UNIVERSITEIT VAN AMSTERDAM
Effect of refugee
accommodations on
house prices in the
Netherlands
Master Thesis
Boris Velthuijs- 2 - Abstract
This analysis on the effects of refugee accommodations on house prices yields results that imply a refugee centre is associated with a decreased selling price of properties under certain conditions. Using a rich dataset on housing characteristic and the locations of current and past refugee centres in a hedonic difference-in-difference model it is shown that in particular in rural areas a negative relationship is present. Also, although this analysis has shortcomings concerning control variables and data availability, temporal refugee accommodations are not associated with a decrease in selling price only permanent AZCs are estimated to have significant impact. The effects are somewhat robust when different time frames are used. The sources of the effects estimated seem to be of intangible nature namely the decreased perceived safety and increased prejudice towards refugees, however further analysis on the sources is necessary since this is not covered in this study.
Statement of originality
I, and solely I, Boris Velthuijs, student at the University of Amsterdam at the faculty Business Economics following the Real Estate Finace track, am responsible for the content of this analysis. All inferences made are in light of doing sounds academic research. For further reaffirmation concerning the originality of the content of this analysis, on request, a written statement can be issued as a predicate of originality.
Boris Velthuijs 10850899
- 3 -
Table of content
1. Introduction ... 4 2. Literature review ... 5 -2.1 Hedonic modelling ... - 5 - 2.2 Externalities ... - 6 - 2.2.1 Nuisance ... - 7 - 2.2.2 Criminality ... - 9 -2.2.3 Perceived safety and prejudice ... - 10 -
2.2.4 Other possible externalities ... - 12 -
2.3 Hypothesis ... - 13 -
3. Data ... 14
-3.1 Data preparation... - 15 -
3.2. Descriptive statistics ... - 18 -
4. Methodology ... 20
-4.1 A standard hedonic model ... - 20 -
4.2 Hedonic model in experimental setting ... - 21 -
4.3 Potential validity issues ... - 25 -
5. Empirical results ... 27
-5.1 A standard hedonic model ... - 27 -
5.2 Hedonic model in experimental setting ... - 34 -
5.3 Regression results over distance ... - 36 -
6. Robustness... 39
-6.2 Robustness over time ... - 43 -
6.3 Robust results summarized ... - 47 -
7. Conclusion... 48
8. Ethics ... 50
9. Bibliography ... 51
10. Appendix ... 54
-A) Assumptions of models used ... - 54 -
- 4 -
1. Introduction
Ongoing conflicts and inhumanities in too many areas in the world force the population of countries in which these conflicts occur to abandon house and land trying to reach a safer place for themselves and their families at least for the time the conflict lasts. Examples of these conflicts are unfortunately plenty, but in particular the humanitarian crisis following the Arabic Spring is one of significant impact on European countries. Refugees who try to seek peace travel to Europe despite the great risks involved in the journey and the uncertainty of shelter in the desired country. The Netherlands is confronted with some of the largest streams of refugees in recent history. State Secretary Klaas Dijkhof estimated the total amount of refugees will be between 50.000 and 60.000 people (Borst, 2015), the largest number in more than 20 year (COA, 2016). The Dutch COA is responsible for the shelter of these people. In 2015 almost 50.000 people are accommodated in asylum seeking centers, from now on referred to as AZC’s, almost the double amount of the year 2014 (COA, 2016). Numerous provocative pro and con demonstrations show these AZC’s are a great point of discussion in the Netherlands because they might affect the quality of the neighborhood the AZC is located in. Common heard complains concern the safety, the pressure on amenities, like schools and medical care, nuisance and potentially the cultural differences of the newcomers in the neighborhood. If indeed these complaints are well grounded, it can affect the neighborhood in various ways. One of these effects could be a decreased willingness to pay and therefore a lower transaction price in the vicinity of a refugee accommodation. If an AZC has significant negative effects on house prices this might lead to financial claims by the owner of the residences in the neighborhood of the AZC. Municipalities and the COA, who are responsible for approving locations of an AZC, need to be aware of a possible effect of an AZC on house prices for legal purposes. Therefore it is important to test and possibly quantify the effect of an AZC on house prices. This research extends the literature concerning externalities affecting house prices and can also be useful to appraisers.
Research question
Does providing shelter for refugees in the form of (temporal) AZC’s affect house prices in the vicinity?
The data used in this analysis is a very rich dataset on housing characteristics in the Netherlands provided by the NVM over the period 1990 up and until 2015. In combination with data from the COA a hedonic difference-in-difference method can be applied to the data. The results indicate that no significant effect is present over the whole dataset, subsets however show different estimations. The most striking results are obtained when comparing rural to non-rural areas and permanent to temporal shelters. Property prices in rural areas in general decrease if an AZC is nearby and only a permanent shelter has a significant impact on house prices. The most important implication is, since the effect seems to be of intangible nature, to inform inhabitants of a neighborhood better and to include them in the process of refugees integrating in the Dutch society.
The remainder of this study if structured as follows: a literature review, section 2, will present the potential channels by which a refugee accommodations can influence its
- 5 -
surroundings, the data and preparation of the data is discussed in section 3, methodology and the validity of the models used are presented in section 4, section 5 discusses the main results obtained, a series of robustness checks will be performed in section 6, section 7 concludes the analysis and section 8 provides some insight in the ethical aspects of this study.
2. Literature review
2.1 Hedonic modelling
Hedonic models assessing house prices are used for many purposes. In the Netherlands for example hedonic models are used by the government to estimate a certain value of a residential property for taxation goals. Others apply the model to estimate insurance or market values of portfolios whereas some use its analytic power to analyse policy changes, events or externalities. Based on the principle of comparable sales, hedonic techniques approach a property as a basket of characteristics with each characteristic adding its own value. If sufficient comparable properties and corresponding characteristics are gathered regression analysis can be applied to quantify what each characteristic adds to the total sum of value. Usually the dependent variable is the transaction price of a house, or a transformation of it, and the independent variables are determinants of the price. The independent variables can be summarized by the following categories (Francke, 2014):
Legal rights, which ownership rights are involved in the transaction.
Conditions on sale, under what circumstances is someone selling or buying.
Market conditions, general conditions at time of sale like demographics, economic activity in a country or relatively low oil prices.
Location, the (near) immediate surroundings of a property.
Physical characteristics, the property itself including lot.
