What is the influence of crime on house prices in Amsterdam and how does prostitution influence this relation?

(1)

1

Universiteit van Amsterdam

What is the influence of crime on house prices in Amsterdam and how does prostitution influence this relation?

Abstract

This paper examines the relation between crime, house prices and prostitution. Data relating to 68,908 housing transactions, crime statistics for the period 2003-2014 and information regarding the closing of prostitution properties are used to do a hedonic regression. The crime rates of neighborhoods are weighted according to the distance of a transaction from the central point of the neighborhood. The transactions close to the prostitution properties and within 1,5 years before closing are compared to the transactions after the closing. A very small effect is found for crime and a somewhat larger effect for the prostitution properties. The coefficients of the crime variables mainly changed by including the current properties, and not by including the closed properties.

Student: Arthur Croes

Student Number: 10095624

Thesis Supervisor: Erasmo Giambona

Faculty: Business and Economics

Study: MSc Business Economics

Double Specialization: Finance & Real Estate Finance

(2)

2

Acknowledgements

I would like to thank my supervisors Erasmo Giambona and Rafael Perez Ribas for providing me with the possibility to write an interesting thesis. I would also like to thank the Nederlandse Vereniging voor Makelaren for providing data regarding housing transactions. Finally, I would like to thank the municipality of Amsterdam for being cooperative and providing data regarding the prostitution properties.

Statement of originality

Hereby I , Arthur Croes, declare that I have written this thesis and that I have complete responsibility over the contents of it.

I confirm that the text in this thesis is original and that I have not used other sources than mentioned in the text and references.

The faculty of Business and Economics is solely responsible for supervising until the point of handing in the thesis, not for the contents.

(3)

3

1. Introduction

Crime combat and prevention is an important part of municipal policy. A significant part of the budget is usually allocated to the safety department, which is responsible for combatting crime. Policy is aimed at improving living conditions and providing a safe environment for the residents of the city. If it fails, high crime levels and unemployment might cause people to leave the area, and the consequential vacancy would provide more opportunity for crime. High levels of crime are likely to prevent new residents to move in to the neighborhood, or at least make them more hesitant. This reduced interest in properties located in high crime areas should be reflected in the value of these properties. Another way to look at it is a quantification of the benefits of municipal policy aimed at crime. How much value people attach to a reduced crime level in their neighborhood will be reflected in the house prices. The main research question is thus the following:

What is the effect of crime on house prices?

There have been several studies that sought to quantify the willingness to pay for low crime in a similar fashion as this research. Thaler (1976) was the first to research the relation between crime and house prices. He used data from Rochester and found that the cost of the average property crime was roughly $500. More recent studies (Lynch & Rasmussen, 2001; Gibbons, 2004) have also found a negative effect of crime on house prices. There are three subquestions that will be answered in this research. These are the following three questions:

What part of the effect of crime on house prices can be explained by the closing of prostitution properties?

What is the willingness to pay of households for a lower incidence of prostitution in the neighborhood?

What are the portfolio implications for households living in high crime or high prostitution incidence areas?

As can be deduced from the questions above, the analysis of the relation between crime and house prices will be supplemented by analyzing the impact of prostitution on this relationship. The possibility for this part of the research is provided by a policy of the municipality of Amsterdam, which provides a natural experimental setting. In 2007 the municipality of Amsterdam implemented a policy aimed at improving the city outlook and reducing crime in the center of Amsterdam. One of the ways the municipality planned to achieve this was by

(5)

5

reducing window prostitution in the center. Reducing window prostitution would give Amsterdam a more diverse and higher grade center and would improve living conditions. Improved living conditions would make it more attractive to live in the center of Amsterdam, which should be reflected in the house prices. In some areas the municipality closed several windows, while in other areas none were closed. If the potential home buyers in the area perceived the presence of a prostitution property as a risk for higher crime, or if they considered it to be a negative factor, the value of the properties in the direct vicinity should show this. Knowledge of the exact date and location of the closing of the prostitution properties can be combined with housing transaction data and crime data to explain part of the variation in house prices by an analysis. To the best knowledge of the author, this relationship between crime, prostitution and house prices has not been researched. This research will provide relevant insight in the value households attach to a crime free neighborhood. As there are existing researches aimed at the same topic, the main contribution of the research will be the focus on prostitution. This focus on prostitution might provide useful insights. In the Netherlands, prostitution is regulated by law, in which the opportunity is provided for municipalities to form their own policy towards it. Several municipalities in the Netherlands have a license system or give properties a prostitution marking in zoning plans. Because the municipalities have the ability to regulate prostitution and in this case Amsterdam is spending a lot of money to accomplish the targets, the research could be of great use to municipalities (Rekenkamer, 2011). When placed in a wider context, this research will add to existing literature by supplementing insight into households willingness to pay for certain urban amenities. Several existing papers have similar subjects. Linden & Rockoff (2008) do research into the effect of sex offenders moving in and out of neighborhoods on house prices. They find that properties located within 0.1 mile of the sex offender fall by about 4%, with the effect vanishing quickly when larger distances are considered. Other examples are the effect of a better school on house prices (Black, 1999) and the effect of foreclosures on prices of surrounding houses (Lin et al., 2007). The relation between crime and house prices will be investigated using both cross-sectional and temporal variation. A concern with a cross sectional analysis is that the crime rate is not exogenous, that the crime rate is possibly dependent on factors that also determine the house prices. This would result in biased estimates, but there are methodological solutions that will be expatiated in the methodology section. The third part of this research will be a small theoretical analysis regarding the consequences of the policy for the portfolio of the households that live in the areas with closed prostitution properties and/or high crime rate areas. As can be

(6)

6

seen in statistics provided by the Central Bureau of Statistics (CBS, 2015), housing is an important part of the portfolio of most households in the Netherlands. The average percentage housing is of total household assets in the Netherlands is 59.7%(CBS, 2015). The 1012 policy of the municipality of Amsterdam might have consequences for the portfolio composition of households living in the area.

The data regarding houses are provided by the NVM (Dutch association of Realtors). After removing outliers with respect to certain variables, the dataset contains 296,850 transactions in the period 1975-2014. This number of transactions is reduced further by selecting the years useful for the analysis and removing transactions with missing observations for important variables. The crime data is publicly available for the city of Amsterdam, and is thus obtained from the website of the department of research and statistics of the municipality of Amsterdam for the years 2003-2014. The data regarding prostitution properties was acquired from the municipality of Amsterdam. A list of closed prostitution properties, together with the date of closure, was requested and granted by the municipality. A list of the current prostitution properties was deduced by the floor plans of the municipality, after removing the closed prostitution properties from the list. An overview of these lists, along with a visual representation on a map, can be found in Table 1 and Figure 1 in the appendix. A more detailed description of the data and preparation of the data can be found in the data and summary statistics section on page 19.

