the STAR project to find the effect of class size on academic performance. Other authors make use of an IV analysis or a regression discontinuity design to find the effect of class size on academic performance, on college entrance rates and on earnings later in life. Fredriksson, Öckert and Oosterbeek (2013) and Angrist and Lavy (1999) use the legal limits on the maximum class size to find the effect of bigger classes. Fredriksson, Öckert and Oosterbeek (2013) use the Swedish system where, in upper primary school, classes bigger than 30 students are not permitted. Angrist and Lavy (1999) base their IV estimate on Israeli law where Maimonides’ rule prevents classes of more than 40 students.
the quality of schooling and house prices, the Dutch, with their free school choice, are less prone to a relationship between house prices and academic performance. Since parents can choose between schools, there is more competition between schools to offer good programs and teachers. Furthermore, in the Netherlands there is no reliance of schools on local taxation; schools in richer districts do not have additional funding available, like in the US, and schools work with similar budgets irrespective of their location.
Furthermore, the use of collective labour agreements creates an equal wage structure among teachers across different locations. However, this also implies that schools located in expensive districts are unable to increase wages to attract skilled teachers. In the long run, this means that richer districts, with more expensive housing, are less attractive for teachers. Since their profession is not location specific, teachers that cannot afford housing, are more likely to move and work elsewhere. This ultimately decreases the availability of teachers in an area with high housing prices, ultimately impacting the size of classes. This thus presents an exogenous instrument for class size.
Through the use of this exogenous instrument, this thesis attempts to identify the effect of class size on the academic achievement of primary school pupils. Dutch pupils are obliged to take a government approved and school independent test at the end of primary school. These test results are used as a measure of academic performance at the end of primary education. Through the use of a spatial weight matrix, this thesis further accounts for the travel possibilities teachers have.
The main results found are insignificant; they indicate a higher likelihood of a negative relationship between class size and academic performance but there is no significant evidence for this negative relationship. This is in contrast to what the main literature suggests. Possible reasons for this difference are the modest time frame used, the differences in the size of districts, unaccounted for property sizes, a potentially non-representative independent variable, and a weak instrument.
2. Literature Review
Several studies have investigated the effect of class size on the academic performance of students. Some studies have even attempted to assess the long-term results of class size on several parameters later in life. Two studies use a regression discontinuity design or instrumental variable approach through using the maximum students legally allowed in a class. In Israel, a twelfth century rabbinic scholar, Maimonides, suggested a maximum class size of 40 students. Maimonides’ rule is still used, limiting the maximum students in an Israeli class to 40. This rule generates a potentially exogenous source of variation and was used by Angrist and Lavy (1999) to estimate the effects of class size. They find a negative association between larger classes and academic performance with the biggest improvement in math scores. A more recent study, that uses a regression discontinuity design based on a maximum of 30 students, in Sweden finds that smaller classes in the later years of primary school, improve cognitive ability tests at the age of 13 and at the age of 16 (Fredriksson, Öckert, & Oosterbeek, 2013). Furthermore, this study finds that smaller classes have an effect on the total level of education completed, as well as wages and earnings in adult life. A cost benefits analysis resulted in an internal rate of return of 18% if classes are reduced from 25 to 20 students.
One of the misfortunes of the STAR program is the minimum school size required for schools to participate in the experiment. A school needs to have at least three classes in each grade, this requires schools to have a minimum size of 60 students per grade to be able to partake in the experiment. Another fallacy of the STAR experiment is that students first located in the bigger classes, were more likely to be sorted in the smallest classes in primary school after the experiment ended. This results into imprecise results at later ages.
The research on the effect of housing prices on school performance is limited because of the reverse causality issue present in the US. However, Blau and Haurin (2017) find that there are small but insignificant net effects on the price of housing on child and young adult outcomes. They find larger effects for some specific population subgroups, like racial minorities, and for certain parameters of later life success but these are exceptions. Another study by Schwartz, Voicu and Mertens Horn (2014) on the effect of choice schools1 on housing prices in New York City finds that “the opening of a choice school weakens the relationship between housing values and the zoned elementary school” (p. 9). They also find that a choice school increases property values in general. Although, the main goal of the paper is not on the academic effect, they also find a significant positive effect of the opening of the choice school on the academic performance, documented via an increase in pass rates, of the schools. However, these two studies fail to identify a clear effect of housing prices on the academic performance of children in general. These studies support that the suggested instrument is exogenous and not endogenously determined.
3. Institutional Setting
In this section, I describe the institutional setting of the education system in the Netherlands. This description is largely based on the work of Renkema, Mulder, and Barnard (2016). Dutch children go to primary school from the age of 4 onwards. The first two years focus more on the adaptation of skills and broad knowledge gain and tend not to focus on the traditional subjects like mathematics and grammar. A focus is placed on learning through playing. After two years, at the age of 6, children move to grade 3 where they start to learn reading, writing and mathematics along with other common subjects. Primary school ends after grade 8, at the age
of 12, when students start secondary education. In this year all students do one of several independent tests approved by the government and chosen by the school to have their level of knowledge and skills independently assessed before the start of secondary school. Based on this test and the advice of their primary school teacher(s), students are admitted to the different types of secondary education paving the path for their later career.
The CITO test is the biggest of these independent tests. In this test students score between 500 and 550 with 550 being an excellent performance. The test consists of two mandatory parts, arithmetic and the Dutch language, and optional part; world orientation. These parts are further divided into different domains. For the Dutch language these domains are; reading, writing, and grammar. For arithmetic they are; numbers, proportions, geometry, and connections. Lastly, for world orientation the domains are geography, history, biology, and society. The three parts are each tested on a different day, making the total test length two or three days depending on the participation of the school in world orientation part. The exact length and duration of the test are dependent on the amount of questions which may differ each year, based on the different requirements of the Dutch government regarding the tested subjects (Ministerie van Onderwijs, Cultuur en Wetenschap, 2015; 2016; 2017; 2018).
About two thirds of all primary school are nongovernmental, most are Protestant or Roman Catholic, both 30% of total schools. Other main orientations include Evangelical, Jewish, Islamic, Hindu and Anthroposofic. This thesis distinquishes between five types in order to control for the specific types: Public, Protestant, Roman Catholic, other religious and other non-religious. Lastly, there is a small group of cooperation schools, combinations of public and nongovernmental schools, they are allocated to the public education group which is most suitable according to Renkema, Mulder and Barnard (2016).