A further specification of location is possible by categorizing based on different location aspects. Physical, socio-cultural/socio-economical and functional environmental aspects are formulated in a thorough investigation on living environments, locations, in the Netherlands (Visser, 2006). Physical environment is about all physical aspects of the location like green spaces, ponds, quality of neighbouring buildings and density. The cultural and socio-economic surroundings concern the composition of a location in terms of culture, educational level or unemployment. Functional environment is mainly about reach ability of amenities and infrastructure. The study showed that when adding physical characteristics and provinces, a proxy for regional effects, to the physical, sociological and functional environments of a property, only about 25% and 30% of the variance in transaction price per square meter is explained by physical characteristics and regional effects for apartments and houses respectively. This was less than expected by the authors. The location linked factors explain 25% and 38% of the variation for apartments and houses respectively (Visser, 2006). The importance of aspects like green surroundings, perceived social status of a neighbourhood and reach ability in the Dutch residential housing and apartment markets is again underlined. It is shown that ‘softer’ measurements like perceived social status or
- 6 -
cultural diversity are too important explanatory variables. In the context of this analysis especially these location characteristics are important since the placement of an AZC can possibly influence its surrounding environment.
If an AZC influences its surroundings, the neighbourhood and its residences are affected. If these effects generate a welfare loss or gain and the people in the neighbourhoods that are affected do not receive compensation from the authority responsible for this welfare effect this is called an externality, a not compensated welfare change resulting from economic activity or a policy change. The placement of an AZC is on itself not an externality, the effects of an operational AZC however can be externalities. The channels, associated with the placement of an AZC, by which an AZC can possible produce externalities will be discussed in the subsequent section.
2.2 Externalities
Literature concerning the effect of externalities on house prices is abundant. Examples are effect of waste transfer stations (Eshet, 2006), noise pollution (Dekkers, 2009) and (Theebe, 2004), traffic (Sirmans, 1992), agriculture (Kuethe, 2012), wind turbines (Droes, 2014) and golf courses (Quang Do, 1995). Many of these externalities are also phrased NIMBY’s, not in my backyard, expressing a negative attitude towards an externality. Take wind turbines for example, some people might not want a wind turbine adjacent to their backyard for practical reasons but are nonetheless in favour of green power. It is an externality, expressing the unintended impact of economic activity on house prices, if not compensated for this impact. However a wind turbine is only a NIMBY if people do not want it to be next to their garden, which can make sense if the view gets blocked by it for example. Though some people might prefer a wind turbine since it allows people to get closer to a self-sustainable, environment neutral, green and well-balanced life. A golf course can also be a NIMBY if a property is located near hole 3 and it is necessary to replace the back windows every Sunday afternoon. On the other hand, golf courses tend to be beautiful green areas with lovely scenery including ponds and trees suggesting it is a perfect extension of the surface of the lot or garden. The golf course is an externality for sure, it has its unintentional negative or positive effect on house prices if the possible effects on house prices are not compensated, but the course is in general only a NIMBY if the balls smash the windows, enthusiastic golf ball collectors aside.
Being an externality, positive of negative, does not imply people do not want it near them. The same goes for the location of an AZC, people might be in favour of supporting war torn refugees by sheltering them, taking possible externalities for granted, while some preferably see this happening near the Greek border or not at all. Over the period of 1997-1999, in the Dutch Randstad region, the effect of an AZC on house prices is investigated by examining 55 AZC’s and 113,574 transactions yielding the conclusion that an AZC is not a NIMBY (Theebe, 2002) (Çankaya, 2002). This conclusion is drawn based on the facts no significant effects on house prices are measured and no differences are present in time on market between properties located near an AZC and not. In the light of the recent influx of refugees and the heated protests when an AZC announcement is made, the matter is ready for a new analysis.
- 7 -
Anyway, the opening of an AZC can have impact on the neighbourhood by multiple possible externalities that are associated with the placement of an AZC. To assess the externalities associated with an AZC, risk analysis reports ordered by two Dutch municipalities, Breda and Zutphen are used as an indication of what might change in the living environment when an AZC is up and running. Inhabitants ex ante are in general apprehensive of their perceived safety, nuisance and criminality (J. Kuppens, H. Ferwerda, 2016). Safety concerns perceived safety of current inhabitants. Nuisance is quite broad but examples are: traffic, pollution, noise, drugs usage and herding. Criminality covers crimes like burglary, dealing drugs, threatening, and theft. Perceived safety is a ‘soft’ measure compared to nuisance and criminality since the latter can be measured more objectively. A similar report on the AZC in Breda provided a conclusion based on a seven month analysis and can be summarized as follows: garbage and noise complaints rose a bit, traffic situation became more hazardous, the perceived safety did not change that much according to the inhabitants. Although a few stated that (drug related) crimes and street vending, mobile phones, were more prominent after the AZC became active (Gemeente Breda, 2015). The subsequent section describes the possible channels by which an AZC and its inhabitants can affect the neighbourhood.
2.2.1 Nuisance
Nuisance is a broad term and captures most things that are not heavily prosecuted but are annoying to residents if confronted with. Those things that are heavily prosecuted will be discussed in the criminality section. An overview of nuisances that are related to an AZC is presented.
Noise can be a horrible disturbance of a peaceful night sleep resulting in, if frequent disturbances occur, either sleeping with earplugs or some other possibly expensive noise reduction measure. If adjacent to a highway, railroad or airfield it is shown that these properties, if exposed to high levels of noise ranging from 65 decibel and louder, are priced at a discount compared to more quite locations. Properties located in a quite environment, lower than 40 decibel, are even priced at a small premium (Theebe, 2004). This study examined over 160,000 properties in the western part of the Netherlands also referred to as the Randstad over the period 1997-1999. Another Dutch studies performed in the geographical area surrounding the Dutch main airport Schiphol analyzed 67,000 transactions in the 1999-2003 period. The results are in line with the study performed by Theebe. A one unit increase in decibel above a certain threshold, 45, 60 and 55 decibel for aircrafts, trains and cars respectively yield a significant drop in the transaction price by 0.8%, 0.72% and 0.014% respectively. Interestingly this also works the other way around, under certain conditions the authors find a reduction of 1 decibel in general leads to an increase in total house value of €574 million in the region (Dekkers, 2009). One result from the risk analysis performed by the municipality of Breda is a rising level of noise in the surroundings of an AZC. Logica l since a large concentration of people 400 in Breda, interact and produce sound. Especially in the summer more people will spend their time outside and therefore the noise disturbance is bigger in this season (Gemeente Breda, 2015). The noise levels cut-offs used in the mentioned studies range from 45-65 decibel. A person yelling produces about 90 decibel and
- 8 -
a normal human voice talking about 60 decibel (Alpine, 2016). So if all the people living in an AZC shout when they talk and produce this sound constantly it is a serious issue for the residential values surrounding the AZC location. However the report from the municipality Breda mentioned complaints of people talking and children playing, indicating minor to moderate disturbance. It is hard to compare people’s voices to aircrafts since it can be reasoned that the aircraft is more disturbing because it is a more unpleasant sound than children playing, given they produce the same decibels. On top of that, the noise from an AZC is usually produced in daytime or in the evening, (air) traffic sounds are more constant during the day implying more nuisance. Altogether, the sounds refugees produce can be disturbing to the neighbourhood but may be less intense than an aircraft passing by, therefore a minor negative effect on house prices is possible.
