The Impact of the Quality of Primary
Schools on Housing Prices: An Analysis of
Amsterdam
Written by Orjan Ohlsen, July 2016
Student number: 6081061
Supervised by prof. dr. M.K. (Marc) Francke
University of Amsterdam, Amsterdam Business School
MSc Business Economics, Finance track
Statement of Originality
This document is written by Student Orjan Ohlsen 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.
ABSTRACT
For this study I utilized prices of house transactions between 2012 and 2014 in Amsterdam to measure the impact of primary school quality on house prices. The average proficiency test score and an added value score are used to measure primary school quality. The Cito-test is the nationwide standard proficiency test for primary schools. The RTL-score is used as a measure for added value. The capitalization of school quality in housing prices has never been studied in the Netherlands. Based on the data of 197 primary schools and 21,628 house transactions, this study finds support that parents value the average proficiency test scores per school. Little support was found that added value of a primary school is capitalized in house prices. A 1 percent increase in a school’s average Cito-test score raises house prices by 0.364 to 5.762 percent, depending on the model specification.
TABLE OF CONTENTS
1. INTRODUCTION 5
1.1 REASON OF THIS STUDY 5
1.2 STRUCTURE 6
2. LITERATURE 7
2.1 METHODS USED IN PRIOR RESEARCH 7
2.2 MEASUREMENTS USED IN PRIOR RESEARCH 10
3. DUTCH PRIMARY SCHOOL EDUCATION SYSTEM 12
3.1 PRIMARY SCHOOLS 12
3.2 URBAN ADMISSION POLICY 13
4. EMPIRICAL APPROACH 14
4.1 THE MODEL 14
4.2 CREATING WEIGHTED AVERAGE OF SCHOOL QUALITY 16
4.3 CONTROL VARIABLES 17
5. DATA DESCRIPTION 19
5.1 DATABASE 19
5.2 DESCRIPTION OF THE SOLD HOUSES 19
5.3 DATA ON SCHOOL QUALITY 20
5.4 DESCRIPTION FROM MATCHING HOUSES TO PRIMARY SCHOOLS 22
5.5 DATA FROM PRIMARY SCHOOLS IN GENERAL 24
6. RESULTS 25
6.1 THE EFFECT OF THE CLOSEST PRIMARY SCHOOL 25
6.2 THE EFFECT OF A WEIGHTED AVERAGE OF SCHOOL QUALITY 28 6.3 THE EFFECT OF THE BEST PERFORMING SCHOOL IN AN AREA 29
6.4 SUMMARY AND CONCLUSION 30
7. ROBUSTNESS CHECKS 32
7.1 RESULTS USING A 500 METER RADIUS 32
7.2 RESULTS USING A REGULAR AVERAGE 33
7.3 RESULTS USING HOUSES WITH THREE OR MORE ROOMS 33
8. CONCLUSION 35
8.1 EFFECT OF PROFICIENCY TEST SCORE ON HOUSE PRICES 35 8.2 EFFECT OF ADDED VALUE ON HOUSE PRICES 36
8.3 LIMITATIONS OF THE RESEARCH 37
8.4 SUGGESTIONS FOR FUTURE RESEARCH 39
9. REFERENCES 40
APPENDIX 42
1. INTRODUCTION
Over the last thirty to forty years, numerous studies were performed focusing on the relation between housing prices and the quality of schools. A wide range of international research, though mainly focused on US data, confirms a significant positive relation between housing prices and the performance of schools (Black and Machin, 2010).
1.1 CONTRIBUTION OF THIS STUDY
In the US, public primary schools belong to school districts. This means pupils can only apply for public primary schools in the district they live. This adds importance to the relationship between the quality of public primary schools and housing prices. In Europe it is less common to have school districts. That might be the reason why there is a lack of this type of research in Europe. Moreover, this type of research has never been done in the Netherlands. Since town structures and the system of primary schools of cities in the US and the Netherlands are incomparable, you cannot generalize the results found in prior research to cities in the Netherlands. For example, the distances in Amsterdam are smaller than large cities in the US. In the Netherlands pupils have to travel 700 meters to a primary school on average1, while pupils in the US travel 5.8 kilometers on average2, pupils in the UK3 travel 2.6 kilometers on average and pupils from Australia4 travel 4.2 kilometers on average. So there appears to be a gap in the current literature, which asks for research to be conducted on Amsterdam. By doing so, this study will be a unique study focusing on a more densely populated Dutch city. The general research findings from this study may be applicable across other more densely populated European cities with a similar primary school system.
Due to the implementation of a new urban admission policy for primary schools in 2015, it became of special interest to study the relation between housing prices and the quality of primary schools in Amsterdam. This urban admission policy applies only to Amsterdam. The policy gives attendance priority to 8 primary schools closest to a child’s address. Therefore, the question arises what the impact is of the quality of primary schools on housing prices in Amsterdam. Differently said: are parents willing to pay a higher price for houses that are surrounded by higher quality primary schools? To calculate the price of the quality of
1 Source: CBS, Central Bureau for Statistics.
2 Source: National Household Travel Survey. US Department of Transportation. 3
primary schools we use housing prices. Rosen (1974) popularized this method, called the hedonic pricing model, to measure the value of public goods by house prices. Local goods, like primary schools, are not traded on competitive markets. So the increase in housing prices due to an increase in the quantity or quality of a public good can be interpreted as the marginal value of this public good. Earlier research also used housing expenditures to measure the willingness to pay for local public goods, like primary schools (Sheppard, 1999). This study could be of particular interest for investors, homebuyers, real-estate managers or policy makers. Since the attendance policy has recently been implemented, the effect of the implementation of this policy might not yet be fully visible in housing prices. This means a difference-in-difference method, that shows the effect of a new policy, will not produce the desired output. In real estate, it is common that the number of sales changes first and afterwards the housing prices. This thesis does not focus on the effect of the policy; rather it focuses on the impact of the quality of primary schools on house prices. The expectation is to find a positive effect of school quality on housing prices in Amsterdam. Since the same effects were found in the US as well as Australia (Davidoff and Leigh, 2008), the UK (Gibbons and Machin, 2008) and France (Fack and Grenet, 2010).
1.2 STRUCTURE
The structure of this thesis is as follows. In Section 2 a theoretical outline is made of prior research in the field of school quality and its effect on housing prices. In Section 3 follows a short description of the Dutch primary school education system and the new urban admission policy. In Section 4 the methodology used in this study is described by stating the hedonic regression and the included variables. Section 5 describes and summarizes the data used for this study. In Section 6 the results of the methods undertaken are reviewed and analyzed. Then in Section 7 some robustness checks are done. Finally in Section 8, results are summarized, conclusions are drawn, links are made between the results and the theoretical outline and suggestions for future research are initiated.