One of the main consequences of this duality system is the absence of school districts. Students are, unlike in the US and the UK, not restricted to a (selection) of school(s) based on the location of their house and parents are free to choose to send their children to any school of their preference. Theoretically, it is possible to send your child to school at the other side of the country. However, often a school is choosen from a subset relatively close to the house so that travel time is limited, but nevertheless, there is a degree of competition. It is precisely this free school choice that is crucial in the determination of an exogenous instrument. The free school choice gives rise to a certain level of competition between schools so that all schools try to offer the best programs, teachers and teaching methods. Additionally, funding is not dependent on local taxations but only dependent on the national government subsidy given to schools. Together, these two imply that district characteristics are not a final determinant in the composition of schools. In the US, students in richer neighbourhoods, with higher housing prices, perform significantly better than students in poorer neighbourhoods. However, since local taxation partially determines the schools funding, there is a relationship between the quality of school and the wealth in a school district (Loubert, 2005). These richer schools have more financial liberty and options. Also, parents base their housing decision on the quality of schools2. In the Netherlands no such relationship exists because local taxes are not a determinant of schools and the school quality is not relevant in the housing decision because there are no clear school district by which parents are bound; they have a free choice.
It is imported to note that it is allowed for non-public schools to deny access to students who are not aligned with their orientation, however this is a rare event. Additonally, capacity constraints can force a school to have to reject children. If this is a regular occurance
2Many studies in the US focus on a relationship between school performance in terms of academic achievement
municipalities have the option to grant people priority access to their closest schools, in Amsterdam for example this means that children have priority to the top 8 closest schools (Gemeente Amsterdam, z.d.). However, a top 8 of schools still povides much more opportunities and choice than the district systems used in US and UK studies. Furthermore, this is only the case when capacity constraints are faced by schools and gives the freedom to apply to schools further away.
Apart from the government subsidies which are equal between schools based on observable characteristics, Dutch teachers also get paid according to a collective labour agreement (CAO). In this agreement the teacher salaries are negotiated for all primary school teachers in the Netherlands. Consequently, there are little differences in salary across different locations. The details of the agreement are complex but pay rises with experience and further qualifications. It is most relevant to stress they are not determined by location. Therefore, schools cannot compete based on salary to attract additional teachers and teachers have to factor in housing prices in their work decision. Not their salary but their occupation decision has limited them in the availability of mortgage and thus in their final housing decision. In districts where housing prices are too high, it is no longer possible for teachers to live as they are unable to afford the housing possibilities in that area so they move away. Consequently, it is expected that schools in the regions with high housing prices, have a scarcer labour pool to attract potential workers from. Since it is unlikely that it is possible to compete based on salary, it is expected that these schools have less teachers available and that therefore, classes are bigger.
4. Data
4.1. Data Source
The data set consists of all primary schools in the Netherlands in 2014 until 2017 acquired from DUO, The Education Executive Agency of the Dutch Ministry of Education, Culture and Science. DUO creates yearly files with the average academic performance of schools (DUO, 2018), files with the school’s student numbers (DUO, 2018a) and files with data on the school’s staff, including teachers, directors and support staff with their age spread (DUO, 2018b). All schools are assigned a unique number called BRINnummer by the government through which each school can be identified (DUO, 2019).
The data on the primary schools has been connected to public data from Statistics Netherlands (CBS) on neighbourhoods and city districts (CBS, 2019). Statistics Netherlands reports a broad range of variables on the neighbourhood and district level for previous years. To make the connection, a special file created by Statistics Netherlands for DUO in 2016 is used where all the individual addresses in the Netherlands are linked to the neighbourhood and city district codes3 (CBS, 2016). However, in 2016, Statistics Netherlands changed the definition of neighbourhoods and city districts in Amsterdam. Historically, Amsterdam had 7 boroughs which Statistics Netherlands defined as city districts in their analyses. The actual city districts in Amsterdam were therefore reported as neighbourhoods. In 2016, Statistics Netherlands decided to align Amsterdam with the rest of the Netherlands removing the boroughs from their reports. Simultaneously, some changes were made to the borders, this is shown in Appendix A. To have a consistent measure of local activity it means that for Amsterdam the neighbourhood level has been used in 2014 and 2015, whereas for 2016, 2017 and for the rest of the Netherlands, the city district level has been used.
4.2. Main Variables
To compare the academic achievement of schools, I look at the average of the independent tests all students have to take at the end of primary school. However, the Dutch government approves several tests every year out of which schools can choose and they use different scoring methods which causes comparability to be troublesome. It cannot be ruled out that the standard deviation of the tests is not comparable to each other and some of the tests are only taken by a small amount of schools resulting in imprecise means and standard deviations. The biggest test is the
3 For 103 schools the connection could not be made, for these the neighbourhood and district code was only
CITO test, which was taken by 81% of schools in 2014 and by 57% of schools in 2017. The schools partaking in the CITO test show significantly different characteristics; the number of students, and number of teachers are higher and class sizes are bigger in CITO taking schools whereas the share of students that is exempted from the final test is lower in CITO taking schools. Furthermore, most of the district characteristics in which the school is located are different; the number of inhabitants and the population density, the share of the population below the age of 14 and the share of immigrants in the total population are higher for CITO taking schools. The number of owner-occupied homes as opposed to rental property is higher for non-CITO schools. Lastly, the WOZ-value of the districts of CITO taking schools, is significantly higher than that of non-CITO schools. The full detail of the differences can be found in Appendix B. These differences are a potential source of the insignificant results found.
The variable of interest, class size, is unknown in this dataset. DUO only provides the number of students in their final year per school but does not provide data on the average class sizes. Using the number of students in the final year is problematic because it is unreported how many classes a school has in the final year. However, DUO does provide the total number of teaching staff and the total number of students in a school. Therefore, the pupil/teacher ratio of a school is used as an estimate of average class size in a school, which for the remainder of this thesis is called class size.
In order to identify the housing possibilities in a local district, this thesis makes use of the WOZ value (real estate valuation). All properties in the Netherlands have a WOZ value based on the selling price of houses in the close proximity with similar features. They are used by municipalities for taxing purposes so rented properties are also given a WOZ value. This has the advantage that the data used the average of all properties. The WOZ value is not a perfect estimator of the housing price since some features are not always independently observed or are unique for the property, like maintenance, and they may differ yearly. However, since the number of properties is large and an average is used, the law of large numbers assures that the WOZ-value is an accurate predictor of the house prices in the Netherlands.
4.3. Sample restrictions
All primary school are included in the dataset. However, some schools are omitted, in this section, these are discussed. Special primary schools for children with additional needs and disabilities are excluded from the dataset because students travel further distances for these schools. In order to ensure that all who need it can come to one of these schools, special transport is arranged for these children, for which parents can get government funding. This funding only covers the distance to the closest school with places available. Consequently, they do not have completely free school choice. Additionally, these children are taught in smaller classes, so student/teacher ratio is lower than in normal primary school because of their additional needs. Lastly, the academic achievement of these children, is due to their special circumstances, not comparable to children of normal schools.
Furthermore, schools that have multiple locations are omitted from the dataset as the data on the staff is only available for the whole organisation and not specified per location. Since all locations are not necessarily in the same district, city or municipality so they are removed from the dataset. Furthermore, there are 11 schools that classify as a driving school, meaning they do not stay in one location. Consequently, they cannot be matched to a particular location and thus have also been omitted from the dataset.