Traffic surely can affect an environment by means of noise, pollution and safety. Evidence that traffic is priced into houses is presented in a study performed in Baton Rouge Louisiana. The study showed that properties located high traffic streets, compared to low traffic streets, are priced at a discount. The busier the road, the higher is the discount (Sirmans, 1992). Since refugees fled their home country abandoning their home and possessions a traffic increase surrounding an AZC will not be because of the refugees. Therefore the sources noise and pollution of traffic nuisance can be eliminated. Safety however is applicable in the context of an AZC. Several news items mentioned the lack of knowledge about traffic rules by refugees as well as their capacity to ride a bike safely, the latter is especially important in the Netherlands, a leading bike nation. The traffic safety issue is confirmed by investigations by the municipalities of Breda and Zutphen. If hazardous situations occur more frequently around an AZC location this can be a significant source of nuisance and it is possible to image parents with children would rather live a block or 2 away. The willingness to pay for a property located near a dangerous road or junction can be lower. It is not factually analyzed that roads or intersections around refugee shelters are more dangerous but the reports and news bulletins can be an indication for it. But since this is not factualized for all shelter locations, locations differ very substantially in context of infrastructure, traffic safety is not expected to have an impact on house prices in the vicinity.
Pollution, defined as rubbish in streets and parks, is not the nicest thing to see when viewing a property. Also if not cleaned, parks and streets might start to smell if consistent pollution is present increasing the nuisance from pollution. Green parks, ponds or forests are shown to add value to a residential property located nearby. In the category physical environment, in a very extended analysis on Dutch house price determinants, the green aspects of a neighbourhood add most value to a house price (Visser, 2006). This is consistent with a case study performed in Besançon, France. Visible trees for example add almost 2% in value in this analysis (Baranzini et al., 2008). If these attributes to an environment are polluted they do not longer add the same value as if neat. Given the importance of non-polluted green areas, it is possible that consistent pollution from refugees affect the house prices indirectly. The municipality of Breda investigated these concerns and concluded that no degradation of the public space was present 7 months after the opening (Gemeente Breda, 2015). Since no indications are present the surroundings of all the AZC’s in the Netherlands
- 9 -
are extra polluted by the inhabitants of it, the effect of refugees on pollution of a town can be regarded as minimal.
2.2.2 Criminality
Crime is defined here as a offending the law thereby possibly face severe punishments, theft, burglary, violence of vandalism are examples. Crime is part of the location and neighbourhood fraction in explaining the values of properties in hedonic modelling. In an extensive literature analysis performed in 2003 by the National Association of Realtors, the biggest real estate union of the United States, crime is among the top 30 used explanatory variables in the reviewed hedonic models and in the top 5 of most frequently used environmental neighbourhood and location factors (Macpherson, 2003). In most cases the crime variable had a significant negative effect on house prices, sometimes it was recorded as insignificant and only once crime had a positive effect. The Dutch NYFER, a research institute performing applied scientific research, was commissioned by the municipality of Utrecht to evaluate the value of safety in Utrecht in the context of house prices. Based on hedonic modelling and assumptions about the interpretation of the coefficients it is concluded that if in Utrecht no violence or related crimes are present the value of this ‘violence free’ city is estimated to be €1,3 billion, or a €13,000.- per property discount is present because of violence. An increase of 1 incident per 10.000 inhabitants decreases the value of a residential property by €1,087.50 in general (NYFER, 2004). Stockholm, Sweden was subject of a study on the impact of crime on apartments in 2008 incorporating 9,622 transactions in different neighbourhoods. The results indicate a 1% increase in the total crimes results in a drop of 0.04% of apartment values (Wilhemsson, 2011).
Critical notes concerning the impact of crime on house values are formulated by Mark Cohen, he indicates several sources by which the estimated effect of crime on house prices, or estimated discount on willingness to pay, for a property can get overstated. First, individual neighbourhood data on crime lacks or is not present to researchers and therefore they are forced to use an index of crime to estimate the effect. The use of an index makes it harder to separate effects of different crimes which might be applicable or useful in a certain area of interest if for example in a certain neighbourhood certain crimes are over/under represented in the index. Also the use of which kind of index is essential in estimating effects. Total crime indices produce different results than violent crime or property crime indices if used. Second, most studies on crime’s impact on house prices fail to incorporate specific characteristics of a neighbourhood possibly overestimating the effect of crime on properties. If for example these studies do not control for the amount of noise, pollution or reach ability, the effect of crime can get overstated if more criminal district are in general located in more polluted, suburban or noisy areas (Cohen, 1990). The last point might also concern the placement on an AZC and is important to keep in mind when interpreting the estimated effect of an AZC on house prices.
Even if estimates of the effect of crime on house prices are possibly overstated, it can be a significant externality when an AZC is placed. It is a great concern and point of discussion when an AZC location is announced and needs to be recognized as a possibly serious effect. At this point, however, no hard facts are yet presented about the increase/decrease of crimes
- 10 -
surrounding an AZC. The municipality of Breda showed in a report including the actual number of police reports that the number of incidents dropped since the AZC was placed. Noticeable since across all categories, ranging from high impact crimes to theft of conveyance, the drop after the opening of the AZC is present (Gemeente Breda, 2015). It should be mentioned that the COA hired private security and the police presence was increased by the municipality in the period after the opening implying the area was potentially safer from this point on.
Increased crimes can deter a neighbourhood, influencing the willingness to pay to live in that area, but since no ‘hard’ and fact based increase is detected, it is too early to conclude what kind of effect an AZC has on criminality in its environment. Keeping the possibility open this externality has its effect, at this point an AZC opening is not assumed to cause crime rates to change. Contrary, it might even spur the crime rates to drop since increased police presence is potentially associated with an AZC opening.