Answers to the research questions will be sought using a hedonic regression model. Supply of housing is assumed inelastic and consumers differ in preferences for characteristics. In equilibrium consumers with the same preferences obtain the same utility, because better characteristics are compensated by a higher price. Regressing the transaction prices on house and neighborhood characteristics and a crime variable would provide an indication for the attractiveness of lower crime to homebuyers. A problem with the analysis regarding crime is that the municipality provides crime statistics for certain area’s which are of a relatively large size. As a result there is only a limited number of area’s for which there are crime statistics available in the city of Amsterdam. If each transaction would be assigned the crime rate of the area in which it is located, the analysis would be limited as differences in the occurrence of crime within an area would be ignored. In order to be able to surpass this problem each individual transaction gets a weighted crime rate. The central point of each district is determined, along with the distance of each house to each central point. The crime rate of the districts will be weighted by the distance to the house, which will cause each transaction to be

(7)

7

assigned a different crime rate. The same procedure is applied to variables related to demographic composure. An important benefit of this procedure is that the factors that perhaps both determine the house price and the neighborhood characteristics are largely removed. This is because the crime rates depend on statistics from a much larger area with this procedure. Regarding the prostitution properties, the exact date is used to compare the transaction prices of houses located within a certain distance before and after the closing of the prostitution properties.

The rest of the thesis is organized as follows. First, a short overview of past developments regarding prostitution policy in the Netherlands is provided, with specific focus on the 1012 policy of the municipality of Amsterdam. After this the existing literature about related topics, as mentioned above, is described. The third section is for the methodology, after which the data and descriptive statistics will be discussed. Following are the results of the analysis, together with robustness checks. Subsequently the theoretical section regarding the portfolio implications for households will be presented. Finally a conclusion will be drawn.

(8)

8

2. Literature Review and Background

2.1 Background

2.1.1 Prostitution in the Netherlands

When the Netherlands was occupied by France in the 19th century, prostitution was regulated

to some extent. After the French left the criticism of this policy grew, leading to the criminalizing of brothels in 1911(Penal Law, 1911). Despite brothels being prohibited in this law, prostitutes were not penalized (Brants, 1998). According to Brants (1998), it did not take long before this law was not fully enforced anymore. Large Dutch cities adopted an unofficial policy of tolerance, local authorities concentrated prostitution in certain areas and tolerated private houses (Outshoorn, 2012). In the 1970’s the sex industry expanded beyond the red light districts of the large cities. Drugs entered the world of prostitution and the sex-industry was commercialized (Brants, 1998). In an attempt to combat this, the municipality of Rotterdam tried to regulate prostitution (Outshoorn, 2012). Attempts to regulate were denied by court, stating that they could not regulate what is prohibited by law. Municipalities felt they had insufficient means to combat prostitution. According to Visser et al. (2000), this was one of three motivations for considering a new law. The second motivation was that women should have the right to choose the profession of prostitute. The third motivation was grounded in the growing trafficking of humans, which was linked to the prostitution branch. Changing the law would allow the authorities to persecute traffickers in a more active and strict way.

2.1.2 Amsterdam

Between 1970 and 1980, the municipality of Amsterdam developed a policy of tolerance. Under this policy the municipality withheld from acting against brothels that met certain requirements. Besides this policy, brothels were subjected to the ‘contigenteringsstelsel’, which meant that prostitution was only allowed in areas with a history of prostitution. Within these areas, the number of brothels was not allowed to rise. However, there was no active monitoring to see whether the brothels complied with the conditions. The limited expansion was regulated through zoning plans, but the rest of the tolerance policy was not really enforced. On the first of January 1996, Amsterdam introduced declarations of eligibility with rules concerning the location, accommodation and the operational management (Municipality of Amsterdam, 2000). With respect to the operational management, the municipality judged the eligibility based on the following requirements: No criminal activity related to prostitution, no prostitution by minors, no public disturbances, no illegal foreign prostitutes and good care for the health of the prostitutes (Municipality of Amsterdam, 2000). The way prostitution was handled resembled

(9)

9

policy as if it was legal, and in Amsterdam the policy hardly had to change after the legalization (Brants, 1998).

After the introduction of the regulation allowing prostitutes of certain countries to work in the Netherlands, there was a large inflow of Eastern European prostitutes into Amsterdam. Simultaneously there was a large increase in windows in use, primarily in the Red Light District (van Wijk et al., 2010). The number of windows in the Red Light District increased from around 200 to around 400. At first escort companies were not included in the policy of the Municipality of Amsterdam, but after incidents in this sector the license policy was enlarged to include escort companies in 2008 (Municipality of Amsterdam, 2008)

In 2006, the municipality of Amsterdam started using the law Bibob to screen the prostitution branch (Bieleman et al.,2006). The Bibob law was introduced in 2003 and was aimed at the testing of holders or appliers of licenses and subsidies by a national bureau (Kamerstukken, 2003).The municipality of Amsterdam used the Bibob law for 108 sexual institutions in the center to be reviewed by bureau Bibob in 2006 and 2007 (Municipality of Amsterdam, 2009). Of these 108 institutions, 58 licenses were suspended. The municipality of Amsterdam

continued to try to combat crime in the red light district. On the 4th of December 2007, the

municipal council accepted the proposed project 1012 (Municipality of Amsterdam, 2007).