Lastly, in the creation of the spatial weight matrix, some schools could not be located accurately. Almost 500 observations are dropped to create the spatial weight matrix. Only schools that identified on the street, postcode or, preferably, house number level are included.
4.4. Descriptive Statistics
Table 1; descriptive statistics Mean SD SD-within Students 226.4 128.03 18.84 Managers 1.13 .77 .33 Teachers 12.27 7.58 1.78 Support Staff 1.50 2.71 1.21
Students final year 28.7 17.7 5.4
Students Exempted for the test 2.47% 10.05% 7.35%
CITO score 535.06 3.95 2.37
Number of inhabitants 13614 15295 746
Population density 2599 3386 262
WOZ-value 216 69 9
Class size (students/teachers) 18.51 3.91 1.87 % inhabitant younger than 14 17.1 3.3 .562
% western immigrants 8.5 4.7 .45
% non-western immigrants 10 12.4 .66
% immigrants from Morocco 1.73 3.5 .2 % immigrants from the Netherlands
Antilles and Aruba
.6 1.1 .13
% immigrants from Surinam 1.6 3.4 .19
% immigrants from Turkey 1.8 3.6 .21
% owner-occupied homes 61 14.8 1.5
% high income household 21.5 8.3 .76
% low income household 36.7 10.7 .7
% teachers younger than 35 31.9 17.7 7.2
Students per managers 203 145 88
Students per support staff 270 687 535 % teachers with a fixed contract 93 9.9 6.7
% female teachers 86 10.3 4.3
the district statistics are skewed to favour bigger districts. The descriptive statistics are per school implying that bigger districts, who tend to have more schools, are overrepresented in the sample and the averages. Therefore, the statistics are not representative of the average district in the Netherlands but representative of the district in which the average school is located.
5. Methodology
The effect of the size of classes on academic performance is endogenously determined; better schools attract more students so that their class size increases. However, in smaller classes there
is more personal attention from the teacher so that the individual progress might speed up. On the other hand, bigger classrooms have more opportunities for students to help each other, and therefore perhaps increase the performance because they teach and learn from each other. Therefore, the effect of class size cannot easily be tested through a regular OLS estimation. There is a need for an exogenous instrument to identify the effect of bigger classes.
Using the established Dutch institutions, this thesis attempts to find an exogenous instrument that can identify the effect of class sizes on academic performance. The Dutch system of free school choice detaches the connection of housing prices from school performance. However, there is still a connection between housing price and class size. As housing prices rise, it becomes more difficult for teachers to live in that area. As previously mentioned, the Dutch use of collective labour agreements makes it difficult for schools to offer teachers in more expensive districts a higher salary. This makes that areas with high housing prices, might not be interesting for teachers to live. The low availability of teachers, forces schools to create bigger classes to accommodate all the students so that class sizes increase. This gives rise to the instrument of housing prices, estimated by WOZ values in the Netherlands, in order to investigate the effect of class size on schooling.
The model that is estimated by the two stage least squares (2SLS) is;
𝐶𝐼𝑇𝑂𝑖𝑡 = 𝛼0+ 𝛼1 𝐶𝑆𝑖𝑡+ 𝐴𝑖𝑡 𝛼2+ 𝐼𝐷𝑖+ 𝜀𝑖𝑡 ( 1a )
𝐶𝑆𝑖𝑡 = 𝛽0+ 𝛽1 𝑊𝑂𝑍𝑖𝑡+ 𝐵𝑖𝑡 𝛽2+ 𝐼𝐷𝑖 + 𝑊𝑋𝑡𝜃 + 𝜐𝑖𝑡 ( 1b )
where, 𝐶𝐼𝑇𝑂𝑖𝑡 denotes the average CITO score of school 𝑖 in year 𝑡. 𝐶𝑆𝑖𝑡 denotes the observed class size. 𝐴𝑖𝑡 is a vector of controls that might affect the average CITO score on the school level (the share of female teachers, the share of teachers that have a fixed contract, the share of exempted students) and some district characteristics like population density, percentage of immigrant born people and their origin and the percentage of people under the age of 14. 𝐵𝑖𝑡 is a vector of control that might influence the optimal class size chosen by schools like the number of available managers and support staff. 𝜀𝑖𝑡 and 𝜐𝑖𝑡 are the error terms assumed to be uncorrelated with the estimators. 𝑊𝑂𝑍𝑖𝑡 is the independent variable of interest and is the average WOZ-value of the district in which the school is located in year 𝑡. Both equations include fixed effects at the school level, 𝐼𝐷𝑖, to include for unobserved characteristics.
rise in a particular street or neighbourhood. Teachers have a wide variety of traveling opportunities ranging from bikes and cars to busses and metros. Therefore, we would expect that if average house prices in neighbouring districts rise, the teacher availability is also affected for other schools in close proximity. Similarly, if prices next door decrease, the pool of teachers is likely to increase because they can now afford to live there and commute to other nearby schools. In order to control for these effects, it is necessary to include control variables for all used variables based on the values of other schools. This list of variables can be created through multiplying the values of the neighbouring schools with a corresponding spatial weight matrix (Elhorst J. P., 2013).
The spatial weight matrix is based on the fact that there is competition between schools for teachers and schools that are further away pose a smaller threat then schools that are nearby. For this reason, a distance decay spatial weight matrix is used or, more specifically, an inverse distance matrix is used: 𝑊 = 1𝛿, where 𝛿 is the distance between two schools. This matrix is then row-normalised by convention (Elhorst J. P., 2013). The impact of the matrix is that schools close by have a high impact on the availability and thus receive bigger weights, the row normalisation further ensures that schools located further away receive values approaching zero. We would expect that in extremely expensive districts, if prices rise, this no longer has an effect on teacher availability; since it was already impossible for teachers to live there, the rise in housing prices makes it no more impossible. However, if the consequence is that prices in neighbouring districts thus rise, it does affect the availability of teachers. This effect, the spatial lag effect, is a separate effect from the increase in housing prices in the district and needs to be identified as such. Consequently, the inclusion of the spatial lags is a crucial determinant in the availability of teachers for a particular school.
Since, it is only expected that the explanatory variables have an effect on the teacher availability and that the performance of schools does not affect each other, nor do the error terms of schools depend on each other, the chosen spatial model is an SLX-model (Elhorst & Halleck Vega, 2017, English translation).
5.1. Non-linear models
housing prices on the availability of teachers thus is highest in the middle segment districts. In order to accommodate for this effect, the WOZ-value is squared to estimate class size.