2.2.3 Perceived safety and prejudice
Next to the ‘hard’ measures of criminality, softer and more intangible measures like neighbourhood reputation or perceived safety are also shown to have significant impact on residential housing markets. In a study on effects of safety and perceived safety 50 municipalities in the Netherlands are investigated by NYFER, a research institute performing applied scientific research. Two findings are interesting. First, a strong and significant correlation exists between perceived safety and number of reported violent incidents exists, 40.50% of the perceived safety can be explained by the number of reported violent crimes. This analysis relates objectivity, official reported crimes, and subjectivity, the perceived safety, to house prices. Second, perceived safety or feeling of safety is strongly significant as an explanatory variable on house prices, controlled for a broad range of other determinants, it has a 99% significant negative impact on house prices (NYFER, 2004). A study performed in Utrecht in 2006, using the results of 1,400 respondents who filled in a questionnaire with the researcher in person, showed, using logistic regression techniques, that perceived neighbourhood reputation had a significant negative impact on people’s intention to move implying an increase in perceived reputation decreased the willingness to move and vice versa (Permentier, 2009). People incorporating perceived reputations or safety in effect incorporate personal feelings and emotions in their decision to buy a property. The willingness to pay for properties located in perceived ‘undesired’ locations is expected to be lower than other properties.
Another intangible and ‘softer’ determinant of house prices comes in the form of prejudice. Prejudice is defined as: ‘a feeling or an attitude toward a group or individual’. Prejudice is present if house prices are influenced by a change in the percentage non-whites in the neighbourhoods (Kiel, 1996). Developments of prejudice in the cities of Chicago, Denver and Philadelphia were investigated over the years 1978-1990, by combining data of the American Housing Survey and data on neighbourhood quality. The results indicate the presence of discrimination and prejudice, but their presence and effect is in general declining over the investigated period (Kiel, 1996). If the definition of the presence of prejudice, prejudice being present if house prices are influenced by a change in non-white population, is
- 11 -
applied in a Dutch context, prejudice seems to be present in the Netherlands as well. The already mentioned extensive analysis on determinants of house prices in the Netherlands performed by Visser and van Dam incorporates the percentage of non-western people as an explanatory variable. The effect of a 1% increase in non-western inhabitants results in drop in house prices of about €430.- per square meter and is significant at the 95% confidence level (Visser, 2006). Big advantage in this study is that these results are controlled for a very broad range of physical, social and functional environmental variables next to the physical characteristics of a property strengthening the estimated results.
Also intangible and linked to prejudice and a possible channel of an AZC influence on house prices is the channel of attitude towards different cultural norms or religions. As mentioned in the introduction the debate about shelters and refugees is characterized by overheated demonstrations of the pro and the con sentiments. Refugees are incorrectly related to sex offending, criminality and being members of an armed force by people who lack wisdom and empathy for war victims. Sentiments like these can play a role in the Dutch housing market because market participant might not be willing to pay as much for a residence located near a refugee shelter and thereby influence the transaction price downwards. It is not the first time a sentiment towards an ethnic group influenced the real estate market in the Netherlands, although temporarily. The study of the effect of Theo van Gogh’s murder by a fundamentalist in 2004 on the Amsterdam housing market showed a significant decline in list prices of houses located in a neighbourhood with more than 25% o f Muslim background compared to neighbourhoods comprising less than 25% Muslim (Gautier, 2009). The results imply attitude or certain emotions played a role in the Amsterdam housing market after the event. If market participants are prone to the sentiments stated, refugee accommodations might affect house prices. However the active players in the housing market are usually higher educated and this sub-group is associated with a more tolerant and less xenophobic attitude towards refugees (Adam, 2015). The general perspective of refugees on same sex marriage though might be hard to cope with for some (potential) home owners in a tolerant country as The Netherlands. The prejudice of refugees in this case is a source of concern for potential homeowners possibly lowering their willingness to pay for a property located near an AZC.
Perceived safety, prejudice and attitude are show to be important in explaining property prices. Prejudice and attitude are closely linked, the effect of the van Gogh murder shows that people who consider buying a property can have certain pre-defined opinions about a part of the population and these opinions can be strengthened in the aftermath of a crisis situation like a radical murderer. At this moment we are in the middle of such a crisis situation namely the widespread use of violence in a false religious flag operation in the Middle East. Media provide abundant coverage of how horrifying the Middle East can be and people might attach a negative feeling towards inhabitants of the Middle East. People might also incorrectly perceive refugees as being an unsafe part of the population especially since Europe at this moment is struggling with terrorist attacks further deterring the reputation of the Middle East population. The public image of refugees drops per news item relating refugees to violence, sexual assault or crime. A recent example of the public image of refugees being disgraced is
- 12 -
the New Year’s evening in Keulen, Germany. Starting fresh in the new year, immediately the media dived onto the incident that happened during the night before. Hundreds of women were sexually assaulted, over 800 people got robbed, violence was used against police officers and heavy firework shot all over the place by what the media was referring to as men with an Arabic or north African origin, including registered refugees (NU.nl, 2016). Immediately the public image was formed and reinforced by European media. After a thorough analysis however it turned out, month(s) later, that predominantly illegal Algerian and Moroccan men were responsible for the chaos in Keulen. The most influential public image provided in the media concerning refugees turned out to be wrong and the rectification of course did not create the mass hysteria in Europe as it was at January 1st 2016. This stigmatization of refugees can contribute to the prejudice towards refugees. If people indeed view immigrants as more dangerous or unwanted in a neighbourhood, without further empirical research to back their perception, an AZC placement can have a negative impact on house prices simply because market participants themselves perceive it as problematic.
2.2.4 Other possible externalities
Other possible unintended effects of the placement of an AZC on neighbourhoods and country that are not formulated in the context of risks associated with the placement of an AZC are discussed below.
Pressure amenities
If refugees receive education, especially kids, this might put pressure on schools or day-care facilities in the neighbourhood. It seems logical that a reduced level of amenities lowers the quality of the neighbourhood, thereby lower the likability of the location and possibly lowering housing prices. Studies on the relationship between house prices and the quality of schools show that high quality schools are priced into the house prices in the neighbourhood (Leigh, 2008) (Abbigail J. Chiodo, 2010). Even though the studied educational systems are hard to compare to the Dutch system, the idea of value adding amenities is clear and pressure on these provisions might result in lower house prices. The already mentioned review of hedonic house prices determinant also showed that school quality is popular in the reviewed studies and that perceived school quality had a consistent significant impact on house prices (Macpherson, 2003). Extra pressure on waist collectors or physicians might have a comparable effect since a drop in a quality of the stated amenities can reduce the value of the neighbourhood. However these effects are not researched yet in the context of an AZC placement and therefore amenities are not expected to be an externality.
Long term demand
The history learns a significant part of the refugees is here to stay, influencing the housing market in the long term. Basic economics tells more demand means a higher price ceteris paribus. A study performed in Spain concluded that in the long run immigrants push up the housing prices by means of population growth (Sanchis-Guarner, 2014). The Netherlands is confronted with the biggest population growth in the first half year since the year 2000 mainly because refugees are granted a permit making them official immigrants. Over 99,000 newcomers arrived and 64,000 left, combined with the natural growth the total population rose by 43,000 in one year (NOS Teletekst, 2016). It is not a proven effect of AZC’s however
- 13 -
the AZC’s facilitate immigration based on international human laws. It is not a confirmed determinant of house prices because most hedonic models are applied locally, but on a national level more people induce higher demand of housing possibly spurring the prices. This effect however is not expected to be of importance in this analysis it just shows a long term perspective.