2.1.3 Project 1012

The reason the municipality of Amsterdam started this project is that they determined the presence of a criminal infrastructure in the center of Amsterdam, that was supported by the large concentration of criminogenic services as for example window brothels and coffee shops

(Municipality of Amsterdam, 2009a_{). The municipality also stated that the area with the postal}

code 1012 was one of the least safe and filthiest areas in Amsterdam (Municipality of Amsterdam, 2008). In order to improve the character and reduce criminal activities in the area, the municipality implemented project 1012. The project consists of three parts (Municipality of Amsterdam, 2011). The first part is the improvement of public spaces. The Damrak and the Rokin for example, near Central Station, are aimed to be redesigned in order to improve the entrance to Amsterdam. This is the most important target of the public spaces improvement but there are other and smaller goals. The second part of the project is called ‘Sleutelprojecten’, a number of key projects that should change the outlook of certain small parts of the area or renovate certain buildings. An example is the key project ‘Cluster Oudekerksplein’. It is aimed at redesigning the Oudekerksplein, for example by granting a permit for a high-end restaurant

(10)

10

and giving former prostitution properties a new function. This is related to the final part, which is called ‘Straatgerichte Aanpak’. The reduction of criminogenic services, as mentioned above, is aimed at 18 streets in the area. Services such as smartshops, coffeeshops and also window prostitution are considered low grade services and are thus aimed to be reduced in number. This part of the project is the most important one for this research, since the reduction in prostitution windows allows this research to be carried out. The municipality of Amsterdam seeks to accomplish this reduction either by acquiring real estate, as the properties discussed in the previous paragraph, or by agreeing with owners on a new function and higher quality

(Municipality of Amsterdam, 2009a). Besides buying prostitution properties, as will be

explained below, 26 coffeeshops have to close in the area (Municipality of Amsterdam, 2009a).

Of these 26, 4 closed based on public disturbances, 9 closed in 2013, 6 were planned to close in 2014 and 7 were planned to close in 2015 (Municipality of Amsterdam, 2014). Whether all of these coffeeshops were closed is uncertain.

The target of the restructuring is to concentrate window prostitution in two areas: around the Oude Nieuwstraat and around the Oudezijds Achterburgwal. From the original 477 windows, 293 are planned to remain in use for prostitution. In order to be able to redesign the windows,

the zoning plans of the municipality had to be changed. On the 8th of February 2011 the

procedure of redesigning window brothels started. 62 properties were to be closed, of which 52 are in the historical center. The 62 properties contained 192 windows. Of the 62 properties, 30 properties were already acquired by the municipality and were not used for prostitution anymore. These properties were leased out on landlease contracts to housing corporations. The municipality of Amsterdam provided a list of all prostitution properties that were closed as part of the 1012 project. This list can be found in table 8 in the appendix on page 54.

(11)

11 2.2. Literature Overview

The analysis that will be conducted in this thesis can be placed in the wider context of households willingness to pay for neighborhood amenities. There have been a large number of studies that have tried to capture the value households attach to certain amenities by analyzing the differences in property values near the amenities. There is a problem with these type of analyses that has been discussed often (Black, 1999; Gibbons, 2003 among others). There is a possibility that the factors that determine the value of a property also influence the independent variable. Taken crime as example, it might be that lower income households are more likely to commit crime than high income households. Since lower income households generally live in less expensive houses, a negative relation between crime and property values would be found if one did not take into account this problem. neighborhoods with high crime levels attract lower income households. Since one cannot include all characteristics regarding the property, neighborhood and owners, omitted variable bias such as this is a large concern when doing research with hedonic regressions. This is a challenge for all authors doing research into the value of urban amenities using variation in property values.

Black(1999) wrote a paper on parental valuation of elementary education. In her research, she used elementary test scores as the dependent variable. With schools, the problem arises that better schools tend to be located in better neighborhoods. Research into the relation between house prices and elementary test scores will find a positive relation, but this does not necessarily reflect parental valuation of a good elementary education. The solution of Black(1999) to this problem wass to compare houses on opposite sides of attendance boundaries. These are geographical lines that determine which elementary school ones child will attend. Variation in demographic variables and taxes is limited because the houses are located in the direct vicinity of each other, so the only large difference is the elementary school. She found that a 5 percent increase in elementary test scores leads to an increase of around 2.1 percent in property values. She stated that this number is around half the effect found in hedonic regressions that do not make use of such a methodology.

Thaler (1978) was one of the first to do research into the impact of crime on house prices. He used a hedonic model (Rosen, 1974) to estimate the effect of an increase in crime rates on house prices. The author used crime statistics on the level of tracts and assigned each house the value of the crime rate of the tract in which it was located. He found that an increase by one standard deviation in crime decreases the average price per home by 3%. Thaler (1978) made use of crime statistics on tract level. The analysis of Thaler (1978) might also be prone to biases. The

(12)

12

problem mentioned earlier is not taken into account in his research. Subsequent to his paper, other authors have used various methods to quantify the value households attach to a low crime neighborhood.

Lynch and Rasmussen(2001) used a database containing information on over 2800 house sales to analyze the relation between crime and house prices. Besides using the number of index crimes as a dependent variable, they weighted the seriousness of offences as an alternative measure. They stated that using neighborhood boundaries, such as census tracts, for neighborhood data would result in biased results. This is because a large number of observations would get assigned the same values for certain variables, while within these census tracts there might be significant differences. In order to surpass this problem they used a geographic information system to define a unique neighborhood for each transaction. The data they used was only available at a block group level, so they weighted the data of block groups within one mile of a house to generate different neighborhood data for each observation. The authors assigned the value for crime to each house according to the police beat in which the house was located and used a different specification weighting the crimes by seriousness. A 1% increase in the number of violent crimes results in a 0.05% decrease in property value. For the logarithmic transformation of property crimes the authors found a positive coefficient. Lynch and Rasmussen(2001) attributed this to higher reporting rates in high property value areas. The results from the analysis suggest that crime has a limited impact on average house prices, but that high crime areas do contain discounted houses. When adding an indicator variable equal to one for houses located in the two deciles with highest crime, they found a significant effect. The average house located in a high crime neighborhood is discounted from 94,000 to 57,000. A positive coefficient for property crimes is also found by Gibbons (2003), who conducted a similar research. He explained the positive coefficient by reasoning that properties with higher values are more attractive targets for burglary than lower valued properties. Gibbons stressed the importance of the problem proposed earlier. Contrary to the research of Lynch and Rasmussen, which analysis might still be prone to the problem mentioned earlier, he proposed a methodological solution. Gibbons (2003) used postcode sector level data to propose an Instrumental Variable equation, with the spatial lags of crime as an instrument. In this way the author controlled for the spatial distribution of unobserved price factors. With a standard OLS regression the factors determining the values of the properties might also determine the neighborhood characteristics in the immediate vicinity. The spatial lags of crime are assumed to be uncorrelated to the unobserved property characteristics that determine the house price,

(13)

13

after removing spatial correlation. The crime variable observations were weighted by the distance to the property, with the weight inversely determined by the distance. The area for which the number of residential property crimes were calculated is within 250 meters of the property. According to the author, the best specification is one of the three where the regression estimates are differenced from their locally weighted average. Result was that an increase of 5 crimes per km² in the expected density of reported criminal damage leads to a drop in property value of 1.6%. The incidence of burglary had no significant effect on house prices.