2. LITERATURE
In this section the methodologies and findings of prior research will be discussed. There is an extensive collection of studies investigating the value people assign to public goods. Tiebout (1956) was one of the first to come up with a theory. He explained that consumers pick a community that optimally suits their preference pattern for local public goods, like schools, and amount of tax they have to pay. They will tend to move to this community. Oates (1969) found results consistent with Tiebout’s theory. According to the research of Oates, people on average are willing to pay more to live in a community with high-quality public services. Over the past forty years, substantial research has been conducted about the specific effect of the quality of primary schools on housing prices. Isolating this effect is not easy and prior research points out that there is no single way to approach this problem. Historically, there are a few methods used studying the effect of school quality on housing prices. They all utilize the hedonic pricing model approach. A hedonic pricing model values a public good using house prices. Three different methods will be discussed below.
2.1 METHODS USED IN PRIOR RESEARCH
The first method uses panel data. Downes and Zabel (2002) connected changes in average housing prices in Chicago to the change in quality of the closest school per address between 1987 and 1991. To prevent biased estimates, they controlled for all other unobserved determinants of house prices. They found a 1 percent increase in mean reading test scores led to 1 percent increase in housing prices. Clapp et al. (2004) also utilized panel data between 1994 and 2004 in Connecticut to measure the effect of change in school quality on housing prices. They found that a standard deviation increase in math score led to a 7.4 percent increase in house prices using a hedonic regression, while using town and tract fixed-effects models only led to a 1.4 percent and 1.3 percent increase in house prices respectively. The risk with this method is that changes in the quality of schools or house prices might be caused by unobserved changes in the neighborhood. Moreover, manually controlling for changes in neighborhood variables is difficult and time consuming.
The second method applies basic hedonic pricing regression. This method regresses the quality of primary schools on housing prices, controlling for all other house and neighborhood characteristics by including them in the regression. Brasington and Haurin
(2006) regressed house value on proficiency test scores and added value5 in several school districts. They controlled for neighborhood characteristics and house characteristics by including them in the regression. They found that an increase of one standard deviation in average proficiency test scores raised housing prices by 7.1 percent. Furthermore, Hayes and Taylor (1996) found a 1 percent increase in math scores increases house prices by 0.26 percent. Brasington (1999) finds that a 1 percent increase in the pass rate for a reading test in the fourth grade leads to a 1 percent increase in house prices. Finally, Crone (2006) discovered that 1 standard deviation increase in math and reading test scores increased house prices by 1.7 percent to 2.4 percent. While valuing the quality of schools, one has to consider variables related to the quality of schools as well as the price of houses. Better schools tend to be located in better neighborhoods. It is therefore complicated to estimate the causal effect of school quality on house prices.
The third methodological approach is called the boundary discontinuity model and derives from Black (1999). Black (1999) compared the prices of houses located at both sides of the boundaries of school attendance districts. These houses are located in the same neighborhood, maybe even in the same street, but since they are in different school districts, they cannot apply for the same schools. With this method you can isolate the effect of school quality on house prices and automatically control for most neighborhood characteristics. Black (1999) finds that parents are willing to pay 2.5 percent more for a house if the average test scores in the school district increase with 5 percent. On the other hand using a basic hedonic regression method led to a 4.9 percent increase in house value. This suggests that the results of the basic hedonic regression method, as used by Brasington and Haurin (2006), are biased upwards. This is explained because house and neighborhood characteristics are positively correlated with school quality and house prices. Hence many studies utilize this boundary discontinuity design to control for neighborhood effects including Fack and Genet (2010), Bayer et al. (2007), Davidoff and Leigh (2008) and Kane et al. (2006). The latter paper found differences in housing prices along the school boundaries in Mecklenburg County, North Carolina. The effect they found of school characteristics was only one-quarter of the effect implied by non-boundary studies. This might imply that non-non-boundary studies do not control for some variables that are taken into account by boundary discontinuity models. Bayer et al. (2007) used a model that embeds a boundary discontinuity design to control for the correlation
5 Added value is the additional knowledge a school imparts to children, besides their innate characteristics and knowledge acquired outside the school.
between school and neighborhood attributes. They also found a positive and significant relation between housing prices and the quality of schools.
Because Amsterdam is not divided into school attendance zones, there are no boundaries along which houses from the same neighborhood differ in school attendance possiblity. A boundary discontinuity model is therefore not applicable in this study. However, not all research employed or approved this design. Clapp et al. (2004) made critical comments since attendance districts change over time and because the boundary effect became weaker on the edge of urban areas. Hence, for this thesis, the second method will be undertaken by creating a basic hedonic pricing regression. To isolate the effect of the quality of primary schools on housing prices, control variables for house and neighborhood characteristics will be included in the regression. In order to decrease the upward bias caused by omitted variables described by Black (1999), sufficient relevant control variables have to be added to the regression. Another common problem in this field of research is reversed causality. Did the quality of schools influence the house prices or are higher quality schools located in wealthier neighborhoods? Black and Machin (2010) analyzed a wide variety of studies that value school quality through housing prices. Some of these studies incorporated instrumental variables in their method, though this research showed it was difficult to find credible instruments for quality of schooling. An explanation of the reversed causation problem is the local financing models for public schools. These models provide an incentive for higher income households to invest in primary schools to improve their quality. So higher income neighborhoods, where presumably house prices are higher as well, can influence the quality of schools (Nechyba, 2003). Epple and Romano (1996) describe that increasing public and private school funding or tuitions leads to higher quality education levels.
To cope with the problem of reversed causality, researchers took a partial-equilibrium perspective that assumes marginal house buyers are not going to have a significant effect on neighborhoods and school quality (Black and Machin, 2010). In the Netherlands, private funding covers only 4.5 percent of the total funding so it might be less relevant. Though to exclude reversed causation by private funding, a partial-equilibrium perspective is taken during this research. All in all, there is a lack in current research that allows modeling to overcome this endogeneity problem.
2.2 MEASUREMENTS USED IN PRIOR RESEARCH
This thesis focuses on the effects quality of schooling conveys on the pricing of adjacent houses. Prior research conducted in the US compared houses in different school districts with one another. In Amsterdam, it is possible to apply for every primary school and no limitations or restrictions apply depending on the proximity of residential address to location of school. Downes and Zabel (2002) matched houses in Chicago to the closest primary school. In Amsterdam, the effect of the closest school may be biased if there are a few other schools close to the house, which is highly plausible since distances in Amsterdam are most likely smaller than in Chicago. Hence, the houses will also be matched to primary schools by taking an average of all primary schools in an area around the house.