Local economy
Some small municipalities were quite unhappy when the COA decided to close their refugee accommodation mainly because of financial reasons. Every asylum seeker gets an allowance per week to spend on goods of their choosing. Small towns warmly welcome these spending as a boost to the local economy. Also the municipalities get an allowance based on the number of inhabitants (Althuisius, 2013). With these funds the municipality might invest in amenities and make the place more attractive possibly affecting house prices in a positive way. However it is not expected that people who seek a home in rural areas are willing to pay more for a location that is in the vicinity of an AZC just because additional revenues of a municipality can spur investments in neighbourhoods. Therefore no effects are expected on the residential real estate market.
2.3 Hypothesis
Summarizing the possible externalities associated with the placement of an AZC yields the following channels that are expected to be of influence when an AZC is opened:
Nuisance, in the form of increased noise and pollution
Criminality, in the form of increased perceived criminality or unregistered crimes Perceived safety and prejudice, in the form of decreased perceived safety, increased
prejudice and a more negative attitude
Aim of this study is to analyse and quantify the effect of a placement of an AZC on house prices in the vicinity. All the possible externalities linked to an AZC are bundled in a single treatment indicator. In what way each channel may affects the house prices precisely and what the main determinants of the possible effects are is beyond the scope of this analysis.
Taking all these aspects into consideration the following hypothesis is tested:
A refugee accommodation does not significantly affect house prices in the vicinity
Testable hypothesis in regression: H0: βtreatment = 0
- 14 -
3. Data
The data used to estimate the effects of an AZC on housing prices are house prices, corresponding characteristics and current and past locations of (temporal) refugee accommodations over a time span of 25 years.
Time frame
The current Dutch asylum legislation is based on the Vreemdelingenwet 2000. Proposed in 2000 and enacted in 2001, this law foresees in the rights and procedures an asylum seeker has in the Netherlands. Despite some amendments since its inception, it still provides the guidelines concerning the AZC’s in the Netherlands. Therefore the moment the law was enacted is a sound indication of a starting point regarding the analysis on this topic because the legislation concerning AZC’s should be clear to the housing market participants from this moment on and did not change dramatically ever since. However, to filter out pre-existing trends and to extend the analysis, data before 2001 is preferred. Therefore the timeframe of the analysis is from 01-01-1990 up and until 31-12-2015. Another advantage of this broad period is that the last big refugee crisis, the war in Yugoslavia in 1994, is also included and that situation was fairly comparable to the current one in the sense of a huge increase when the crisis occurred and the heated debates back in those days are remarkably comparable (COA, 2016).
House characteristics should be important determinants of the price of the house like size, lot size the house is built on, number of bathrooms, construction year, type of house, quality of house, maintenance and so on. Of course location is very important since the research concerns effects of being on a specific location. Data on these house characteristics is provided by the Nederlandse Vereniging van Makelaars en Taxateurs NVM, the biggest union of brokers and appraisers in the Netherlands. The aim is to collect data on transaction prices because this provides the best insight in the market value and the market’s possible reaction to an AZC placement.
The COA provides facts of the refugee accommodations on their website. Location and number of people in the accommodation are of prime importance and can be subtracted. Also the type of shelter, temporal or permanent, to who shelter is provided, underage or adult, and the purpose of shelter, asylum procedure started or denied, can be of importance and is provided by the COA. At the time of writing 113 AZC locations are known in the sample period. There are more than the 113 locations mentioned in the whole time frame of the analysis but the COA was not inclined to provide detailed data on the closed AZCs. Very unfortunate because it would strengthen the analysis and provide insights in the 2 most recent Dutch refugee crises.
The CBS Statline database, the Dutch government bureau of statistics, can provide data on location specific characteristics like for example unemployment, demographics or crime rate to control for these factors when estimating the effect of an AZC on house prices. These characteristics can also be covered by the use of time and location or entity fixed effects.
- 15 -
3.1 Data preparation
Before the analysis can take place the data has to be prepared in order to suit the analysis. The NVM data in particular needs preparation before suited for the analysis. The NVM provided a unique raw dataset containing about 1,083,000 observations with 32 different variables per observation. The procedures necessary to estimate a hedonic pricing model based on the NVM data will be in line with the guidelines provided in Casametrie (Frankce, 2010) and in the reader Hedonic Price Models (Francke, 2014) by Marc K. Francke.
To create a sample containing useful observations to analyse the effect of a refugee accommodation on house prices some observations need to be dropped. The aim is to create a sample of properties which are roughly comparable to each other, hedonic pricing models are after all a form of sales comparison techniques. The necessary editing of outliers will be discussed per variable.
Transaction price is the dependant variable in the regression analysis. It is the best reflection of the residential real estate market’s response to the placement or closing of a refugee shelter. To estimate a sound model explaining transaction prices, a transformation of transaction price to its natural logarithm is necessary. Advantages of a transformation are firstly that some relationships between price and for example surface become linear so they can be estimated by ordinary least squares. Secondly, the transformation allows the law of diminishing returns to be implemented in the model. A 1 unit increase in value of, say, volume is priced less when volume is already relatively high. Thirdly the relative standard deviation is used instead of the absolute one implying the weight attributed to outliers is reduced. A very high price has more statistical impact than a small one and a transformation to natural logarithm reduces these relative differences and standard deviations. The fourth reason to use natural logarithms is in line with the third argument namely the error term is closer to normality which is necessary for linear regression analysis. The reason is that outliers impact the outcome, and therefore its error term, less in a natural logarithmic transformation (Francke, 2014).
Lot size is specified in m², the raw dataset of the NVM contains 366,116 observations with lot size equal to 0. The NVM explanation table of the data states an observation of 0 on lot size means the lot size is not known. However, these curious observations can largely, 95,7%, be explained by the fact that these properties are located in flats or on ground/upper floors of spliced houses and therefore lot size is hard to measure. Therefore the decision is made to leave these observation in the dataset. When a natural logarithm transformation has taken place, all the unknown observations, lot size equalling 0, are recorded as a . since ln(0) is mathematically impossible. These observations do not add information about lot size but about other variables and therefore are useful to keep since this analysis is not about a causal interpretation of the hedonic price estimators. To keep these observations in the dataset after the transformation to a natural logarithm it is necessary to replace the dots by zeros otherwise the statistical software Stata does not use these observations. It is important to also include an indicator variable in the regression indicating if the observation was subject to the replacement of dots by zeros when the models are estimated and the natural logarithm of lot
- 16 -
size is used. This is necessary to account for discontinuity in the variable resulting from the mathematically impossible calculation of the natural logarithm of 0.