𝐶𝐼𝑇𝑂𝑖𝑡 = 𝛼0+ 𝛼1 𝐶𝑆𝑖𝑡+ 𝐴𝑖𝑡 𝛼2+ 𝐼𝐷𝑖+ 𝜀𝑖𝑡 ( 2a )
𝐶𝑆𝑖𝑡 = 𝛽0+ 𝛽1 𝑊𝑂𝑍𝑖𝑡+ 𝛽2 𝑊𝑂𝑍𝑖𝑡2+ 𝐵𝑖𝑡 𝛽3+ 𝐼𝐷𝑖 + 𝑊𝑋𝑡𝜃 + 𝜐𝑖𝑡 ( 2b )
where, all variables are as previously defined and 𝑊𝑂𝑍𝑖𝑡2 is the square of 𝑊𝑂𝑍𝑖𝑡.
Lastly, in smaller classes, peers are less likely to help each other so that the smallest class sizes might also be underperforming. An average classes would be performing best. To test for this effect, not only class size is instrumented by the WOZ value, but class size and its square are instrumented by the WOZ value and its square:
𝐶𝐼𝑇𝑂𝑖𝑡 = 𝛼0+ 𝛼1 𝐶𝑆𝑖𝑡+ 𝛼2𝐶𝑆𝑖𝑡2+ 𝐴𝑖𝑡 𝛼3+ 𝑊𝑋𝑡𝜃 + 𝐼𝐷𝑖 + 𝜀𝑖𝑡 ( 3 ) where, all variables are as previously defined and 𝐶𝑆𝑖𝑡2 is the square of 𝐶𝑆𝑖𝑡. They are jointly instrumented by the house price and the square of the house price together with the spatial weighted house price and its square.
5.2. Instrument validity
The Hausmann test, comparing the full linear models, showed a Chi-square statistic of 5.78. The results indicate that we can be 95% confident that the fixed effects model is better suited to estimate the effect of class size on academic performance than the simple OLS model. This indicates that there is an endogeneity problem. Therefore, I use an instrument for class sizes, the housing prices measured by WOZ-value. In the previous sections I have made an argument for why housing prices are an exogenous instrument for class size. There is no connection between quality of schools and housing prices because parents have a free school choice and are not limited to nearby schools, this means that the quality of schools does not affect the housing decision and subsequent price. Furthermore, as all parents can choose schooling, even if schools in richer districts were better, it would be possible for parents from other districts to bring their children there and profit from this better education. However, since schools in richer districts do not have more funding available, there is no reason to believe that they are better than schools in other districts.
and as their job is not location specific, they are very likely to find a job in a district closer to home. If many teachers do this, there might be a shortage of teachers in the more expensive areas so that class sizes are bigger simply because there are not enough teachers available. This means that housing prices influence the class sizes but do not affect academic performance in any other way. The relevance of the instrument can be tested for. The effect of housing prices on class size should be big enough for the prediction to matter. The first stage estimations provide the opportunity for this. In all models, the F-statistic in the first stage estimation is higher than, showing that jointly the instruments are a relevant predictor.
5.3. Robustness tests
However, there are several factors that might affect the availability of houses for teachers in a specific way. To test if any of these factors affect class sizes, and ultimately student performance, several robustness tests are done.
First, it might be that it is only after a certain threshold is reached that an effect can be observed. With low housing prices, it is possible for teachers to buy a house and an increase in price does not affect their ability to find a suitable home. It is only when housing prices are already at the limit of affordability for teachers that an increase in prices makes it impossible to find suitable housing meeting their needs. To test if such a threshold exists and if this has any effect, a dummy variable is included. The dummy is equal to one if a school is located in a district where the housing price is higher than the threshold and 0 otherwise. The econometric specification then looks like this:
𝐶𝐼𝑇𝑂𝑖𝑡 = 𝛼0+ 𝛼1 𝐶𝑆𝑖𝑡+ 𝐴𝑖𝑡 𝛼2+ 𝐼𝐷𝑖+ 𝜀𝑖𝑡 ( 4a )
𝐶𝑆𝑖𝑡 = 𝛽0+ 𝛽1 𝑊𝑂𝑍𝑖𝑡+ 𝐴𝑖𝑡 𝛽2+ 𝛽2 300,000𝑖𝑡 + 𝐼𝐷𝑖 + 𝑊𝑋𝑡𝜃 + 𝜐𝑖𝑡 ( 4b ) where 300,000 is the threshold dummy which can only take on the value of 1 and 0 and all other variables remain the same. The amount of the threshold is specified at €300,000 because that is approximately the mortgage that a couple consisting of two starting teachers can afford4. This maximum mortgage limits their opportunities of buying a house; houses cheaper than €300,000
4 Their salary is slightly lower than €35,000 individually; €70,000 as a couple, multiplied by 4.5 to get the
are affordable, houses more expensive are difficult to buy. Through the addition of this dummy, it is tested if there might be a structural break in the data.
Also, older teachers most likely have bought a home years ago and recent developments in the price of housing are unlikely to affect their location decision. The effect is expected to be most prominent among younger teachers who are currently looking to find housing and are unable to find suitable housing. Therefore, a negative relation between the WOZ-value and the availability of younger teachers could signal a potential future problem. The model to test if the problem is more prominent among younger teachers looks like:
𝐶𝐼𝑇𝑂𝑖𝑡 = 𝛼0+ 𝛼1 𝑦𝑜𝑢𝑛𝑔𝑡𝑒𝑎𝑐ℎ𝑒𝑟𝑠𝑖𝑡+ 𝐴𝑖𝑡 𝛼2+ 𝐼𝐷𝑖 + 𝜀𝑖𝑡 ( 5a ) 𝑦𝑜𝑢𝑛𝑔𝑡𝑒𝑎𝑐ℎ𝑒𝑟𝑠𝑖𝑡 = 𝛽0+ 𝛽1 𝑊𝑂𝑍𝑖𝑡+ 𝐴𝑖𝑡 𝛽2+ 𝛽2 300,000𝑖𝑡+ 𝐼𝐷𝑖+ 𝑊𝑋𝑡𝜃 + 𝜐𝑖𝑡 ( 5b ) where 𝑦𝑜𝑢𝑛𝑔𝑡𝑒𝑎𝑐ℎ𝑒𝑟𝑠𝑖𝑡 is the share of younger teachers, below the age of 35, as percentage of the total staff at a school. This test is a control test to see if the effects observed can be attributed to the rising housing prices.
Lastly, the effects are most pronounced in the region where the housing prices are highest. In the Netherlands, the Randstad is the densest region and the heart of the Dutch economy. Especially in Amsterdam, the housing prices have been booming (Lennartz, 2018). Therefore, the effect is most likely stronger in the Randstad. Consequently, I test if there is a different effect when the sample is restricted to schools located in the Randstad.
6. Results
Given the sometimes close proximity of schools, some schools are even located in the same building, it is expected that schools experience similar effects so that the spatial lag has the same sign.