To measure quality of schools, prior research mainly used three measurements: per pupil expenditure, proficiency test scores and added value. The first measurement is mainly used in earlier research. It derives from Tiebout’s theory that people value communities based on their expenditures on public goods, like primary schools. Hanushek and Taylor (1990) showed that using this measurement for schooling quality shows biased results. Furthermore Hayes and Taylor (1996) used all three measurements but could not find an effect for per pupil expenditure. They could find an effect for school achievement. Also Downes and Zabel (2002) find that parents do not value school inputs, like per pupil expenditures, but they rather value school outputs, like test scores. Weimer and Wolkoff (2001) use ELA-scores, the English Language Arts exam, to measure school quality. They find a positive and significant effect on housing prices in Monroe County, New York. Furthermore, only 4.5 percent of the total funding of primary schools in the Netherlands is private. This means 95.5 percent of the funding derives from the government, based on the number and distribution of the pupils.6 This way the government tries to equalize the per pupil expenditure in the Netherlands. Added value is the additional knowledge a school imparts to children. Hayes and Taylor (1996) argued that the capitalization of school quality in housing prices derived as a result of the school’s marginal effect on outcomes. This is more than just a test score, but a measure of improvement due to the school. Hayes and Taylor (1996) use a value added model that decomposes average achievement into school district and the expected effect derived from parents. They found a 1 percent increase in the marginal effect of schools leads to a 0.26 percent increase in house prices. Meyer (1997) states that measuring school quality by test
scores leads to biased results because they are contaminated with non-school factors that contribute to the pupil’s performance. So in order to isolate the contribution of schools on pupil’s achievement, Meyer (1997) states that a value added measure should be used instead of pupils’ test scores. Brasington (1999) criticized the measure used by Hayes and Taylor (1996) because they did not measure the same students’ improvement over time. The value added created by Brasington (1999) is the test score in the twelfth grade minus the test score in the fourth grade. He took the average of a school district. However, Brasington (1999) did find a positive and significant effect for average proficiency test scores per school district, but not for the value added.
Downes and Zabel (2002) also used three measures for school quality. Their value added was calculated in the way Brasington (1999) did, by using test scores taken by the same cohort of pupils at different grade levels. They could only find significant support for test scores and not for pupil expenditures and added value. Brasington and Haurin (2006) used measures of per pupil expenditure, test scores and value added to measure school quality. They utilized the value added measure of Hayes and Taylor (1996), Brasington (1999) and Downes and Zabel (2002). Both expenditure and test scores were capitalized in the house prices. Test scores appeared to be the most consistently valued measure of school quality, while the value added measure did not give significant results.
Value added seems to be the most rational attribute to measure school quality, since it displays the ability of a school to improve children. Though when parents value primary schools they will look at information available. Test scores per primary school are public information, but value added generally is not. This could be an explanation why value added did not give significant results in some of the studies. There is an ongoing discussion and debate as to whether test scores or value added should be used to measure the quality of primary schools. In this thesis, we use both measurements and compare the results. A side note should be made that the value added used in this study is available to public as well.
3. DUTCH PRIMARY SCHOOL EDUCATION SYSTEM
This section will provide information about primary education in the Netherlands and the new urban admission policy in Amsterdam.
3.1 PRIMARY SCHOOLS
Dutch children can attend primary school at the age of 4. They are legally obliged to attend primary school at age 5. Primary school education lasts for 8 years. Furthermore the law of compulsory education obligates children to follow primary school and secondary school until the age of 18, or if they at least graduate in higher general secondary education. There are more than 7.000 primary schools in the Netherlands. Most of these schools are regular primary education (bao), though there are also schools focusing on special primary education (so) and continued special primary education (vso). These last two types of schools are developed for children with learning disabilities or children who need extra care. After 8 years of primary school pupils have to take a final exam to measure their level of education. The most common exam undertaken by primary schools is a proficiency test called the Cito-test (nowadays called “The Final Test of Primary Education”). The average results per primary school are public information. The final test in the eighth grade is often the only test Dutch pupils have to take. This means that for most pupils there is no test in between first and eighth grade that could be used to measure the improvement of Dutch pupils.
All the primary school can be split into two categories: public primary schools and particular primary schools. Public primary schools are founded and controlled by the government and are not based on religious or philosophical denominations. All children are able to attend these schools. 32 percent of all primary schools in the Netherlands are public. Particular primary education is not controlled by the government and is founded on religious ideas or an educational philosophy. These schools have the right to reject pupils if they do not endorse the religion or philosophy of the school. 68 percent of all primary schools in the Netherlands have particular education (30 percent Catholic, 30 percent Protestant and 8 percent Other). The average school size in the Netherlands is steady around 220 pupils. In Amsterdam there are more than 200 primary schools. Around 46 percent of them are public, 18 percent is Catholic and 17 percent is Protestant.
For pupils, it is costless to attend any primary school. The government does the financing of the primary schools. The money a school receives depends on the number and background of
the pupils. Pupils who need more attention, like non-Dutch pupils or pupils whose parents have a low educational background, are more costly. The government therefore spends more money on schools that have more pupils that need extra attention. Besides governmental funding, schools can be funded privately by people (mostly parents) or companies. A school could ask parents for a contribution for extracurricular activities, people can just donate money and people can sponsor the primary school. The latter means that the sponsor gets a favor in return. This type of funding is restricted by some regulations since this is something the government treats carefully. In 2014, the government did 95.5 percent of the funding, parents did 3.6 percent and companies did 0.9 percent of the funding. Pupils are not geographically restricted to apply for a primary school. Though distance and quality are important factors for parents to pick a primary school (Karsten et al., 2002).
3.2 URBAN ADMISSION POLICY
Distance and quality are two key factors of future importance following the implications surrounding the new urban admission policy for primary schools. The policy came to effect in the 2015/2016 school year for more than 200 primary schools in Amsterdam. This urban admission policy results in one central system where pupils born after July 1st 2011 can apply for primary schools. Parents register their children for the primary school they prefer most. Besides that they have to name at least five alternative primary schools, arranged from most preferred too least preferred. The central system then registers all the applications and automatically places all pupils in primary schools based on the stated preferences. If a primary school has enough space for all the applicants, the pupils will automatically be placed to this primary school. This happens in more than 75 percent of the cases. However, if there are more applicants for a primary school than they can place, they will have to select pupils. The selection of pupils is based on a priority system. Pupils will get priority to attend the eight primary schools closest to their home address, based on a walking distance. Furthermore, pupils get priority to attend a primary school in the following cases:
• If a sibling attends the same primary school when the pupils reaches the age of 4. • If the pupil went to a preschool that is part of the primary school.
• If the pupil attends an integrated center for children (IKC) and the primary school belongs to this center.
• If a parent of the pupil has a contract of service of indefinite duration at the primary school.
4. EMPIRICAL APPROACH
In this section, the models are explained as well as the creation of the weighted average of school quality. Furthermore, the control variables are briefly summarized and details explaining their expected effect are included.
4.1 THE MODEL
The hedonic pricing regression reflects the market value of house characteristics, neighborhood characteristics and the quality of primary schools in house prices. House characteristics, like lot size and type of house, are house specific. Neighborhood characteristics, like average income, are equal for an entire sub-district. Amsterdam is divided in 8 districts, which are subdivided into 97 sub-districts. There are two different regressions used in this study. The first regression contains fixed-effects for sub-district characteristics and the second regression controls for sub-district variables by including them in the regression. The first hedonic house price function used in this study takes the following form:
(1) 𝑙𝑛 𝐻!" = 𝛽!+ 𝛽!𝑋! + 𝛽!𝑌! + 𝑆 + 𝐷 × 𝑇 + 𝜀!