Outliers in the dataset are the observations with a huge lot compared to the others making these hard to compare to the bulk and are therefore dropped. Decision rule in this case is to drop observations larger than 7,030 m² or the 0.75% highest lot sizes in the raw dataset. Variance, skewness and very importantly kurtosis drop significantly yielding a distribution in line with a normal distribution. Hedonic pricing models in general use OLS regression techniques requiring the distribution of the observations to be approximately normal enabling linear estimations of β parameters (Francke, 2014). Also the transformation of lot size to its natural logarithm is essential in enabling linear estimations. Because at some point an extra m² lot size is not valued as much as it is when lot size is relatively small. Called the law of diminishing returns, this implies a non-linear relationship between transaction price and lot size and that is why a natural logarithm is applied to re-linearize this relationship (Frankce, 2010).
NVM data provides the amount of m² garden. Over 53% of the observations do not possess a garden. In this case it is reasonable to apply the information in a way that suits the data namely by incorporating the garden variable as an indicating variable if garden is present or not. However after checking the regular hedonic pricing model the indicator yielded a negative value and using the natural logarithm of m² garden resulted in a positive estimator, which is in line with economic theory. Therefore the natural logarithmic transformation of m² garden will be used which is also in line with arguments concerning the law of diminishing returns and is more convenient to interpret however the latter is not a prime target in this analysis. Because 53% of the observations do not possess a garden, m² equals to 0, the natural logarithm of 53% of the observation cannot be calculated. After the transformation to natural logarithms has taken place, the observations not possessing a garden will display a . in the dataset for mathematical reasons. These dots need to be replaced by zeros to keep to make sure the data is used. Just like it is the case for ln(lot size), an indicator variable needs to be included in the regression, equalling 1 if a dot is replaced by a zero and 0 otherwise, to take into account the discontinuity in this variable arising from mathematical calculations.
The surface of a residence is measured in m². Unfortunately 50,287 observations possess a surface of an unrealistic high 99,999 m² indicating the surface in unknown. These observations need to be deleted because transformation to the natural logarithmic will not solve the problem, ln(positive number) = positive and will be recorded as a number and not as a . which was the case for lot size. To arrive at a normalised distribution of surface and for comparability reasons, observations lower than 25 m² and higher than 499 m² will be removed from the dataset because these are too small/large compared to the bulk of the dataset. In total, excluding the observations of which no surface is recorded, 0.36% of the observations are dropped in the raw dataset. The surface variable is converted to natural logarithms since the law of diminishing returns is applicable to surface and because the distribution of surface approaches normality if converted, which is necessary for linear regression analysis.
- 17 -
To the variable volume, measured in m3, the same arguments and methods are applied as to the variable surface. The difference in deciding when to drop an observation is a factor 2.5, this is the assumed median height of a ceiling measured in meters. Observations with a volume smaller than 60 m3 and larger than 1,250 m3 are therefore dropped and the volume variable is converted to its natural logarithm. The arguments that back this decision are the law of diminishing returns, comparability and approaching a normal distribution. As was the case with some surface observations, the volume of 42,253 observations is unknown. Excluding the unknown observations, the percentage of raw data dropped is 1.29% by this adjustment.
Transactions recorded as an auction will be removed from the dataset since it is unknown if it concerns a voluntary or a forced sale. If in a transaction the cost of sale is on the account of the buyer and cost of sale is on account of the seller the observations will remain in the dataset, the first is the standard in the Netherlands and the latter probably concerns newly built properties. Only 0.02% of the raw data is lost due this adjustment.
The number of stories varies from 0 to 12 in the NVM dataset. Only properties with 4 stories or lower will remain in the analysis in effect dropping 0.38% of the raw dataset. The comparability of properties with 6 floors and 1 floor is negligible and does not suit the type of regression analysis. Because ideally the aim is to collect a dataset with properties that are almost the same except for the fact that one part is in range of an active AZC and the rest is not thereby allowing a proper estimate of the effect of a refugee shelter on house prices ceteris paribus.
199 rooms in a property that sold for €600,000 seems a bargain just like €262,000 for a 58 room single-family home. These kinds of outliers seem to be data errors and will be removed from the dataset. The number of rooms can be an indication of the luxury of the property, since an extra room does not precisely measure the impact of an extra room on house prices but can be a more general measure of the class which can be useful (Francke, 2014). Number of rooms in this analysis will be limited to 11, comparability argument applies, dropping 0.2% of the raw data.
Total loss of observations attributable to handling outliers equals 7.09% of the raw data or 76,787 observations.
Calculating distance between an object and a refugee accommodation is done in the way ancient civilisations calculated right angels, by the theorem of Pythagoras. X and Y coordinates provided by the NVM on the locations of houses are compared to the coordinates of a refugee accommodation. Calculating the difference between X and Y coordinates of an AZC and an object yields two straight lines on a grid by which the two point are connected. Since the angle these two lines make is 90 , the theorem of Pythagoras is used to calculate the direct distance from an object to the shelter location. For municipalities in which more than one shelter is accommodated the direct distance towards the closest shelter is used in the analysis.
- 18 -
Some of the observations had no data on the coordinates and therefore no distance can be estimated. Also because AZC data on some of the municipalities is missing and the COA was not able to provide this data no distance could be calculated. Therefore observations that had no distance registered after the data editing were dropped. 113.364 observations were lost due to this modification. In total 17.56% of the raw data provided by the NVM is not used resulting in a dataset of 893,013 observations.
3.2. Descriptive statistics
The used variables in the analysis of the NVM dataset will be presented and shortly discussed in Table 1. The number of observations for each of the listed variables is 893,013, this applies to the whole prepared dataset used in this analysis.