Other papers have looked at the relation between crime risk and property values in a different way. In these studies, the change in crime risk is a result of a sex offender moving into a neighborhood. Pope (2008) used Megan’s law to obtain information on the residential locations of sex offenders. This is a law that requires information about sexual offenders to be made public. Besides the residential location, information about transactions and houses was used and a database which contained the date the sex offender moved out. This dataset provided the author with a quasi-experimental setting where the experiment was reversed. While the specification of crime risk is different than in this thesis, the paper is comparable with the part of the analysis regarding the prostitution properties. The prostitution properties can not really be seen as a natural experiment, since the location of the properties is not random, but still provides a natural setting to which a similar methodology can be applied. Pope (2008) created two subsets out of the transaction data. Subset one contained houses that were never near a sex offender or only once near a sex offender. Subset two was created by dropping all houses further than 0.3 miles from a sex offenders residential location from subset 1. He created dummy variables that indicated whether the nearest sex offender was located within 0.1, 0.2 or 0.3 miles of the house, and whether the house was sold within a year of the sex offender moving into the neighborhood. Characteristics of the sold houses were included, together with year dummies. The results indicated that after a sex offender moves into the neighborhood, prices of properties located within 0.1 miles of the offender drop by 2.3%. Between 0.1 and 0.2 miles the coefficient is insignificant. Prior to the sex offender moving into the neighborhood there was no statistical significant difference between the different distance dummies. After the sex offender moves out of the neighborhood house prices rise again. Linden & Rockoff (2008) conducted a similar research. Using more distant properties as a control group, they did a difference in difference analysis. They also found a significant effect only for the properties located within 0.1 miles, results for larger distances were insignificant.

(14)

14

It is not certain that these distances will also be the optimal distances for this analysis but it provides an indication at which distance the average household experiences some kind of influence of changes in certain characteristics. The distance chosen by Linden & Rockoff reflects the distance at which offenders are legally obliged to inform neighbors. Since in the red light district of Amsterdam there often is a prostitution property not very far away it is likely that the possible effect will also dissipate fairly quickly when larger distances are used. Buyers of homes might not be that concerned with the closing of prostitution properties half a mile away. However, the entrance of a sex offender into a neighborhood can influence the perceived crime risk but will not impact the general outlook of the neighborhood. Prostitution properties are different because there is both a visual aspect as well as other factors related to it. The public that visits a certain area will change if there are less prostitution properties and there might be lower incidents of disturbance. Perhaps more comparable in this respect are studies that have investigated the impact of foreclosed properties on house prices.

Lin et al. (2007) researched what effect foreclosed properties have on prices of surrounding properties. Using a hedonic model, they estimate both distance and time effects. After controlling for a sample selection bias, the effect on house prices within 0.1 km of a recently foreclosed property is -8.7%. For a radius of 0.2 km this effect is -4.7%. For a property foreclosed between two and five years before the sale, the largest distance effect is -5.5%. The effect the authors found is significant within a distance of 0.9 km. With this research the problem of endogeneity also exists. A possible explanation for the results is that declining property values in a neighborhood cause more residents to default on their mortgage, and that it is not the direct consequence of properties foreclosing. Harding et al. (2009) point this out and control for this bias by using a large sample of repeat sales transactions. The authors jointly estimate the change in overall price level and the contagion effect with the repeat sales specification of a hedonic model. They find that having a property in the process of foreclosure in the vicinity can result in a drop in value of up to 1% per nearby property. Nearby in this case means within 300 feet (91.44 meter). Between 300-500 feet the effect is significantly smaller and at a distance of more than 500 feet (152.4 meter) the effect is insignificant. The differences in distance for which the effect is significant between the two researches might be caused by the difference in methodological approach.

The literature described above provides some indication regarding the size and sign of the coefficients of the crime variables. For the coefficients of the prostitution variables and their impact on the coefficients of the crime variables it is more difficult to form expectations. The

(15)

15

methodology can be compared best to the articles in which Megan’s law is used to estimate the impact of crime risk on property values. In these researches however, the authors suggest that living near a perpetrator would lead a household to experience higher crime risk. In this way the house prices are affected. With the prostitution properties it is probably a stretch that households only perceive a decreased crime risk after nearby prostitution properties were closed. The closing of nearby properties could affect factors like disturbances, incidents of violence or theft, but it will probably have a limited effect on burglary. The effect on the different types of crime might not be the only or foremost channel through which the property values are affected. Other factors such as less pedestrians walking by at possibly inconvenient hours, less noise and a cleaner and nicer looking surrounding area can also be assumed to play a large part. Though these factors are likely to be related to perceived crime risk, they are not of importance with the researches making use of Megan’s law. In this research the closing of the properties will only be used as a factor to explain deviations in house prices and crime rates and not as a direct indicator for crime risk.

(16)

16

3. Methodology

In this section, the methodology of the analysis will be discussed. First the analysis regarding crime will be discussed, after which the analysis with respect to the prostitution properties will be expatiated. The data available allows for both an intertemporal and cross-sectional analysis of households valuation of lower crime. For this research the hedonic model of Rosen (1974) is used. Supply of housing is assumed inelastic and consumers differ in preferences for characteristics. In equilibrium consumers with the same preferences obtain the same utility, because better characteristics are compensated by a higher price. The decision to buy a particular house is dependent on a whole range of factors. These do not only consist of property characteristics like house type and size but also location characteristics are important. School quality, socio-economical and racial composition of the neighborhood and crime rate, among many others. As discussed prior to this section, the relation between these types of amenities and property values has been researched often. A full list of the amenities and characteristics that influence the purchasing decision of a household is impossible to obtain. The problem that has been discussed earlier in this thesis is that the impossibility of obtaining such a list is likely to produce biases in the results obtained from the analysis. There might be some factors that influence several determinants of house prices, but are unobserved. The basic regression formula used in this analysis is as follows:

ln 𝑃 = 𝛼 + 𝑋_𝑖𝛽 + 𝑍_𝑗𝛾 + 𝐶𝑟𝑖𝑚𝑒𝛿 + 𝜀_𝑖𝑗 (1)

The variables will be estimated using OLS. The dependent variable is the logarithm of the house price. Following the standard of existing hedonic studies, as well as the researches discussed in the literature review, the logarithm of the house price is chosen as a dependent variable instead

of the house price. Xi is a vector of observable property characteristics, Zi is a vector of

neighborhood characteristics, Crime is the weighted density of crime in surrounding areas and