Where ln Hij is the natural logarithm of the price of house i in sub-districts j. Data on housing
prices have a long right-hand side tail and a logarithmic function gives a better fit to linear models (Conroy and Milosch, 2011). Xi contains house characteristics of house i and Yi is the
measure of school quality belonging to house i. The quality of schools is measured in two different ways. Either by a proficiency test score, in this research the average Cito-test score of school s near house i (Ci,s) or by the value added measure of school s near house i (Vi,s).
Value added is measured using an RTL-score. Information about the RTL-scrore and its derivation can be found in the Appendix. The primary schools are matched to houses in three ways as follows:
• Houses are matched to the closest primary school, as per Downes and Zabel (2002). • Houses are matched to all schools within a 1,000 meter radius from the house by a
weighted average, assigning higher weights to primary schools that are in closer proximity.
The second method is done to replicate a school district around a house and take the average of this “school district”. The latter is done under the assumption that parents only look at the best primary school and value this if it is close enough to their house. Distances in Amsterdam are small and there are more than 200 primary schools. The competition between primary schools is high. Therefore, instead of considering just the nearest primary school, considering all schools within a circle around a house seems appropriate.
Furthermore, sub-district fixed-effects are included in the regression by creating a dummy for all of the 97 sub-districts. This is denoted with an S in the regression. Lastly, time fixed-effects interacting with the districts in Amsterdam, denoted with 𝐷 × 𝑇, are included in the regression. Since the housing market fluctuates over time and location, it is important to control for both factors. In Figure 1, we can see the general development of the housing market in Amsterdam. The house market moves in an upward trend, though the house prices fluctuate around the upward trend. This figure shows the importance of including time fixed-effects. T contains 12 dummies: one for every quarter between 2012 and 2014 and D contains dummies for the 8 districts in Amsterdam. Hence, this interaction creates 96 dummies. The objective is to measure the effect of Yi on Hi. The hypothesis, as stated before based on prior
research, is to find a positive effect of index Yi on Hij.
The second hedonic house price function used in this study controls for sub-district variables by including them in the model. The regression takes the following form:
(2) 𝑙𝑛 𝐻!" = 𝛽!+ 𝛽!𝑋! + 𝛽!𝑌! + 𝛽!𝑍!+ 𝑇 + 𝜀!
Where ln Hij is the natural logarithm of the price of house i in sub-districts j. Xi contains
house characteristics of house i and Yi is the measure of school quality. Zj contains the
neighborhood characteristics of sub-district j. Lastly time fixed-effects, denoted with a T, are included to the regression. T contains quarterly dummies just like Equation (1). The objective is to measure the effect of Yi on Hi. The hypothesis aims to find a positive effect of index Yi
on Hij.
To cope with the problem of reversed causality, it is assumed that marginal house buyers are not going to have a significant effect on neighborhoods and school quality. Still, finding a causal effect using these models will not be possible. Therefore, the focus will be on the relation between quality of primary schools and housing prices.
4.2 CREATING WEIGHTED AVERAGE OF SCHOOL QUALITY
Amsterdam has over 200 primary schools. The Cartesian coordinates of these schools are collected and matched with the Cartesian coordinates of houses in Amsterdam. Both coordinates from houses and schools are used to calculate the distance between houses and primary schools. This way we can determine which primary school is the nearest to each house. Moreover, this distance is also used to calculate the weighted average of all schools within a 1,000 meter radius per house. The distance between a house and a primary school is calculated as follows:
(3) 𝐷!,! = (𝑥!− 𝑥!)!+ (𝑦
!− 𝑦!)!
Where Di,s equals the straight line distance between house i and school s within the meter
radius of the house. The x-coordinate and y-coordinate of school s are defined as xs and ys
respectively. The x-coordinate and y-coordinate of house i are defined as xi and yi
respectively.
Now we can calculate a weighted average of the quality of all the schools with a 1,000 meter radius of the house. The closer the school is to a house, the heavier the weight. The 1,000 meter radius is twice the average distance pupils in Amsterdam have to travel from their house to their primary schools7. Then a 500 meter radius is used to check for robustness, since the average distance of the closest primary school in the dataset is 285 meter (see Table 1A). The weights are defined like this:
(4) 𝑤!,!= ! !!,!
! !!,! !
Hence, the weighted average of the quality of all the schools measured by proficiency test score, the Cito-test score, takes the following form:
(5) 𝑌!,! = !𝑤!,!∗ 𝑐!,!
And the weighted average of the quality of all the schools measured by the value added, the RTL-score, takes the following form:
(6) 𝑌!,! = !𝑤!,!∗ 𝑣!,!
Equation (5) and (6) both sum all the schools within a 1,000 meter radius from the house.
4.3 CONTROL VARIABLES
In order to isolate the effect of school quality on housing prices, control variables for house characteristic and sub-district characteristics have to be added. The house characteristics and neighborhood characteristics are shown in Table 1A on the next page and Figure 2 and Figure 3 in the Appendix. It is expected that both living area and lot size have a positive effect on housing price. The distance to the city center is also included in the table. Amsterdam’s city center is rich in public goods and, like any other city center, the district with the most central location. Other parts in Amsterdam are easy accessible from the center and therefore I expect a negative sign of the coefficient. So the smaller the distance to the city center, the higher the price of the house. Furthermore, the proximity to a primary school is included. According to Conroy et al. (2016) living close to a primary school can have a negative effect on housing prices, because of so-called nuisance effects (noise, parking and traffic congestions). This negative effect declines with distance, suggesting this might be an important variable to control for. A dummy is made for houses within a 50 meter distance of a primary school and included in the regression. Based on the results of Conroy et al. (2016) the coefficient of this dummy is expected to have a negative sign.
The dummy variables for elevator, attic, swimming pool and indoor parking are also expected to have a positive effect, since these are indicators of luxury within the house. Lastly the indoor condition and the outdoor condition, both ranging from 1 to 9, are added up to create the variable “House condition”. Three categories are made: bad condition (scores form 2-13), average condition (scores from 14-15) and good condition (scores from 16-18). Figure 3 in the Appendix depicts the associated distribution corresponding to each category. The expectation is that better condition leads to higher house prices.
The neighborhood characteristics are all shown in the descriptive statistics in Table 1A. It is expected that the average income in a sub-district, the percentage of owner-occupied houses and high education are expected to have a positive effect on the house value. On the other hand, the percentage of unemployment, non-western immigrants and low education are expected to have a negative effect on housing prices. Furthermore, there might be a few interaction effects between variables. Parking for example, is very expensive inside the belt highway (A10 road), North-Amsterdam excluded, and parking facilities are scarce. Therefore, a house with parking facilities or indoor parking might be valued more in certain districts in Amsterdam. The impact of a parking facility or indoor parking is expected to be
greater in the center, south, west and east district. A dummy variable will indicate whether a house is located in one of these districts.
Table 1A: Descriptive Statistics of houses sold in Amsterdam during 2012 to 2014.