To find the effect of house prices on the academic performance, I start with the estimation of the reduced form of the model. The results of this estimation are provided inTable 3. In column (1), the results are shown for the simple regression model without additional controls but with clustered standard errors and in column (2) these results are shown when the data is structured as a panel. There is strong evidence that the house price has a positive effect on the academic achievement of children. Schools that are located in richer districts perform better. The addition of time fixed effects, shown in column (3), and the addition of the squared value of house price, in column (4), support this statement where there is a strong, significant positive effect of the value of housing on the academic performance. This means that there still is a lot of sorting of likeminded people. Richer people tend to live together, and poorer people are more likely to live together and the children of richer people are more likely to perform better in school than the children of poorer people. Through the addition of further controls on the district level (column (4) and (6)), the effect of the house price becomes smaller. These controls focus on density, origin of the inhabitants and the share of young of people. An F-test performed to show if the origin of inhabitants is a relevant determinant of their academic achievement is significant at the 1% level.
However, the addition of fixed effects on the school level alters the results, in column (7) and (8). The coefficient for house price goes down and the standard error shoots up resulting in an insignificant coefficients. When moving to a fixed effects model, the estimation incorporates the constant effects of schools, like orientation, but also the district characteristics that remain constant over time. Also, effects that do not stay constant over time, but do not change much, like the composition of inhabitants, are now incorporated. Consequently, the effects of income segregation are now incorporated in the model. The consequence is that a rising WOZ-value within a district, does not affect the academic performance significantly. Lastly, the introduction of the spatial weight matrix to account for the effects of schools in close proximity, increase the coefficient of house prices (see column (9)).
Table 2: estimates of the first stage class sizes via the linear model 1 2 3 4 5 6 7 8 Log WOZ-value 2.169*** (.103) 2.030*** (.188) 1.046*** (.310) 1.932*** (.134) .852*** (.143) 1.718*** (.616) 1.405** (.632) 1.252* (.644) Time FE X X X X X X X District controls X X X X School controls X X X X Fixed Effects X X X W * Log WOZ-value .037 (.0561) Wald Chi2 148 247 148 837 1365 R2 .0274 .0009 .0036 N 16159 16159 15891 16119 15852 16159 15852 15356
Table 3: Reduced form estimation of CITO-score via the linear model 1 2 3 4 5 6 7 8 9 Log WOZ-value 4,365*** (.155) 4,276*** (.143) 4,189*** (.142) 2.982*** (.164) 4.140*** (.140) 2.989*** (.162) 1.176 (.855) .584 (.891) .948 (.908) Time FE X X X X X X X District controls X X X X School controls X X X X Fixed Effects X X X W * Log WOZ-value -.086 (.079) R2 .0998 .0998 .1041 .1473 .1249 .1579 .0873 .0004 .0002 N 16203 16203 16203 15934 16119 15891 16203 15852 15356
Notes: in column (1) the data is not sorted in a panel, in all later columns, it is. Columns (2) – (6) are random effects models, columns (7) – (9) are fixed effects models. School controls are the students per managers and support staff, share of female teachers, share of teachers on a fixed contract and the share of exempted students. District controls are population density, percentage of immigrants (per origin) and the share of people under the age of 14. The reported
𝑅2is the 𝑅2of the overall regression. Standard errors are in parentheses. Asterisks indicate that the estimated are significantly different from zero at the ***
In column (1) of Table 4 the most basic model is tested where class size is instrumented with house price. The estimated coefficient of 2.02 is smaller than the estimated coefficient of OLS (4.28). Which means that the effect of the class size on the academic performance is smaller than the effect of house price in the reduced model; further evidence of sorting in districts. The addition of further controls in columns (2) through (5) show similar effects to the reduced model except for the effect of adding school fixed effects. The effect of adding school fixed effects is much more pronounced in lowering the coefficient than in the reduced model, see Table 3 column (5). This is partially because the students per managers and the students per support staff are a main contributor to the size of classes as funds can only be invested once. This effect is much stronger than in the reduced model where they only affected the complete academic performance.
However, similar to the reduced model, when adding the fixed effects, the coefficient of class size is positive but insignificant as can be seen in column (6). Therefore, within schools, the effect of an increase in the class size, affects the performance positively. The addition of the school and district controls in column (7) show that the effect of class size on academic performance is negative, though the result is not significant. This more complete model shows that the effect of class size is more likely to be negative than positive on the academic performance of children. The addition of the spatial lag does not alter the coefficients.
6.1. Non-linear model
In this section, I evaluate the effect of an increase of house prices on class sizes under the assumption that class size does not linearly depend on house prices. The structure of this section follows the same order as the previous section; first the effect of house price on class size, then the reduced form and lastly, the effect of the estimated class sizes on the academic performance.
Table 4: Second stage estimation of CITO-score via the linear model 1 2 3 4 5 6 7 8 Class size 2.0166*** (.101) 1.9398*** (.214) 2.3169*** (.801) 1.1089*** (.073) .646*** (.069) .695 (.565) -.4327 (.322) -.441 (.348) Time FE X X X X X X X District controls X X X X School controls X X X X Fixed Effects X X X Spatial lag X R2 .0364 . 0380 .0313 .0391 .0792 .0274 .0316 .0318 N 16159 16159 15891 16119 15852 16159 15852 15356
insignificant. However, the spatial lags are significant at the 5% level meaning that there is a significant impact of neighbouring districts’ house prices on class sizes.
In the reduced model the addition of the square of house price alters the results, as shown in column (1) of Table 6. The coefficients for house price and its square are jointly significant at the 10% level. In column (2) it can be seen that the effect is positive when moving to a fixed effects model. Through the addition of the spatial weight matrix, the class size can be more precisely estimated since the joint significance of this model, column (3), is higher. The signs of housing price and its spatial lags are the same. This means that the effect of increasing housing prices of neighbouring districts has a similar effect of the teacher availability in the district compared to increases in the house prices near the school.
In the second stage regression, when including the square of house prices to test if there is a quadratic effect on class size in Table 7. The effect of class size is estimated as positive without further control. When including the controls however, I find that the coefficient moves becomes negative but is still insignificant as shown in column (2). However, in column (3), there is a significant negative effect of class size on the performance of children, but the square is positive.
The addition of the spatial weight matrix has little effect on the basic model results; column (4) show little effect. However, in column (5) the coefficients of the extended model are reported where the coefficients of class size and its square are significant individually at the 5% and 10% level respectively. However, jointly they are not significant meaning that the effects cancel each other out partially due to their opposite signs. The marginal impact of an increase in the WOZ-value, as estimated in column (3) of Table 7 is at the average -0.8125 which is like expected. For the smaller and bigger classes effects are -1.1726 and -0.4537. This means that the estimations are unexpected with smaller effects for already bigger classes. It is expected that a decrease in class size impacts children the most. However, when including the spatial weight matrix, the marginal effect at the average is -0.7998 which is similar to the effect without spatial lags. The effect of neighbouring districts thus might be smaller than initially expected.