𝜀𝑖𝑗 is the error term. The vector of neighborhood characteristics include among other variables

the demographic composition of the area in which the house is located. Percentages of the population within certain age brackets, percent native and immigrants and average income are examples of neighborhood control variables. Following Lynch & Rasmussen (2001), each observation gets assigned a value based on the distance to the central points of each neighborhood. Neighborhood characteristics of neighborhoods located at a larger distance of the transaction will get a smaller weight in the value for the observation. This is to better take into account within area differences in neighborhood characteristics. Contrary to Lynch &

(17)

17

Rasmussen (2001), the crime variables will also be weighted by distance in this analysis. Crime rates of neighborhoods that are closer to the property will naturally hold a larger part of the weighted crime rate than crime rates of more distant neighborhoods. Distant neighborhoods will have nearly no impact on the weighted crime rates. The crime rates, and neighborhood variables, will be weighted according to the following function:

𝑐𝑟(𝑥) = ∑ 𝑤(𝑥)𝑢𝑖

𝑁 𝑖=1

∑𝑁𝑖=1𝑤𝑖(𝑥)

Where 𝑤𝑖(𝑥) = 1

𝑑(𝑥,𝑥𝑖)𝑝 , 𝑑(𝑥, 𝑥𝑖)𝑝 is the distance between point x and xi and p is the power

parameter. The crime rates are for neighborhoods and not for locations, so the centroids for

each neighborhood are calculated. These centroids function as location xi in the function. As is

apparent from the function above, the choice of p is of importance for the results of the process. A large p will decrease the influence of more distant neighborhoods, where a small p will make more distant neighborhoods more important. Because the impact of crime rates of distant neighborhoods on the crime rate at a certain location is likely to be rather small, the value chosen for p will reflect this and is thus chosen as such. The value of p is set to 2 in this analysis. The choice remains arbitrary and the sensitivity of the results with respect to p are tested in order to attain the impact of this decision. Besides p, a cut-off distance is used at which the neighborhoods are ignored in the determination of the values. The cut-off distance is set to 2 km in this analysis. This cut-off distance is necessary because neighborhoods far away can not be expected to have any influence on the crime rate, not even a small influence. Determining the crime and neighborhood variables by this procedure has several benefits. First of all, assigning rates to each house based on the neighborhood in which they are located would limit the variation in the sample. While there are a decent number of neighborhoods in Amsterdam, the number is too small to be able to do this analysis. Besides this the prostitution properties are concentrated in the center of Amsterdam, which consists of only a couple of neighborhoods. Using neighborhood crime figures would prevent the prostitution properties to be used in this thesis. Determining the variables in this way is also useful because it limits the problem proposed earlier. If house prices and crime rates are jointly by the same unobserved property values than using the crime rate of a certain area would result in a bias. By weighting the crime rates of different neighborhoods, a larger area is used to determine the crime rates. The crime rates of other neighborhoods can be assumed uncorrelated with the unobserved property characteristics in a certain neighborhood. Of course the crime rate of the neighborhood in which

(18)

18

a property is located will still have a weight in the crime rate assigned to the property, since the central point is highly likely to be within 2 km of the property. But the neighborhood in which a property is located already holds far more houses than just the houses in the direct vicinity of a house. This means the correlation between the neighborhood crime rate and unobservable property characteristics of a certain house within the neighborhood was already limited. The bias is thus reduced, but not completely removed. It is assumed that the bias will be modest because the local weighted crime rate will depend on much more data than just the number of crimes in the direct vicinity of the house. The crime variables are defined in several different forms, the analysis will determine which form generates the best results. Crime index rates for several categories are used, as well as the number of reported incidents on these crimes. Also obtained are objective as well as subjective safety indices. The different types of crimes can be expected to show some correlation. It is not unlikely that this correlation is too large to include multiple crime variables in one specification. The correlation between the variables will be checked to see if this is the case. If so, than the different types will each be used in a separate regression. A more detailed description of the crime variables can be found in the data description section on page 19.

Because there the bias was not completely removed, it is important that the location is controlled for adequately. This will be done in several different ways. In the first regression, no controls for spatial correlation are included. In the second regression, neighborhood fixed effects are included. The second regression formula will look like this:

ln 𝑃 = 𝛼 + 𝑋_𝑖𝛽 + 𝑍_𝑗𝛾 + 𝐶𝑟𝑖𝑚𝑒𝛿 + 𝑌𝑒𝑎𝑟𝐷𝑢𝑚 + 𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑟ℎ𝑜𝑜𝑑𝐷𝑢𝑚 + 𝜀_𝑖𝑗 (2)

Other controls for location, such as distance to the central station and longitude and latitude will also be included in separate specifications. In the regression formula above, year dummies are also included. It is important that market conditions are controlled for, since fluctuations in house prices have been fairly large during the sample period. The aggregate house price fluctuations or trends are controlled for by including year dummies. Other possibilities include half-year dummies and quarter dummies. To control for certain characteristics that only influence a part of the sample the standard errors are clustered on neighborhood or district level. Next the prostitution variables will be included in the analysis. This part of the analysis is done to answer two of the questions posed in the introduction. The first is what the change in the impact of crime is when one controls for the closed prostitution properties. In order to assess this the regression has to be expanded with variables related to the closing of the prostitution

(19)

19

properties. To be able to do this, first the treatment prostitution has to be defined. The problem with this is that the houses located closer to the prostitution properties might differ from houses located further from the properties. This is likely even, since the properties are located in the center of Amsterdam, where property values are higher. The data provided by the municipality of Amsterdam contains the date at which each prostitution property was emptied. The prices of the transactions within a certain range of the property before and after are compared, to control for preexisting differences between the houses located closer and further away. The regression formula with prostitution variables included is the following:

ln 𝑃 = 𝛼 + 𝑋𝑖𝛽 + 𝑍𝑗𝛾 + 𝐶𝑟𝑖𝑚𝑒𝛿 + 𝑃𝑟𝑜𝑠𝜃 + 𝑌𝑒𝑎𝑟𝐷𝑢𝑚 + 𝑃𝑟𝑜𝑠𝑃𝑜𝑠𝑡 + 𝑃𝑟𝑜𝑠𝐶𝑢𝑟 + 𝜀𝑖𝑗 (2)