Mean Median Standard Deviation
Minimum Maximum Count
Housing Value
House price 290,512 218,000 252,496 26,500 6,195,000 21,628
House Characteristics
Living area (in m2) 85.3 73 49.6 12 890 21,628
1 if unit has a lot 0.0973 0 0.2964 0 1 21,628
Lot size (in m2) 174.1 122 332.4 14 8,386 2,105
Distance to city center (in m) 3375 2946 1960 137 11,497 21,628
Distance to closest school (in m) 285 258 169 6 3,300 21,628
1 if unit has primary school in 50m
0.0177 0 0.1317 0 1 21,628
1 if unit is apartment 0.8934 1 0.3086 0 1 21,628
1 if unit has elevator 0.1495 0 0.3566 0 1 21,628
1 if unit has attic 0.0323 0 0.1769 0 1 21,628
1 if unit has swimming pool 0.0009 0 0.0304 0 1 21,628
1 if unit has indoor parking 0.0217 0 0.1457 0 1 21,628
1 if unit has a parking facility 0.0935 0 0.2912 0 1 21,628
1 if unit is monument 0.0357 0 0.1856 0 1 21,628
1 if unit is monumental 0.0329 0 0.1783 0 1 21,628
Sub-district Characteristics
Average income (x1,000 euro) 32.4 29.5 8.2 23.3 64.6 21,628
% Unemployment 10.7 10 4.0 3 26 21,628
% Non-western immigrants 28.6 23.9 16.6 4 73.7 21,628
% Aged 65+ 11.9 10.7 4.8 3.3 30.2 21,628
% Families with children 23.5 21 8.4 5.8 52.3 21,628
% Owner-occupied houses 30.1 27.7 10.0 10.5 87.9 21,628
Average time till moving (years) 8.6 8.5 1.4 2.7 15.8 21,628
% Low education 23.6 21 11.4 3 55 21,628
% Medium education 32.5 32 5.1 21 49 21,628
% High education 43.8 46 14.9 10 72 21,628
Observations 21,628
This table looks at statistic values of most variables in the regression. They are split up in the dependent variable (house prices) and control variables (house characteristics and sub-district characteristics). The sub-district characteristics are calculated from 97 sub-districts in Amsterdam. Note: To calculate the distance to the city center, I took Damsquare (Dam) as the most central point. The X,Y-coordinates in meters are (121353, 487353). Furthermore the percentage of unemployed people is measured by the amount of people between 15 and 64 years old who make use of social benefits (WWB, IOAW and IOAZ), disability support (<80% disable) or the Unemployment Insurance Act (WW).
5. DATA DESCRIPTION
The data used for this research needs a description. This section declares where the data comes from and shows the summary statistics.
5.1 DATABASE
Data about Dutch real estate are available in the database of the Dutch Association of Real Estate Agents and Real Estate Experts, called NVM. This contains transaction prices of houses between 2012 and 2014. Transaction prices reflect what people want to pay for a house. The logarithm of the transaction price is the dependent variable in this research. Furthermore, the dataset contains information about house characteristics, such as: size, condition, district, period of construction, location expressed as coordinates etc. This data can contribute to control for variables that are correlated with the value of a house, other than the quality of schools. Size and condition for example are expected to have a positive effect on housing prices. The coordinates in the dataset are Cartesian coordinates of the postal code of the house. So they are not house specific, but specified per postal code.
The individual houses are matched to sub-districts they are located in. Data about characteristics of the sub-districts can be found in the datasets of the City of Amsterdam. In their publications of research and statistics of Amsterdam they divide the city into 8 city districts and 97 sub-districts. These sub-districts are part of 8 city districts: (A) Center, (B) Westpoort, (E) West, (F) New-West, (K) South, (M) East, (N) North and (T) South-East. See Figure 6 in the Appendix for a geographical overview. Every sub-district is indicated with one of these 8 letters followed by two numbers (for example: A01). Every sub-district has a unique set of characteristics, like average income, percentage of unemployment, percentage of non-western immigrants, the level of education, etc. Hence, all houses get matched to a sub-district and thus to specific neighborhood characteristics. Together with the house characteristics they form the control variables.
5.2 DESCRIPTION OF THE SOLD HOUSES
As shown in Table 1A on the previous page, there were 21,628 transactions in Amsterdam between 2012 and 2014 after deleting a few observations (see the data selection in the Appendix for an overview). The mean price of these transactions was 290,512 euro. There is substantial spread in the transaction prices, ranging from 26,500 to 6,195,000 euro. The
apartments. Around 90 percent of the transactions concerned an apartment. The rest of the houses, around 10 percent, are not apartments and often have a lot. Out of the 21,628 transactions, 2,105 of them concerned houses with a lot displaying an average size of 174.1 square meters. The average distance to the closest primary school of the houses in this dataset is 285 meters. The distances range from 6 meter to 3,300 meters. Figure 4 in the Appendix reveals that the distribution of the distances is left-skewed. Almost every house transaction between 2012 and 2014 had a school within a 1,000 meter radius (99.3 percent). Only 1.77 percent of the houses had a primary school within 50 meter distance. Not so many houses are provided with luxury like swimming pools or indoor parking (0.09 percent and 2.17 percent respectively). Parking facilities in general are hard to find in Amsterdam. Over 90 percent of the house transaction in 2012-2014 had no parking facilities.
Table 5 in the Appendix illustrates the frequency distribution of different types of houses. Upstairs apartments are the most prevailing in Amsterdam. From the houses sold in Amsterdam between 2012 and 2014, more than 58 percent were upstairs apartments. Around 14 percent were downstairs apartments. Figure 2 of the Appendix shows that more than 30 percent of the sold houses were constructed in the period between 1906 and 1930. More than half of all the transacted houses were constructed before 1944. Finally, around 68 percent of the houses are of average condition, 20 percent of the houses are in good condition and just 12 percent are in bad condition. This is displayed in Figure 3 of the Appendix.
On average, the house transactions come from neighborhoods where the average spendable income is around 32,400 euro per household per year. Though, most of the transactions come from neighborhoods that have an average spendable income lower than 32,400 euro, since the median income is 29,500 euro. Besides that, the transactions come from neighborhoods where on average live 11.9 percent people over the age of 65, 23.5 percent families with children and 28.6 percent non-western immigrants. The latter has a relatively high standard deviation of 16.6 percent. This means there is more spread in the distribution of non-western immigrants over the sub-districts and therefore might have a bigger influence on the difference in housing prices.
5.3 DATA ON SCHOOL QUALITY
The Dutch Ministry of Education publishes average scores on the final proficiency test of all primary schools, the Cito-test. The Cito-test measures the skill level of language, arithmetic, information processing and world orientation (optional). These scores range from 501 (lowest
score) to 550 (highest score) with an average of 535 and a standard deviation of 10. In this research the average Cito-score of three years, 2012, 2013 and 2014, was used in order to limit the effect of variation in scores caused by randomness. Kane and Staiger (2001) estimated that 28 percent of the variance in proficiency test scores of average-sized primary schools in North Carolina is due to sampling variation and about 10 percent is due to other non-persistent sources, like widespread diseases or distractions in the school. If a school happens to have a really bad or good graduation year, this will not reflect the true quality of the school. For this reason the average test scores of three years was used, or two years if only two years were available.