5 An average class is 18.51, see Table 1. The marginal effect at the average is: 𝛽
𝑐𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒+ 𝛽𝑐𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒2∗
𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑐𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒
6 A small class is defined as being 2 standard deviations smaller than the average, see Table 1. The marginal
effect at the small class size is: 𝛽𝑐𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒+ 𝛽𝑐𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒2∗ 𝑠𝑚𝑎𝑙𝑙 𝑐𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒
7 A big class is defined as being 2 standard deviations bigger than the average, see Table 1. The marginal effect
Table 5; First stage regression in the non-linear models
1 2 3A 3B 4 5A 5B
Dependent Variable Class size Class size Class size 𝐶𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒2 Class size Class size 𝐶𝑙𝑎𝑠𝑠 𝑠𝑖𝑧𝑒2
Log WOZ-value 52.724*** (7.164) 7.809 (19.16) 7.809 (19.16) 258.013 (787.484) 4.085 (19.309) 4.085 (19.309) 91.891 (792.821)
Log (WOZ − value)2 -2.060***
(.291) -.258 (.772) -.258 (.772) -7.678 (31.711) -.115 (.777) -.115 (.777) -1.219 (31.922) District controls X X X X X X School controls X X X X X X Fixed Effects X X X X X X W * Log WOZ-value .865* (.447) .865* (.447) 44.262** (18.344)
W * Log (WOZ − value)2 -.068*
(.036) -.068* (.036) -3.529** (1.498) Wald Chi2 247 R2 0.0011 0.0011 0.0002 .0042 .0042 .0018 N 16159 15852 15852 15852 15356 15356 15356
Table 6: Reduced form estimation of CITO-score in non-linear models 1 2 3 Log WOZ-value 11.341* (6.632) 44.8167* (27.011) 42.453 (27.192) Log (𝑊𝑂𝑍 − 𝑣𝑎𝑙𝑢𝑒)2 -.291 (.27) -1.782 (1.088) -1.672 (.1.095) District controls X X School controls X X Fixed Effects X X W * Log WOZ-value .0584 (.629)
W * Log (WOZ − value)2 -.0119
(.051)
R2 .1042 0.0000 .0125
N 16203 15852 15356
6.2. Robustness tests
When adding the control for the threshold, as the effect might only be observable in districts with higher housing prices, the results do not change. In Table 8, the results are reported when a threshold dummy is included for several of the previously specified models. Concluding, there is no evidence for a structural break at €300,000 housing and that the effect is much more gradually dependent on the price of housing.
The next control test is by investigating the results of the schools that are located in the Randstad. All districts in the provinces of Noord-Holland, Zuid-Holland, Flevoland and Utrecht are included in this definition, therefore taking the definition of the Randstad broader than normally defined to include for travel opportunities. In full model, the coefficients in the cities that are only located in these four provinces, are positive as opposed to the negative coefficients for the whole country. This effect is shown in
Table 9, when compared to column (2) and (4) of Table 7. The coefficients of class size have changed sign but remain insignificant. The effect of house prices on class size is thus widespread and not limited to the Randstad only.
Table 7; second stage estimation of CITO score in non-linear models 1 2 3 4 5 Class size 1.578*** (.133) -.3932 (.319) -1.6639* (.861) -.3077 (.311) -1.506** (.747) Class size2 .046 (.029) .0386* (.022) District controls X X X X School controls X X X X Fixed Effects X X X X Spatial lag X X R2 .0382 .0306 .0041 .0277 .0137 N 15852 15852 15852 15356 15356
Another robustness test is only measuring the share of younger teachers. Since the effect is most likely only measurable for younger teachers, as older teachers are less likely to need to make housing decisions, the same tests have been performed with house price instrumenting for younger teachers to see if there are potential future problems policy should be made for now. If there are already effects on the availability of younger teachers, the problem will grow over time as the share of the current teachers declines. The results shown in Table 10 report on the findings for the effect house prices has on academic performance through the availability of younger teachers. When looking at a model including all control variables and fixed effects, as shown in column (1), it can be seen that there is a negative effect of young teachers on the academic performance based via house price. However, this result is highly insignificant. When assuming the squared effect of house price. Column (2), the effect is more pronounced and less insignificant, though not significant at the 10% level. However, when adding the spatial weight matrix, column (33), a higher R-squared is obtained and the effect is smaller.
teachers is insignificant. Therefore, the results when instrumenting with the share of young teachers, do not provide strong reasons to believe the results are fundamentally different.
Table 8; Control test on the CITO-score with an added dummy for Threshold
1 2 3 Class size -.354 (.315) -1.506** (.747) Class size2 .039* (.022) Log WOZ-value 39.475 (27.865) X X
Log (WOZ − value)2 -1.559
(1.125) X X Spatial lag X Threshold -.342 (.438) X X R2 .0000 .0314 .0137 N 15852 15852 15356
Table 9: Robustness test by limiting the sample to schools located in the Randstad
1 2 Class size .2941 (.694) .6508 (.723) Spatial Lag X R2 .0187 .0133 N 6680 6582
Notes: Columns (1) is the reduced form whereas column (2) and (3) are IV estimates. All models include school controls (the students per managers and support staff, share of female teachers, share of teachers on a fixed contract and the share of exempted students), district controls ( population density, percentage of immigrants (per origin) and the share of people under the age of 14) and time fixed effects. The reported 𝑅2is the 𝑅2of the overall regression. Standard errors are in parentheses. The reported 𝑅2is the 𝑅2of the overall regression. Asterisks indicate that the estimated are significantly different from zero at the *** 1% level, ** 5% level, and * 10% level
Notes: All columns are fixed effects models and are non-linear. All models include school controls (the students per managers and support staff, share of female teachers, share of teachers on a fixed contract and the share of exempted students), district controls ( population density, percentage of immigrants (per origin) and the share of people under the age of 14) and time fixed effects. The reported 𝑅2is the
𝑅2of the overall regression. Standard errors are in parentheses. The reported 𝑅2is the 𝑅2of the overall
Table 10; Robustness test if the instrument is replaced with the share of teachers below the age of 35
1 2 3 4 5
Teachers younger than 35 -5.514 (31.169) -21.672 (16.237) 10.397 (39.174) -3.962 23.539 (30.249)
Teachers younger than 352 -60.781
(64.335)
-43.515 (44.794)
Log (WOZ − value)2 X X X X
Spatial Lag X X
R2 .0002 .0005 .0005 0.0016 .0001
N 15852 15852 15851 15356 15356
6.3. Interpretation of results
All in all, the results do not point towards a conclusive result of the effect of class size on the academic performance of children. This is unlike what the main literature suggest, (Fredriksson, Öckert, & Oosterbeek, 2013; Krueger & Whitmore, 2001; Angrist & Lavy, 1999) where a positive effect of smaller classes can be found. Some careful analysis of this problem is thus necessary to identify where the differences within this analysis come from and how they relate to the findings in the literature. Five possible reasons can be identified as to why the results are not unambiguous; small time frame, big districts, property size, a non-representative independent test, and a weak instrument.
dataset uses four consecutive years on the school level. Their number of observations is thus higher reducing their standard deviation.