The formula above is the same as (1) except that the prostitution variables have been added. Pros indicates the number of prostitution properties that was closed within a certain distance either in the 1,5 years before or the 1,5 years after the house was sold. The distances 300m, 200m and 100m, as used in the researches discussed in the literature review, will be used in the first regressions. These distances will be differed to see differences in results. Prospost indicates that the house was sold within the same distance as Pros after the prostitution property were closed. ProsCur is the number of existing prostitution properties within the same distances as the set of variables Pros. The coefficient of the ProsPost variable is the coefficient that contains the prostitution property treatment. It shows the effect of the change in the zoning of the properties on the surrounding property values. This specification requires the influence of the number of properties within a certain distance with changed zoning to increase linearly, which might not be accurate. It is difficult to assess what the difference in impact of each additional prostitution property within a specified range is. The set of properties with changed zoning is too small to deduce the correct form of the prostitution variable. In the third specification the PROS variable is replaced by a set of proximity dummies. These dummies indicate the distance to the closest prostitution property. The problem with this specification is that the distance to the closest prostitution property does not take into account the possibility of multiple properties being within a certain range. It is not unrealistic that if there is an effect, the effect will be larger if there is a cluster of former prostitution properties. Therefore a dummy variable for the second closest prostitution property is added. The list of prostitution properties provided by the municipality, table 1 in the appendix, also shows the number of windows for most of the prostitution properties. There are some significant deviations in the number of windows per property, the smallest number of windows is 1 while the largest is 8. It can be expected that these two properties have a different impact on transactions nearby. The prostitution variables

(20)

20

are replaced by variables containing the number of windows to see whether there is a difference in effect. This part of the analysis will be done in the final section of the thesis. The municipality did not provide the number of windows for all properties. Of the properties used in this analysis for only three property the number of windows is missing.

The resulting regression coefficients from specification (2) and (3) will also show the effects of the closing of the properties on the crime variables. The differences in coefficients of the crime variables are analyzed to see whether there is a significant change as a consequence of including the prostitution variables.

Another problem could be the assumption of inelastic housing supply. If the assumption is incorrect, consumer preferences will partially show up in quantity, not in demand. However, since the analysis is done for the more central parts of Amsterdam, the housing supply can be assumed inelastic because there is not much room for new houses. The problem of multicollinearity could also arise with this analysis. In the case of the crime types it can even be expected, but this can be solved by doing separate regressions for the different type of crimes. For the other variables the multicollinearity will be tested, after which one of the variables that show multicollinearity will be removed.

Only sold houses are considered, which is a common problem in hedonic regressions. This could cause sample selection bias. In unfavorable market conditions the houses that are sold could represent the nicest properties or the properties where the owners were willing to accept the largest discount. However, it is difficult to deal with this problem as the number of houses sold in Amsterdam is not large enough to be able to conduct this analysis using only repeat sales.

(21)

21

4. Data and descriptive statistics

4.1 Data

The data consists of housing data, crime data, prostitution property information and data regarding neighborhood composition and facilities. The housing data is obtained from the NVM ( Dutch association of realtors). The data contains information about 296,850 transactions from 1975-2014, and includes important house specific information as for example the transaction price, house size, the number of bathrooms. The dataset also contains the date, a variable related to the type and detailed type of the house. For the more important variables summary statistics can be found in table 1 on the next page. 296,850 is the number of transactions that were left after removing outliers from the dataset. Transactions with a price higher than 3,000,000 or lower than 5000 were removed from the dataset, as well as transactions with a price per square meter of over 10000 or under 100. Transactions with important variables missing were also removed. As the crime variables could only be obtained for the period 2003-2014, the transactions prior to 2003 are dropped from the dataset. Most of the prostitution properties were closed after 2003, so this would not affect the part of the analysis concerning the closing of the prostitution properties. For a large number of transactions, the houses are located in areas that are in the outskirts of the city and crime related data was not consistently available or would take too much effort to reconcile. The reason for this is that for a number of neighborhoods the data was not available for all years or some neighborhoods merged with other neighborhoods, at least for the provision of crime statistics. The remaining number of transactions is 70,533. The crime data was obtained from the department of research and statistics of the municipality of Amsterdam. The crime data was publicly available at the website of the department. As written in the methodology section, the crime data was available at the neighborhood level and for the years 2003-2014. For several different types of crime data was collected. These are burglary, theft, vandalism, public disturbance and incidents of violence. For all these variables the number of reported cases in each of the different neighborhoods was collected for all the available years. The fact that just the reported cases are available is a problem because it can cause a bias in the data. Not every person is equally likely to report a crime, and the characteristics that determine whether a person will report a crime could be related to demographic factors that in turn could be related to the property values. Since the number of reported incidents is the best available data, this problem could not be avoided for this analysis. Data related to the demographic variables such as average income or racial composition were obtained from two different sources. The first is the municipality of Amsterdam and the second

(22)

22

the Central Bureau of Statistics (CBS), who both provide statistics on a neighborhood level. Also obtained from CBS were files containing information about the borders of all neighborhoods. Using this information, the central points of all neighborhoods could be determined, along with the distance of each house to each of the central points.

The prostitution property data is obtained from the municipality of Amsterdam. They provided a list of all properties acquired since the start of the 1012 policy, with the date of the closing. As one can see in table 1 in the appendix, several properties were acquired before the start of the municipal policy. For some properties the date of acquisition is within the sample period, but 11 properties have to be excluded from the analysis because the date is not in the sample period. For these years no neighborhood characteristics or crime statistics were available. Using zoning plans of the municipality, the remaining prostitution properties are obtained.

4.2 Summary Statistics

In the table on the next page, descriptive statistics are shown related to the property characteristics, the crime variables and the demographic variables. A full list of included variables, along with a description, is shown in table 9 in the appendix. It can be seen that for some variables the number of observations is slightly lower than for others, this is because for these variables the data was not available for some years and neighborhoods. The fact that for the variable companies the minimum and maximum are not rounded numbers is due to the fact that these variables are all weighted by the distance to the different central points of the districts. The mean house price in the sample is 296,221.9, with a standard deviation of 203,147.6. The median house price of the sample is 235,500. The average size is 86,68 square meters, with a standard deviation of 44.02. The average year in which the transactions took place is 2008.4, which is an indication that the transactions are roughly equally divided between the two halves of the sample period. Most of the houses in the sample were built before 1945, which is not surprising since most of the houses are in the center of Amsterdam, where most houses were built before the second world war. For the binary variables like central heating or garage the mean can be interpreted as the percentage of houses in the sample that has central heating or a garage. Not surprisingly, not a high percentage of the transacted houses contained a garage. The average number of burglaries per year is 11.99 per 1000 households, with a standard deviation of 4.868. A map with an indication of the number of burglaries per year that each transaction was assigned is shown in figure 2 in the appendix. The high standard deviation is an indication that there are large differences between areas in the incidence of crime, which can also be seen

(23)

23

in the other crime variables. A visual representation of the different areas with an indication of the prices of the transactions can be seen in figure 1 in the appendix.