In the Netherlands most pupils get tested only once in the final eighth grade of primary school. This means there are no prior results to correct for earlier achievements. Hence making an added value measure by subtracting an earlier grade score from the final grade score, like Brasington (1999) and Downes and Zabel (2002) did, is not possible. In 2014, Dutch television news service RTL Nieuws published their own test scores, based on the Cito-test and the population of pupils. Their Cito-test scores are based on weighted Cito-scores. The weights they use are based on the socioeconomic statistics of the child’s postal code (like average income, education and composition of the population with the same postal code). This information is gathered on 1 October 2013. The Cito-scores are gathered from 2012 to 2014. This RTL-score can be used to display the added value by primary schools. The scores range between 4 and 10 with one decimal. The national average is 7 and the standard deviation is 1. We can retrieve this information on the website of RTL Nieuws, who published the adjusted scores in July 2014 (“Hoe vergelijkt RTL Nieuws de Cito-scores van 2014?”, 2014). The derivation of the RTL-score is explained in the Appendix. Not every primary school had information available to calculate the RTL-score. This means that the added value of some primary schools was not available and could not be matched to surrounding houses. Therefore, when regressing house prices on the added value RTL-score, there are less observations. This is visible in Table 1B on the next page.
Table 1B: Descriptive Statistics of the schools matched to the houses sold in Amsterdam between 2012 and 2014.
Mean Median Standard Deviation
Minimum Maximum Count
School Quality Closest Cito-score 534.1 533.4 5.11 521.9 546.1 21,628 Closest RTL-score 6.87 7.03 .83 4.88 9.09 21,027 Weighted Cito-score (1,000m) 534.1 533.8 3.42 524.7 544.3 21,492 Weighted RTL-score (1,000m) 6.89 6.90 .41 5.32 8.11 21,492 Weighted Cito-score (500m) 534.1 533.5 4.30 523.6 546.1 19,499 Weighted RTL-score (500m) 6.86 6.88 .66 5.06 8.53 19,461 Best Cito-score (1,000m) 540.1 540.2 3.79 527.9 546.1 21,492 Best RTL-score (1,000m) 7.90 8.03 .45 6.02 9.09 21,492 Best Cito-score (500m) 536.6 537.3 4.66 524.7 546.1 19,499 Best RTL-score (500m) 7.31 7.31 .73 5.06 9.09 19,461
Number of primary schools 197
This table looks at the different measures of school quality used in the hedonic pricing regression of this research. The distances in the brackets show what radius was used to calculate a weighted average or to find the best performing school. These statistics do not say anything general about the primary schools in Amsterdam, they only show the statistics of schools matched to the houses available in the NVM-database. The closest score is the average Cito-score or the average RTL-score of the closest primary school from the house. The weighted scores are the weighted average Cito-scores or RTL-scores within the range displayed between the brackets. The best score is the best average Cito-score or best average RTL-score of all primary schools within the range displayed between the brackets.
Aouragh et al. (2012) use an adjustment of the Cito-scores to account for the distributions of pupils in the schools in their attempt to measure school quality. They also correct for socioeconomic background because these were the only control variables available. The difference with the RTL-score is that Aouragh et al. (2012) does not correct for the parents’ educational background. Since this is an important non-school factor that could bias the value added of the school, this study will use the RTL-score as the added value measure indicated with Vi,s.
5.4 DATA FROM PRIMARY SCHOOLS MATCHED TO THE SOLD HOUSES
In Amsterdam there are 212 primary schools for regular primary education (bao). Primary schools with less than two years of test scores available in the years 2012, 2013 and 2014 are deleted from the database. There are 15 primary school missing more than one year of test scores, leaving 197 primary schools in the database to calculate average Cito-scores, as shown in Table 1B above. For this research three different methods where used to match primary schools to houses. This resulted in three different subsets as well, shown in Table 1B. The 500 meter radius is included as a robustness check. First every house from the entire
dataset was matched to the closest primary school. Since Cito-scores were available for all 197 included in the dataset, all 21,628 observations are combined with a Cito-score. However, although there are 197 primary schools with available Cito-scores, they do not necessarily have an RTL-score. For 193 primary schools an RTL-score could be calculated. This leaves the number of house transactions that could be matched to a primary school with an RTL-score to 21,027. The average Cito-score and average RTL-score for the closest primary school per house is 534.1 and 6.87 respectively.
Next, all the primary schools within a 1,000 meter are matched to the houses, if this was possible. Some houses are not surrounded by a primary school within a 1,000 radius and therefore there are fewer observations. Though as we know from Figure 4 in the Appendix, almost all the houses have at least one primary school in a range of 1,000 meter from the house. Furthermore, if we consider only the houses with at least one primary school within a 1,000 meter radius, we can create another distribution of the distance to the closest school. Figure 5 in the Appendix displays approximately 91 percent of these houses having a primary school within 500 meters. This means there will be less, but still a sufficient amount of, observations if the radius is scaled down to 500 meters. This 500 meter radius will be used as a check for robustness. In Table 1B, the number of observations drop if a 1,000 meter radius is compared with a 500 meter radius. Just as before, for all the 197 primary schools Cito-scores are available, but only for 193 school RTL-Cito-scores are available. The weighted average Cito-score and RTL-score within a 1,000 meter radius of the houses is 534.1 and 6.89 respectively. For a 500 meter radius this is 534.1 and 6.86. These values are close to each other and also close to the scores of the closest school. This might be because it is a weighted average, meaning that a primary school closer to a house receives a higher weight than schools farther away from a certain house.
Finally, only the best primary school within a 1,000 meter and a 500 meter radius of a house is used to measure school quality. The density of primary schools is high in Amsterdam, with almost 200 primary schools in a city with a surface area of only 219 square meters. Hence, the closest primary school is not the only option for a lot of households. Most parents can easily reach another primary school besides their closest primary school. Therefore this measure of school quality is created as well, under the assumption that parents will pick the best scoring primary school within a certain range (1,000 meter radius). The mean scores of the best primary schools within a 1,000 meter radius are 540.1 for the Cito-score and 7.90 for
5.5 DATA FROM PRIMARY SCHOOLS IN GENERAL
General statistics about the 197 primary schools in Amsterdam are shown in Table 2 below. For these schools at least two years of Cito-scores were available. The average Cito-score per primary school in Amsterdam from 2012 to 2014 is 533.2. This is lower than the normalized national average of 535. The standard deviation is only 5.2, also lower than the normalized national standard deviation of 10. The average RTL-score is 6.9, with a median of 7.1. This is close to the national normalized average of 7. Only 193 primary schools have an RTL-score. This is because the calculation of the RTL-score demands a lot more information than the Cito-score. Four schools were dropped since this information was not available. The average percentage of pupils with low educated parents is 43.7 percent per school. The average percentage of pupils with medium and high educated parents per school are both around 28 percent. The average household income per pupil per school is 30,861 euro.