The second issue that might influence the effectiveness of this study are the unequal sizes of districts. The biggest district with a school location, in terms of population has over 108,000 inhabitants and the smallest, with at least one school, only has 30. For this last one the influx of students is most likely to come from a different, neighbouring district. With a mean of over 13,000 and a standard deviation of over 15,000, see Table 1, the differences across the country are big. Even though, the fixed effect control for the size of districts, the effects of these differences remains on the housing prices. On the one hand, free school choice is a much less relevant thing for parents in rural areas because distances too other school can be too big to provide a reasonable option for them. A consequence of this is that in the rural areas, there might be a stronger endogeneity issue between school quality and housing prices. On the other hand, in the biggest districts, there might be profound within district differences in terms of density, WOZ-value, origin of inhabitants and youngsters. This gives rise to measurement issues. This also shows why the addition of the spatial weight matrix has little effect on the scores since several schools, the ones with the highest inverse distance value, are in the same district and thus observe the same characteristics. The solution to this problem would be to gather the data on the more specific neighbourhood (“buurt”) level. The intention of this thesis was to do it on that most specific level, however, due to the statistical areas of 2014 and 2015 in Amsterdam this was impossible. Excluding Amsterdam, however, was not an option because houses are most scarce there.
Another alternative to why the effect of WOZ-values might not affect the schools so much is that the size of the property is not incorporated. If the average size of a property is higher, the WOZ-value goes up et ceteris paribus. However, the average size is not included in the current measure of WOZ-value where it is a key decision in the attractiveness of a house. If houses are structurally smaller in a district, this might affect the relationship between house prices and class sizes.
located who do not take the CITO test, see Appendix B. Therefore, the effect of WOZ on class size might be underestimated in the sample and thus bias the estimate for the effect of class size on the academic performance of students. In Table 11, the first stage estimation of the full model is redone for all the schools, including the ones without a CITO test. What can be seen is that the effect of house prices on the class sizes is more significant than in the first stage estimation with only the CITO schools.
Additionally, schools without CITO have more students from an immigrant background, both western as well as not western, further complicating their tasks and reducing the academic performance, which does not show in the current statistics. A difference in mentality between these schools can be most clearly observed in the difference of the exempted students. Schools that do not partake in the CITO test have, on average, a bigger share of their students exempted from taking the final test. This means that the academic performance is different between these schools, most likely better in the schools that do not partake in the CITO test because they have more poor performing students exempted; this change in attitude might show that there is more personal time for students so that the individual needs can be better observed. Consequently, it is expected that these classes, due to the additional attention, perform better. As these schools are omitted that could suggest a bias.
Table 11; the effect of house price on class size for all schools, including the schools not partaking in the CITO test
1 2 3 4 Log WOZ-value 1,317** (.547) 12.916 (16.99) 1.215** (.553) 9.639 (17.06)
Log (WOZ − value)2 -.469
(.686) -.341 (.689) W * Log WOZ-value -.012 (.055) .809** (.396)
W * Log (WOZ − value)2 -.067**
(.032)
R2 0.0074 0.0085 0.0098 0.0112
N 21737 21737 21737 21737
The last main possible reason for the weak results is the instrument itself. House prices have steadily risen in the last decades except during the financial crisis. However, the problem of teacher availability has only become apparent recently. Therefore, the threshold despite having an insignificant effect has only been met in the recent years. The effect of the bigger classes due to a shortage of teachers might only appear over time. It can be that the instrument has a weaker statistical link than it has a theoretical one but it might strengthen over time.
Lastly, there are some small additional reasons for the imprecise results, the student teacher ratio is not the same as the actual class sizes experienced by students which can result in a slight measurement error. Or if an area is classified as shrinking area, an area where population decreases, there might be additional funding available for schools so that schools can stay open where they would otherwise have to close because of low student numbers. Combination classes might also affect the academic performance, in these areas schools are too small to offer 8 different classes so they combine students which has an adverse effect on student performance.
7. Conclusion
The effect of class size on academic achievement and later life performance is of great importance to policy makers. Not only is education costly, it is also an important determinant of an individual’s later life success. Knowledge about the effect of smaller classes is thus crucial in optimal policy design. This thesis attempts to contribute to that knowledge.
One of the main problems with estimating the effect of class size on the academic performance is the reverse causality; better schools attract more students resulting in bigger classes. Furthermore, many studies, in the US, are prone to sorting issues due to the existence of school districts. In order to circumvent these problems, this thesis makes use of the Dutch institutional setting to find an exogenously determined instrument. The Dutch constitution guarantees free school choice and equal funding for all schools from the national budget. Furthermore, wages of teachers are determined by a central labour agreement so that teachers in expensive districts do not earn more. Therefore, higher housing prices create a teacher scarcity and this, consequently, causes bigger classes.
effect of bigger classes on academic performance. This is potentially due to data selection; the observations dropped have an effect on the class size. Other potential sources are the small-time frame, the size districts, the effect of property sizes, and a weak instrument.
There is a lot of potential research that can still be done to better estimate the effect of class sizes on the academic performance of children. In the Netherlands, a law has been passed to increase the funding of schools in places where teachers are scarce (Ministerie van Onderwijs, Cultuur en Wetenschap, 2018a). This gives the potential to do a study with a regression discontinuity design for these areas. Other potential studies include a longer time frame or more detailed data on the neighbourhood level instead of the district level. Lastly, the instrument of housing prices could be replaced with an availability-of-housing score, or with the actual renting and housing prices in an area instead of the proxy with the WOZ-value used in this thesis. All these options might further reduce the noise present in this dataset due to which there are no significant results.
All in all, the results do not point towards a conclusive result of the effect of class size on the academic performance of children. Unlike what the main literature suggest, (Fredriksson, Öckert, & Oosterbeek, 2013; Krueger & Whitmore, 2001; Angrist & Lavy, 1999) where a positive effect of smaller classes is found. After careful analysis of the potential problems causing the lack of results, there is more clarity on the endless potential for further research.
References
Angrist, J. D., & Lavy, V. (1999). Using Maimonides' Rule to Estimate the Effect of Class Size on Scholatic Achievement. The Quarterly Journal of Economics, 114(2), 533-575.
Blau, D. M., & Haurin, D. R. (2017). The Effect of the Price of Housing on Child and Young Adult Achievement. Journal of Real Estate Research, 39(3), 289-318.