Table 1. Descriptive statistics NVM dataset
Variable Mean Std dev. Min Max
Ln(TP) 12.0422 0.5670 9.3365 15.6393 Ln(Lot size) 3.5367 2.5470 0.0000 8.8579 Ln(Garden) 2.0671 2.1430 0.0000 6.9058 Ln(Surface) 4.6644 0.3661 3.2189 6.2106 Ln(Volume) 5.7368 0.4055 4.0943 7.1309 NVM grades 5.3946 2.9880 1.00 10.00 House type 11.6873 8.5834 2.00 27.00 Quality (A) -0.2601 1.0360 -1.00 2.00
Open Porch (A) -0.6368 0.5052 -1.00 1.00
Building period 5.0822 2.3295 1.00 9.00 Elevator 0.0944 0.2924 0.00 1.00 N. floors 2.2510 0.8788 0.00 4.00 N. rooms 4.1863 1.3810 0.00 11.00 Attic 0.2243 0.4171 0.00 1.00 Parking 0.3081 0.4617 0.00 1.00 P. internal 0.0485 0.2148 0.00 1.00 Main. Inside 7.0262 1.1250 1.00 9.00 Main. Outside 7.0667 0.9857 1.00 9.00 Monument 0.0103 0.1012 0.00 1.00 Pool 0.0017 0.0417 0.00 1.00
The natural logarithm of transaction price, the dependent variable in this analysis, has a mean of 12.0422, the untransformed average transaction price of €200,216.-. The transformation of this variable to its natural logarithm ensures that the distribution is more in line with normality. The same holds for ln(lotsize), ln(surface) ln(garden) and ln(volume). Lot size has a mean of 3.5367, untransformed 190 m2, and, as displayed in Table 1 in the column minimal value, there are observation that do not possess a lot. These observations, 366,116 in total, concern primarily flats and therefore they remained in the dataset. The average amount of garden is 2.0671, untransformed 43 m2. Both ln(lotsize) and ln(garden) have quite a high standard deviation when compared to the other transformed variables. This
- 19 -
is because about 41% of the observations do not possess a lot and 53% do not possess a garden, indicator variable equalling 0. NVM grades and house type are indicator variables provided by the NVM. The displayed statistics on indicator variables in Table 1 are not that important since they are hard to interpret. Quality and open porch both have an A in parenthesis because this only applies to apartments. The minimal values of these variables indicate a -1, this corresponds to not being an apartment. Elevators are present in 84,279 observations, an attic in 200,337. Parking space applies to 275,114 of the properties and there are 43,316 observations that possess an internal parking facility. 9,240 monuments are in the dataset and 1,552 residences are equipped with a pool.
Table 2. Descriptive statistics experimental variables
Variable Mean Std dev Sum
T100 0.0000 0.0063 35 T250 0.0004 0.0205 374 T500 0.0025 0.0499 2,229 T750 0.0064 0.0800 5,759 T1000 0.0118 0.1082 10,576 Inrange100 0.0003 0.0172 264 Inrange250 0.0039 0.0622 3,464 Inrange500 0.0187 0.1356 16,736 Inrange750 0.0444 0.2059 39,620 Inrange1000 0.0776 0.2675 69,282 Dactive 0.2336 0.4231 208,606 lotDIS 0.3287 0.4698 293,576 gardenDIS 0.5011 0.5000 447,463
Table 2 provides descriptive statistics on the variables necessary to estimate the effect of a refugee accommodation on house prices. T100, T250, T500, T750 and T100 are treatment effects, Dactive * Inrange(distance), in the ranges of 100, 200, 500, 750 and 1,000 meters. The sum noted in Table 2 shows how many observations are ‘treated’ in every range. The Inrange variable measures if the property is in the range of 100, 250, 500, 750 or 1,000 meters from the closest AZC or temporary refugee shelter. 16,736 observations are in the 500 meter range from the closest shelter, however only 2,229 are in effect treated. This is because the Dactive, indicating if an AZC is up and running, is multiplied by the Inrange variable to arrive at the treatment effect. So a property can be located in a 500 meter range but if the shelter has not opened yet, the treatment indicator equals 0 since it is expected that an AZC might have effect on house prices only after opening. The lotDIS and gardenDIS are indicator variables indicating if the observation was subject to editing for discontinuity reasons.
Table 3 present the overview of the different data sets used in this analysis. The main data set is split up into several pairs, each individual dataset is used in the analysis to check whether the result obtained in the main dataset are robust when applied to a dataset with certain characteristics. The data set is split into: apartments versus non-apartments, residences not coded as an apartment in the NVM dataset, Randstad versus non-Randstad region, cities,
- 20 -
more than 100,000 inhabitants, versus villages, less than 100,000 inhabitants, AZC, (semi-) permanent, versus emergency shelter, temporal and before versus after January 1st 2001. Table 3 shows what percentage of total data is used per dataset and how many unique municipalities are in each dataset.
Table 3. Specification of Datasets
Dataset Observations % of total Municipalities % of total
Apartment 313,687 35.13 68 100.00 Housing 579,326 64.87 68 100.00 Randstad 474,971 53.19 20 29.41 Non Rand. 418,042 46.81 48 70.59 City 624,662 69.95 49 72.06 Village 268,351 30.05 19 27.94 AZC 519,979 58.23 32 47.06 Temporal 373,034 41.77 36 52.94 Before 2001 262,182 29.36 68 100 After 2001 630,836 70.64 68 100
4. Methodology
4.1 A standard hedonic model
In order to estimate an effect of a shelter on house prices, first it in necessary to estimate the hedonic pricing model without the treatment effects to check the validity of the model. Explanatory variables for transaction price, according to Hedonic Price Models by Marc Francke (Francke, 2014), are variables on legal rights, conditions on sale, market conditions, location and physical characteristics.
Legal rights concern the rights people buy in a transaction, off course the property itself is included but sometimes land leases apply implying a rental contract of the site the property is located on is transacted and not the ownership. If land-leases are applicable, not uncommon in the Netherlands, this needs to be corrected for. The NVM dataset however does not provide data on this therefore it cannot be checked. But there is no strong reason to assume properties located near an AZC have different characteristics concerning legal rights compared to properties further away. This allows the estimation to be run without controlling for legal rights since no indications for a difference between treated and untreated is present, implying a random distribution of legal rights across observations.
Conditions on sale are important determinants of the selling price, as discussed in the preceding section. Observations that have been sold by auction will be removed from the dataset since it is unknown if the auction is voluntary or not.
Market conditions concern the conditions the market faced when the transaction took place. Conditions like interest rate, economic activity or inflation can affect the transaction price. These variables are not used in this analysis since it is possible to control for them by
- 21 -
controlling for the time period. An intercept per time period is included to incorporate the difference in these kinds of conditions over the years, in effect using time fixed effects.
Location is, at least in real estate, a dominant factor in explaining the value of a property. Incorporating this variable in the analysis can be done in two ways. First by controlling for location characteristics by including them in the analysis, for example including unemployment rate as an independent variable. This is necessary when regular OLS is used when these location characteristics differ over time. Second option is to include indicator variables per location, for example each postal code receives its own intercept thereby incorporating all location effects, like unemployment, of that postcode that are not incorporated in the regression. The scale of these intercept can range from municipality to address level. If each observation receives its own intercept, incorporating location effects, this is effectively introducing fixed effects in the regression and assumes these characteristics are stable over time.