As described above in the methodology section, each house gets assigned an individual value for several different types of crimes based on the distance to the different centers of the neighborhoods in the center of Amsterdam. In the table above the crime statistics are summarized after weighting the statistics by distance for each transaction. For the analysis of the prostitution properties it is useful to see whether the transactions that occurred within a certain distance of the properties are inherently different than the properties located at a further distance from the prostitution properties. To see the differences between the different subsets, the means of the most important variables for transactions that occurred within 300, 400 and 500 meters are shown in table 2, along with the standard deviations.

(24)

24

*_{The crime variables are estimated in different ways, in this overview the crime variables are per 1000 Households}

Table 1 Full sample Summary statistics

Variable Mean (SD) Median Min Max N

price 296221.9 235500 9999 3000000 70533 (203147.6) M2 86.68 75 26 500 70533 (44.02) year 2008.4 2008 2003 2014 70533 (3.316) rooms 3.230 3 0 15 70533 (1.453) floors 1.408 1 1 3 70533 (0.665) Before 1945 0.647 1 0 1 70533 (0.478) 1945-1970 0.124 0 0 1 70533 (0.330) 1971-1980 0.0256 0 0 1 70533 (0.158) 1981-1990 0.0741 0 0 1 70533 (0.262) After 1991 0.129 0 0 1 70533 (0.336) Days on Market 110.2 62 -7 2139 70533 (137.0) luxury 0.162 0 0 1 70533 (0.369) kitchen 0.728 1 0 2 70533 (0.518) # Toilets 1.099 1 0 2 70533 (0.543) # Bathrooms 0.948 1 0 2 70533 (0.397) Central Heating 0.878 1 0 1 70533 (0.328) balcony 0.562 1 0 1 70533 (0.496) isolation 0.960 1 0 2 70533 (0.772) attic 0.0420 0 0 1 70533 (0.201) terrace 0.127 0 0 1 70533 (0.333) lift 0.143 0 0 1 70533 (0.350) garage 0.0604 0 0 1 70533 (0.238) garden 0.228 0 0 1 70533 (0.420) border 0.00206 0 0 1 70533 (0.0453) Burglary* 11.99 11.35 2.637 97.65 70533 (4.868) Theft* 152.1 108.44 24.66 2499.6 70533 (148.5) Violence* 100.8 89.79 13.02 565.9 70533 (52.85) Vandalism* 15.86 14.32 2.057 69.42 70533 (7.078) Public disturbance* 18.01 17.04 2.747 91.13 70533 (7.378) Avg. Income 20.17 18.30 5.379 47.79 59961 (7.052)

Social Benefits(per 1000 households) 83.48 79.44 5.052 219.0 59961

(35.27)

Percent 25-45 0.430 0.43 0.284 0.563 48630

(0.0376)

Companies(per km°) 1449.2 1406.57 43.51 4209.7 70533

(783.7)

Percent Non-western immigrant 0.241 0.2015 0.00332 0.805 69585

(0.132)

Percent Western immigrant 0.181 0.1849 0.0595 0.300 69585

(0.0438)

Percent Native 0.556 0.5926 0.137 0.747 69585

(25)

25

As can be seen in table 2, there are significant differences between the full sample and the houses located within 500, 400 or 300 meters of the prostitution properties. The mean house price is 119,167.8 higher for transactions within 300 meter of the closed prostitution properties than for the full sample. For clarity, the subsets in the table above consist of all houses within a certain distance of one of the prostitution properties that was closed. The date at which the property was closed and the date at which the transaction of the house occurred were neglected in this summary of the data. The goal was to see whether a simple regression might show a difference in inherent characteristics instead of a genuine effect. The differences in characteristics can be explained by the fact that the closed prostitution properties are located primarily in the center of Amsterdam, where the house prices are on average higher. Figure 1 in the appendix shows the locations of the closed prostitution properties. It can be easily seen that the dots are located in the more center districts. The houses located more closely to the prostitution properties are also larger, and the average income is higher. The average income does not increase with a lower distance to the properties, but as the properties are already at a central location at 500 meters it is not a surprise that the impact of this small decrease in distance is limited. For the crime variables the different sets also have large differences in value compared to the full sample. This is not very surprising as crimes such as theft and vandalism are more likely to occur in the more central districts.

Table 2 Summary Statistics for subsets

Full sample Houses<500m Houses<400m Houses<300m

Variable Mean (SD) Mean (SD) Mean (SD) Mean (SD)

price 296221.9 402664.4 414221.6 415389.7 (203147.6) (244298.3) (250410.3) (251695.9) year 2008.4 2008.0 2008.0 2008.0 (3.316) (3.313) (3.332) (3.322) M2 86.68 99.74 103.7 104.3 (44.02) (56.87) (59.44) (58.87) rooms 3.230 3.103 3.159 3.135 (1.453) (1.673) (1.722) (1.728) floors 1.408 1.603 1.629 1.627 (0.665) (0.791) (0.799) (0.798) Burglary 15.37 18.87 20.38 21.36 (56.11) (10.43) (11.42) (12.16) Theft 167.9 639.5 752.1 839.4 (296.5) (402.5) (405.5) (402.9) Violence 100.8 247.8 273.0 292.3 (52.85) (90.28) (90.02) (88.38) Vandalism 18.45 27.71 30.41 32.47 (43.16) (11.12) (11.18) (11.09) Public Disturbance 21.82 36.45 41.06 44.52 (63.52) (16.37) (16.32) (16.02)

(26)

26 Table 2 (continued) Avg. Income 16.72 18.68 18.51 18.28 (9.632) (9.677) (9.595) (9.478) Social benefits 65.40 50.65 48.39 46.76 (41.58) (28.07) (27.33) (26.36) Percent 25-45 0.350 0.391 0.397 0.402 (0.163) (0.164) (0.167) (0.168) Observations 70533 2766 2054 1636

The variables related to the prostitution properties are defined in several different forms, the analysis will show what is the correct specification. In Table 3 there are summary statistics of the several variables. One important thing to note is the difference between closed prostitution properties (PP in the table below) and prostitution properties closed. The first indicates that the variable relates to the distance to the closed prostitution variables and does not take into account the date of the closing. They are merely indications of the number of closed properties that are within a certain range of the house. The second indicates that the properties were closed within a certain time period before the transaction.