Table 2: School characteristics of primary schools in Amsterdam between 2012-2014
Mean Median Standard Deviation
Minimum Maximum Count
Average Cito-score 533.2 531.9 5.2 521.9 546.1 197
Average RTL-score 6.9 7.1 0.8 4.9 9.1 193
Average number of pupils that took the Cito-test per school
33.7 32.3 14.9 6.3 79.3 195
% Pupils with low educated parents
43.7 46.6 10.8 21.1 60.9 195
% Pupils with medium educated parents
28.5 28.0 3.2 20.5 37.0 195
% Pupils with high educated parents
27.8 28.4 12.2 10.6 53.2 195
Average household income per pupil
30,861 28,738 5,553 23,873 53,086 195
Observations 197
In this table the general statistics are shown of the 197 primary schools in Amsterdam. It gives some general information about the test-scores and the background of the pupils.
6. RESULTS
To estimate the effect of school quality on housing prices, Equation (1) and Equation (2) are applied. The quality of primary schools is measured by proficiency test scores (Cito-score) or added value scores (RTL-score). The variables for education level per sub-districts are deleted from the regression, because they are diagnosed with high levels of multicollinearity8. Also, the distance to the city center correlates high with the other sub-district controls. Therefore this distance is not taken into account. Houses are matched to primary schools by three different methods. All methods showed a high adjusted R-squared, which indicates the explanatory power of the regressions is high. The methods used to match houses to primary schools are as follows:
1) Only the closest school is matched to a house.
2a) All schools within a 1,000 meter radius are matched to a house, using a weighted average.
2b) All schools within a 500 meter radius are matched to a house, using a weighted average (robustness check).
3a) Only the best performing school within a 1,000 meter radius is matched to a
house.
3b) Only the best performing school within a 1,000 meter radius is matched to a
house (robustness check).
The results of methods 1, 2a and 3a will be discussed in this section.
6.1 THE EFFECT OF THE CLOSEST PRIMARY SCHOOL
Table 3 on the next page presents the results using method 1. Column 1 and 2 show the results of the regression without controlling for neighborhood characteristics. Living area and lot size have the expected positive sign. Just as the condition, and whether the house contains a swimming pool, elevator and whether it is monumental. Surprisingly parking facilities show a negative effect, but this might be because of omitted variables, like neighborhood characteristics.
8 The values of the Variance Inflation Factor (VIF), a measure for multicollinearity, were 668.37 for low education, 124.34 for average education and 1098.79 for high education. The rule of thumb is that if VIF exceeds 10, multicollinearity is high. This can lead to insignificance in some of the regression coefficients
Table 3: Hedonic pricing regression results using the closest primary school. Dependent variable is the natural logarithm of house price.
No neighborhood controls Sub-districts dummies Neighborhood controls
(1) (2) (3) (4) (5) (6) ln(Cito-score) 9.170*** (55.97) 0.364* (2.23) 0.902*** (5.84) ln(RTL-score) 0.125*** (9.50) 0.00301 (0.28) -0.00540 (-0.54) ln(living area) 0.893*** (219.98) 0.933*** (213.03) 0.842*** (272.62) 0.843*** (267.56) 0.840*** (249.08) 0.840*** (244.32) ln(lot size) 0.0109*** (3.52) 0.0115 *** (3.43) 0.0115 *** (5.19) 0.0120 *** (5.36) 0.0155 *** (6.41) 0.0156 *** (6.38) Dummy for primary school within 50
meters -0.0202 (-1.84) -0.0252 *
(-2.12) -0.0132 (-1.73) -0.0122 (-1.57) 0.00328 (0.39) 0.00594 (0.69) Dummy for average condition 0.129***
(27.12) 0.120*** (23.28) 0.117*** (35.20) 0.115*** (34.36) 0.123*** (33.48) 0.122*** (32.61) Dummy for good condition 0.236***
(42.12) 0.233*** (38.10) 0.210*** (53.62) 0.209*** (52.45) 0.222*** (51.13) 0.221*** (50.03) Dummy for elevator 0.0296***
(5.56) 0.0482 *** (8.29) 0.0304 *** (7.97) 0.0312 *** (8.01) 0.0150 *** (3.61) 0.0153 *** (3.63) Dummy for attic -0.0267**
(-3.14) -0.0353 *** (-3.83) -0.0107 (-1.79) -0.0101 (-1.66) -0.0197 ** (-2.99) -0.0195 ** (-2.93) Dummy for parking facility -0.138***
(-15.71) -0.178 *** (-18.93) 0.0415 *** (6.00) 0.0419 *** (6.03) -0.0348 *** (-5.06) -0.0330 *** (-4.77) Dummy for indoor parking 0.0256
(1.24) 0.0236 (1.06) 0.0262 (1.81) 0.0263 (1.81) 0.0257 (1.61) 0.0265 (1.65) Dummy for monument 0.0379***
(4.67) 0.0661*** (7.59) 0.0144* (2.45) 0.0142* (2.41) 0.0185** (2.93) 0.0201** (3.18) Dummy for monumental 0.0546***
(6.38) 0.0821*** (8.90) 0.0284*** (4.67) 0.0284*** (4.64) 0.0296*** (4.45) 0.0307*** (4.58) Dummy for swimming pool 0.172***
(3.58) 0.193 *** (3.74) 0.121 *** (3.62) 0.119 *** (3.55) 0.107 ** (2.88) 0.104 ** (2.78) % non-western immigrants -0.00200*** (-11.97) -0.00206 *** (-12.22) % aged 65+ -0.0130*** (-31.14) -0.0126*** (-29.67)
% families with children -0.0174***
(-63.67) -0.0175 *** (-63.94) % owner-occupied houses -0.00492*** (-25.83) -0.00497 *** (-25.84)
ln(average moving time) 0.198***
(18.46) 0.189 *** (17.44) ln(average income) 0.793*** (57.46) 0.819*** (59.86) % unemployment 0.00159** (2.82) 0.00164 ** (2.90) Parking facility*district in ring road 0.300***
(26.99) 0.377 *** (31.49) 0.0650 *** (7.31) 0.0673 *** (7.49) 0.141 *** (15.93) 0.148 *** (16.49) Indoor parking*district in ring road -0.0267
(-1.07) -0.0309 (-1.15) -0.0340 * (-1.96) -0.0343 * (-1.96) -0.0469 * (-2.44) -0.0485 * (-2.51)
Dummies for type of house Yes Yes Yes Yes Yes Yes
Dummies for construction period Yes Yes Yes Yes Yes Yes
Sub-district fixed-effects No No Yes Yes No No
Quarterly fixed-effects Yes Yes No No Yes Yes
Quarterly dummies*district dummies No No Yes Yes No No
Constant -49.13*** (-47.85) 8.071 *** (198.62) 6.506 *** (6.32) 8.786 *** (247.12) (0.64) 0.620 6.232 *** (108.78) Observations 21,628 21,027 21,628 21,027 21,628 21,027 Adjusted R2 0.845 0.824 0.927 0.927 0.907 0.908
t statistics in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001. In this table we see the results for regressing the natural
logarithm of housing prices on the quality of the closest school to a house. Dummies for type of house are included. For every house type in table 5 a dummy was created. For all construction periods in Figure 2 a dummy was created as well. Lastly, for all sub-districts displayed in Figure 6 a dummy was created and included to control for neighborhood fixed-effects. There are six columns. Each column belongs to one of the three regressions: one that does not include neighborhood controls, one that includes neighborhood fixed-effect and one with neighborhood controls. Living area and lot size are measured in square meters. Average time until moving was expressed in years and average income in euro’s times 1,000.