CBS. (2016, 5 9). Buurt, wijk en gemeente 2015 voor postcode huisnummer [datafile]. Retrieved from https://www.cbs.nl/nl-nl/maatwerk/2017/38/buurt-wijk-en-gemeente-2015-voor-postcode-huisnummer
CBS. (2019, 3 15). Kerncijfers Wijken en Buurten 2004-2018 [datafile]. Retrieved from
https://www.cbs.nl/nl-nl/dossier/nederland-regionaal/wijk-en-buurtstatistieken/kerncijfers-wijken-en-buurten-2004-2018
Chetty, R., Friedman, J. N., Hilger, N., Saez, E., Whitmore Schanzenbach, D., & Yagan, D. (2011). How Does your Kindergarten Classroom Affect your Earnings? Evidence from Project STAR.
DUO. (2018, April 15). 05. Gemiddelde eindscores bo en sbo [data file]. Retrieved from https://duo.nl/open_onderwijsdata/databestanden/po/leerlingen-po/bo-sbo/bo-sbo-eindscores.jsp
DUO. (2018a, October 1). 06. Historisch overzicht aantal leerlingen per schoolvestiging [datafile]. Retrieved from https://duo.nl/open_onderwijsdata/databestanden/po/leerlingen-po/po-totaal/leerlingen-po-6.jsp
DUO. (2018b, October 1). 02. Onderwijspersoneel in het primair onderwijs in aantal fte [datafile]. Retrieved from https://duo.nl/open_onderwijsdata/databestanden/po/onderwijspersoneel/po-personeel2.jsp
DUO. (2019, January 1). 01. Adressen van hoofdvestigingen van scholen in het basisonderwijs [datafile]. Retrieved from
https://duo.nl/open_onderwijsdata/databestanden/po/adressen/adressen-po-1.jsp
Elhorst, J. P. (2013). Spatial Panel Models. In M. M. Fisher, & P. Nijkamp, Handbook of Regional
Science (pp. 1637-1652). Berlin, Heidelberg: Springer.
Elhorst, J. P., & Halleck Vega, S. (2017, English translation). The SLX-model: Extensions and the sensitivity of spatial spillovers. W. Papeles de Economía Española, 152, 34-50.
Finn, J., & Achilles, C. (1999). Tennessee's Class Size Study: Findings, Implications, Misconceptions.
Educational Evaluation and Policy Analysis, 97-109.
Fredriksson, P., Öckert, B., & Oosterbeek, H. (2013). Long-term Effects of Class Size. The Quarterly
Journal of Economics, 128(1), 249-285.
Gemeente Amsterdam. (z.d.). Op welke scholen heeft mijn kind voorrang? Retrieved 6 10, 2019, from Schoolwijzer Amsterdam:
https://schoolwijzer.amsterdam.nl/nl/po/vragen-en-opmerkingen/vraag/op-welke-scholen-heeft-mijn-kind-voorrang
Krueger, A. B., & Whitmore, D. M. (2001). The Effect of Attending a Small Class in the Early Grades on College-test Taking and Middle School Test Results: Evidence from Project STAR. The
Economic Journal, 111(468), 1-28.
Lennartz, C. (2018, July 3). Huizenprijzen in mondiale steden beginnen te dalen – volgt nu ook
Amsterdam? Retrieved 6 10, 2019, from RaboResearch - Economisch Onderzoek:
https://economie.rabobank.com/publicaties/2018/juli/huizenprijzen-mondiale-steden-dalen-volgt-amsterdam-ook/
Loubert, L. (2005). Housing Markets and School Financing. Journal of Education Finance, 30(4), 412-429.
Ministerie van Onderwijs, Cultuur en Wetenschap . (2015). Regeling toetsreglement van de Centrale
Eindtoets po schooljaar 2015-2016. Ministerie van Binnenlandse Zaken en
Koninkrijksrelaties .
Ministerie van Onderwijs, Cultuur en Wetenschap . (2016). Regeling toetsreglement van de Centrale
Eindtoets po schooljaar 2016–2017. Ministerie van Binnenlandse Zaken en
Koninkrijksrelaties .
Ministerie van Onderwijs, Cultuur en Wetenschap . (2017). Regeling toetsreglement van de Centrale
Eindtoets po schooljaar 2017–2018. Ministerie van Binnenlandse Zaken en
Ministerie van Onderwijs, Cultuur en Wetenschap. (2018). Regeling toetsreglement van de Centrale
Eindtoets po schooljaar 2018-2019. Ministerie van Binnenlandse Zaken en
Koninkrijksrelaties.
Ministerie van Onderwijs, Cultuur en Wetenschap. (2018a). Subsidieregeling regionale aanpak
lerarentekort. Ministerie van Binnenlandse Zaken en Koninkrijksrelaties.
Renkema, E., Mulder, A., & Barnard, M. (2016). Merging Identities: Experiments in Dutch Primary Education. Religious Education, 111(1), 75-94.
Schwarts, A. E., Voicu, I., & Mertens Horn, K. (2014). Do choice schools break the link between public schools and property values? Evidence from house prices in New York City. Regional
Appendix B; CITO vs NON-CITO schools
CITO test No CITO test T-statistic
Observations 17011 7691
Students 234.9 (138.4) 207.8 (135.4) 14.4
Managers 1.17 (.75) 1.04 (.79) 12.2
Teachers 12.7 (7.6) 11.3 (7.5) 13.6
Support Staff 1.5 (2.7) 1.4 (2.8) 4.4
Students final year 29.9 (17.7) 26 (17.2) 16.3
% Students Exempted 2.24 (7.6) 3.03 (14.2) -4.4 Number of inhabitants 14100 (15790) 12534 (14105) 7.5 Population density 2705 (3578) 2371 (2914) 7.7 WOZ 219 (69) 210 (58) 9.8 Class size (students/teachers) 18.7 (3.8) 18.1 (4.2) 9.8
% inhabitant younger than 14 17.1 (3.3) 16.9 (3.2) 3.6 % western immigrants 8.7 (4.8) 8.2 (4.5) 7 % non-western immigrants 10.4 (13) 9.1 (10.7) 8.1 % immigrants from Morocco 1.9 (3.7) 1.4 (2.9) 9.6
% immigrants from the Netherlands Antilles and Aruba
0.65 (1.1) 0.64 (1.1) 0.7
% immigrants from Surinam
1.7 (3.7) 1.4 (2.8) 7.8
% immigrants from Turkey 1.97 (3.8) 1.66 (3.3) 6.4
% owner-occupied homes 60.6 (15) 61.9 (14) -6
% high income household 21.65 (8.3) 20.8 (8.5) 4.3
% low income household 36.6 (10.8) 36.9 (10.6) -1.4
% teachers younger than 35 31.9 (17.6) 31.8 (17.8) 0.4
Students per managers 209 (142) 191 (154) 8.6
Students per support staff 277 (557) 255 (908) 1.95
% teachers with a fixed contract
93.37 (9.3) 93.33 (11) 6.5