As discussed in the data section on property characteristics provided by the NVM dataset, these physical characteristics possess, of course, lots of explanatory power regarding transaction price.
The results of the standard hedonic model are discussed in the empirical results section. After the standard hedonic model is validated the model will be used to estimate the effects of an AZC on transaction prices of residential properties. This extension of the standard hedonic model will be discussed in section 4.2.
4.2 Hedonic model in experimental setting
The method to test the hypothesis is a hedonic difference-in-difference regression estimation following the regression equations:
1) Ln (TP)it = β0 + β1 * Treatmentit + β2 * I. AZC in rangei + β3 * I. AZC activet + β4 * C.housing characteristicsit + ε it
2) Ln (TP)it = β0 + β1 * Treatmentit + β2 * I. AZC in rangei + β3 * I. AZC activet + β4 * C.housing characteristicsit + time fixed effectst + ε it
3) Ln (TP)it = β0 + β1 * Treatmentit + β2 * I. AZC in rangei + β3 * I. AZC activet + β4 * C.housing characteristicsit + time fixed effectst + location fixed effectsi + ε it 4) Ln (TP)it = β0 + β1 * Treatmentit + β2* I. AZC in rangei + β3 * I. AZC activet
+ β4 * C.housing characteristicsit + time fixed effectst + entity fixed effectsi + ε it
The dependant variable is the natural logarithm of transaction prices. Treatment is an indicator variable equalling indicator AZC in range * indicator AZC active. With indicator AZC in range being equal to 1 if an AZC is placed within a certain radius of the house and indicator AZC active being equal to 1 if the AZC is up and running, otherwise the indicators take on the value of 0. The area an AZC has its effect on house prices is expected to be in the range of several hundred meters. The possible channels of noise, pollution and potentially traffic safety surrounding an AZC are very closely tied to the location of the refugee shelter
- 22 -
thus these factors are expected to have a possible effect only in close range of a shelter with diminishing impact if distance increases. The perceived increase in crime, increased prejudice and decreased perceived safety are not so much tied to the location of the shelter because a feeling cannot be measured in meters removed from the shelter. These intangible possible externalities are not expected to diminish over distance but are assumed to be heterogeneous distributed across inhabitants of a neighbourhood or geographical area an AZC is located.
The control variables on housing, C.housing characteristics, enter the equation like a vector of multiple characteristics to make sure the estimation of the treatment effect is precise and controlled for characteristics that can be determinants of house prices. Included control variables on housing characteristics are: the natural logarithm of lot size, garden, surface and volume, indicator variables equalling 1 if an elevator, attic, parking space, garage, swimming pool or monument is applicable the property, multiple indicator variables on NVM-grade, type of property, quality of maintenance inside and outside and building period as well as the continuous variables on number of floors and number of rooms. When the estimation are run for apartments only, instead of house type indicator variables the variables on quality and open porch are used.
Time fixed effects make sure that the estimation is controlled for effect of time, this should eliminate any form of seasonality and trends. Time fixed effect control for factors are constant across entities but change over time. Time fixed effects enter as a set of yearly indicators.
Location fixed effects, assuming omitted variables that differ among locations are stable over time, control for being located in a certain dimension of space. In this case postal codes will be used to incorporate location fixed effects, in effect including an intercept for each minus one unique postal code thereby controlling for being located in a certain postal code area.
Entity fixed effects control for being a certain property and can be accounted for by including intercepts for each minus one entity in the regression. This corresponds to stating that each observed property has its own specific characteristics and by including an intercept these property specific characteristics are accounted for. For example the floor plan of a house in combination with its garden or view can be unique and by including these entity fixed effect this is accounted for. These effects can vary across entities but stability over time is assumed.
The assumptions underlying this model are the parallel trend assumption, meaning that if an AZC was not placed the house prices in this region would have behaved in the same way as places an AZC was never placed. This is a feasible assumption since the AZC announcements in general are a surprise to the neighbourhood and to the residential real estate market. Only the COA and the municipality know in advance where an AZC location might be possible. However, some neighbourhoods in a municipality can differ substantially from others, therefore some very diverse real estate market can exist within a city. Every submarket, apartments, single-family homes or flats residences, has its own characteristics
- 23 -
and makes comparing them, in the sense of a reaction to an AZC, more difficult if not controlled for these features. In this case, since all submarkets and neighbourhoods in a municipality are present in the NVM dataset, the analysis is about the general effect of a refugee centre on the residential real estate prices. Knowing different submarkets and neighbourhoods are present in the dataset, the population is a heterogeneous one and that is why the regression estimates will produce an average causal effect. The causal effect of being close to an AZC can differ per property type or neighbourhood and by taking the aggregate effect of all these submarkets and locations, an average treatment effect will emerge. Regression techniques in a heterogeneous population will, if the treatment is randomly assigned, produce consistent estimators of the average causal effect (Watson, 2011).
The other assumption is that no selection bias is present. This can be checked and controlled for by adding the possible selection criteria. If in this situation randomization is not fully present randomization based on covariates, treatment is random given certain characteristics of a neighbourhood in essence stating the 1st assumption of the regression models mentioned in Appendix A, however can produce consistent estimators of the causal effect (Watson, 2011). Control variables on location are the possible covariates that are necessary to estimate the models properly. A possible way to check if locations of an AZC and the control group significantly differ is to check the Inrange variable. A significant coefficient on the Inrange variable can signal that the treated and the control group’s locations differ substantially. But still it is hard to determine what needs to be controlled for because a significant coefficient means that an average difference across municipalities is possibly present. However because it concerns an average effect in a heterogeneous population, the exact source of this difference cannot be pinpointed because the locations differ on to many aspects.
An overview of the remaining regression assumptions is presented in Appendix A. These assumptions are applicable in a more general regression analysis context. The stated assumptions concerning the parallel trend and selection bias are in particular important for the methodology used in this analysis.
The method is fit to estimate the effect of an AZC on house prices because an AZC announcement can be seen as a policy change or treatment or quasi experiment to one part of the sample while the other part remains untreated. In this way, when the assumptions of the model hold, the treated part of the sample can be easily compared to the non-treated part of the sample in a before and after analysis yielding an average treatment or causal effect.
Possible violations of the assumptions of the model can be a selection bias arising from the location an AZC is placed. A temporal refugee shelter is temporal because the need is very high. This might induce temporal shelters being located in areas with sufficient space or in obsolete properties, sufficient space implies lower value of the location and thereby lower house prices. Obsolete properties can be locations that are currently not in use, like an office or prison, that are transformed to a shelter. Selection bias can arise because an office not used might imply lower economic activity or a location that does not suit modern standards thereby influencing the randomness of the location. To address this problem it might be