We can see that both the proficiency Cito-score and the value added RTL-score of the closest primary school are positively valued by the housing market in this regression. If the Cito-scores improve with 1 percent, housing prices rise with 9.170 percent. If the value added score of a primary school improves with 1 percent, than housing prices will rise with 0.125 percent. Black (1999) found that controlling for more sub-district characteristics leads to a smaller positive effect of school quality on housing prices. In other words, the results of the basic hedonic regression method are biased upwards because sub-district variables are correlated with housing prices and school quality.
In column 3 and 4, dummies for the sub-districts are included in the regression. This way neighborhood characteristic are kept fixed. Including these fixed-effects for sub-districts leads to a decrease in the coefficient for the logarithmic Cito-score from 9.170 to 0.364 and remains significant at the 5 percent level. The coefficient of the RTL-score declined from 0.125 to 0.00301, though was not statistically significant anymore. The lower coefficients are in line with the results of prior research. Clapp et al. (2004) also found that the effect of school quality decreased when they used town and tract fixed-effect models instead of a hedonic regression without fixed-effects.
In Column 5 and 6, sub-district control variables are manually added to the regression, replacing the sub-district fixed-effects. The results show that the proficiency test score of the closest school has a positive impact on house prices, significant at the 0.1 percent level. A 1 percent increase in Cito-score leads to a 0.902 percent increase in house prices. The coefficient of the RTL-score has a negative sign, though is not significant. Compared with the no control regression, the coefficients for school quality decreased substantially. The RTL-score even ended up displaying a negative coefficient, although it did not show any significance.
Notice that the percentage of unemployment in a sub-district has a positive influence on housing prices. This is an unexpected sign. You would expect that a sub-district with a high unemployment rate is unappealing for house buyers and therefore has a negative effect. However, the effect of unemployment might be captured in the percentage of people aged over 65, the percentage non-western immigrants and the average income. One might wonder if house buyers still would be interested in the unemployment rate, if controls are already in place for these other variables. The interaction effect between parking facilities and houses inside the ring road highway is positive and significant at 0.1 percent. Though, the interaction
effect between an indoor parking spot and the houses inside the ring road do not show a positive effect, which is unexpected.
Overall proficiency test score of the closest primary school is capitalized in house prices using this method and the value added of a primary school is not. This might be because parents value Cito-score more than the RTL-score since the Cito-score is historically the traditional measure of school quality. Alternatively it might be because the RTL measure for added value controls for socioeconomic background of the pupils (see the derivation of the RTL-score in the Appendix), while we also include control variables for socioeconomic background of the sub-district in the regression. If we assume pupils go to a primary school nearby their home, we might be controlling twice for this effect.
6.2 THE EFFECT OF A WEIGHTED AVERAGE OF SCHOOL QUALITY
Next we look at the results when we calculate a weighted average of all schools that surround houses in a 1,000 meter radius, corresponding with method 2a. The weighted average is created according to Equation (4) and Equation (5). The results are displayed in Table 6 in the Appendix. The addition ‘a’ behind each column indicates a radius of 500 meter was used. An addition ‘b’ indicates a 1,000 meter radius. In columns 1b and 2b we see the results of a regression with sub-district fixed-effects. The effect of weighted average proficiency test score on house prices is positive and significant: a 1 percent increase in Cito-scores leads to a 2.444 percent increase in house prices. Though the effect of a school’s added value is negative and not significant. Furthermore, in line with the study of Conroy et al. (2016), the dummy for living within 50 meters of a primary school now shows a negative and significant effect on the house prices.
Columns 3b and 4b show the results of adding sub-district controls to the regression. Again the weighted average proficiency test scores show positive and significant results in line with the prior results of Downes and Zabel (2002) and Black (1999). A 1 percent increase in the weighted average Cito-scores leads to a 4.458 percent increase in house prices. Though in column 4b, we find a negative and significant effect of the weighted average value added on house prices. A 1 percent increase in the weighted average RTL-score leads to a 0.0901 decrease in house prices. An explanation for this might be that the RTL-score and sub-district characteristics are correlated. The RTL-score measures the added value by compensating for the characteristics of the pupils’ sub-district. So if more pupils live in a “bad” sub-district, the primary school will receive more compensation on the RTL-score. Hence, houses located in
“bad” sub-districts, where house prices are presumably lower, might be matched to schools with a decent RTL-score, since the school received compensation. This might lead to a negative impact of added value on house prices.
Finally, all neighborhood control variables are highly significant (except for the unemployment rate). The possible reason is discussed in the previous sub-section. Overall, weighted average proficiency test scores have the expected impact on house prices, while the weighted average added value does not show consistent results. A possible argument why added value might not be capitalized in the house prices is stated in the last phrases of the previous sub-section.
6.3 THE EFFECT OF THE BEST PERFORMING SCHOOL IN AN AREA
Table 7 in the Appendix outlines the results of method 3a. For this method only the score of the best primary school within a 1,000 meter range is used as a measure for school quality. Hence, it is assumed that it is important for parents that there is at least one good school in that area. In columns 1b and 2b, the logarithm of housing prices is regressed against school quality and house characteristics, keeping the sub-district effects fixed. Similar to before, a positive and significant coefficient for the effect of the best proficiency test score on house prices is found. A 1 percent increase in the average Cito-score of the best performing primary school within a 1,000 meter range makes house prices increase by 1.222 percent. However, the measure for added value also shows a positive and significant sign. If the RTL-score increases with 1 percent, house prices rise with 0.129 percent. The effects are all significant at the 0.1 percent level.
When we add control variables for sub-districts to the regression instead of fixed-effects, we also get positive and significant results for the effect of the best proficiency test score and value added measure. The values of the coefficient are 5.621 and 0.202 for Cito-score and RTL-score respectively. Both the proficiency test score and the value added measure are capitalized in the house prices. This was not the case with method 1 and 2a. Using method 1 and 2a, the proximity of the primary school plays an important role. In this sub-section, using method 3a, proximity does not play such a big role. The primary school just has to be in a 1,000 meter range, but it does not make any difference if the school is on a 10 meter distance or a 990 meter distance. This might be an explanation why the added-value measure shows a positive and significant impact on house prices. Perhaps parents just value the fact that there is at least one primary school around with a good score for added value.