(27)

27

Table 3 Summary Statistics Prostitution Property Variables

Variable Mean SD Min Max N

# Of PP closed within 1,5y before transaction and between 500m and 350m

0.0449 0.617 0 21 70533

0.0160 0.295 0 14 70533

0.0148 0.303 0 16 70533

# Of PP closed within 1,5y before transaction and within 100m

0.00573 0.156 0 12 70533

# Of closed PP between 300m and 200m 0.157 1.327 0 29 70533

# Of closed PP within 500m 0.829 5.402 0 61 70533

Closest PP closed within 1,5y before transaction (Distance in meters)

2998.7 1407.8 28.40 7442.5 70533

# of current properties within 300 m 0.301 2.667 0 52 70533

Closest current PP (Distance in meters) 2503.4 1390.5 3.704 6690.7 70533

# of houses with a PP closed within 1,5 y prior to the transaction and between 300m and 200m

0.0468 0.0682 0 1 70533

# of houses with a PP closed within 1,5 y prior to the transaction and between 200m and 100m

0.0414 0.0642 0 1 70533

In the table above one can see that only a small percentage of the transactions in the sample concerned a house that was located in the vicinity of the prostitution properties. The bottom two dummy variables indicate that only for 4.68 percent of the sample has a prostitution property closed within 1,5 years prior to the transaction and between 300m. and 200m.. Between200m and 100m the percentage is only 4.14. This limits the scope of the analysis in the sense that it limits the sample that is available for the part of the analysis related to prostitution. While the analysis can still be supplemented with the analysis of the prostitution properties, it is unlikely that the coefficients of the crime variables will differ very much when the prostitution-related variables are included. If any change, it would probably be related to the number of current properties instead of the closed ones. Besides this, one can see in the table that the closest current prostitution property is located 3.7m. from a transacted house. This distance is 28.40 m. for the properties that were closed, taking into account the date at which the properties were closed as well as the transaction date.

(28)

28

5. Results

In this section the initial results of the regressions will be discussed. The checks for robustness and specification checks will be done in the next section. Before the results from the hedonic regressions are discussed, the correlation matrix between the independent variables will be

analyzed. The highest correlation, which could be expected, is between the size in m2 and the

number of rooms. Since these variables are not of essential importance, except to control for differences between properties, they will both be left in the regression. The number of rooms is an important control variable and excluding it will not benefit the analysis. Average income is strongly negative related (-0.59) to Social benefits, which is also a logical correlation. One of these will be chosen based on the explanatory power of the coefficient, average income seems like a more important control variable than social benefits. Some of the type dummies are highly correlated with size, rooms, floors and/or garden, but as the type dummies are purely to control for differences in type this is deemed of no importance. Other high correlations are the number of baths with the number of toilets and the age dummy indicating a building date before 1945 is negatively correlated with lift. The other correlations are considered to have little impact on the analysis. What we have not discussed yet are the correlations between the crime variables. As was expected, some of the crime variables are either highly correlated with one another or highly correlated with some of the prostitution variables. Especially theft, violence and vandalism are correlated with each other, with the former two also highly correlated to disturbance. These three variables are also strongly correlated with the prostitution property variables, while for burglary this is not the case. The strong correlation between the crime variables indicates that it will indeed be wiser to choose one type of crime for each regression. The strong correlation between the prostitution property variables and the crime variables does not need to have a similar consequence, as it is part of the relation we would like to test. Given the high values for some of these correlations, it is wise to test for multicollinearity in our regression results.

Regression formula (1) is run with number of reported burglaries per 1000 households chosen as crime variable. Some of the detailed type dummies will be excluded because the number of observations is too low. Type 4, detached houses, is also excluded because the number of observations is too low. After running the regression, the variance inflation factor (Vif from now on) is calculated. The results indicate that there is some degree of multicollinearity for most variables and a somewhat larger degree of multicollinearity for some variables. The multicollinearity is related to categorical variables, which is logical. If a certain categorical

(29)

29

variable takes on one value, it is necessarily zero for the others. The vif should be highest for the categories with the most observations, which is the case for the variables in this analysis. The only non-categorical variable showing a high degree of multicollinearity is social benefits, which showed high correlation with average income. Social benefits will be excluded but the rest of the multicollinearity can be ingnored. The variable of interest, burglary, has a vif of 1.42. This is very acceptable. Besides testing for multicollinearity, the Breusch-Pagan/Cook-Weisberg test is used to test for heteroskedastic standard errors. The null hypothesis of no heteroscedasticity is rejected so regessions are run with robust standard errors.

The results of the first regression can be seen in column (1) of table 4 on the next page. Most of the resulting coefficients have the expected sign and are highly significant. The R squared is 0.8525 and the F-statistic is 6928.18, indicating that the proposed model fits well and explains a large part of the variance in price. The variable Burglary has a negative sign and a coefficient of -0.0039866. This is contrary to the findings of Gibbons (2003) and Lynch and Rasmussen (2001), but it does make economic sense. Homebuyers do not like burglary so a higher incidence of burglary should have a negative effect on property values. Since this model does not contain time dummies or locational variables, the size of the coefficients are of little importance. Only balcony and percent aged between 45 and 64 are not significant at the ten percent level. The rest of the variables is significant at the one percent level and most at the 0.1 percent level. The negative sign of the variable percent native is surprising. One would expect that a higher percentage of native residents would be positive for the house prices, since many immigrants are working class immigrants or are descendants of working class immigrants. The explanation is in the percent western and non-western immigrants coefficients. Percent western immigrants has a highly positive effect and percent nonwestern immigrants a strong negative effect. The negative sign of the variable border is also surprising, because it suggests that a corner house is less desired than a regular house.

Including neighborhood dummies included in regression (2) seems to lead to some problems. The coefficients all have t-statistics of 0.00. This is probably due to the fact that some of the variables only have positive values for certain neighborhoods. For instance, certain property types might only be present in certain neighborhoods. Instead of neighborhood dummies, district dummies are used in the second regression. Some of the districts contain too little values to be included in the regression, so they are removed. In the second column of table 19 on the two page above the regression results of formula (2), including year dummies and district dummies, are shown.