The influence of induced earthquakes on housing markets : an empirical analysis for the Netherlands

Finance Department Faculty of Economics and Business University of Amsterdam

Master Thesis: Business Economics

Specialisations: Finance and Real Estate Finance

The Influence of Induced Earthquakes on Housing Markets:

An empirical analysis for the Netherlands

Author: Niels Schroers BSc. Student number: 5877377

Supervisor: Prof. Dr. J.B.S. Conijn Date: 25 February 2015

2

Statement of Originality

This document is written by Student Niels Schroers who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

3

Abstract

This thesis investigates the influence of induced earthquakes on the housing market in the northeast part of the Netherlands. In this area, earthquakes occur due to gas extraction. Therefore houses get damaged and people feel insecure about their living area.

Data from the Dutch brokers association (NVM) and the statistics Netherlands (CBS) is used to study the effect of earthquakes on two aspects of a housing market. The first aspect that will be studied is the influence earthquakes have on house prices. This testing will be conducted using a hedonic pricing model. The second aspect study will be done on the number of houses being sold in the market using a difference-in-difference model.

The results of the aspects studied do not confirm a significant connection between earthquakes and their effect on house prices and number of houses sold within the market. In conclusion, there is no scientific evidence suggesting that induced earthquakes have an influence on the housing market in the area studied.

4

Table of Contents

1. Introduction ... 5

2. Literature Review ... 8

2.1 Supply Side Determinants ... 8

2.2 Demand Side Determinants ... 9

2.3 Environmental Externalities ... 10

2.3.1 Externalities from Earthquakes ... 11

3. Methodology ... 15

3.1 Models ... 15

3.2 Treatment and Control Group ... 19

3.2.1 Treatment Group ... 19

3.2.2 Control Group ... 21

4. Empirical Analysis ... 23

4.1. Data ... 23

4.2 Statistical Data Analysis ... 24

4.3 Empirical Analysis ... 31

4.3.1 Results of the Hedonic Pricing Model ... 31

4.3.2 Results of the Difference-in-Difference Model... 35

4.4 Robustness Checks ... 37

4.4.1 Hedonic Pricing Model ... 37

4.4.2 Difference-in-Difference ... 39

5. Conclusion and Discussion ... 42

6. References ... 45

7. Appendices ... 45

5

1. Introduction

Many studies have been conducted on housing price determinants. These factors are of interest to different groups involved in the housing market, such as developers, governments, real estate and financial professionals, and households. Especially for the last group, housing prices play a significant role in their lives. In The Netherlands, 60% of the households are homeowners (CBS, 2013). More than 60% of their total capital is allocated in the property, which cause this asset to be the biggest in the whole portfolio. This means that even a small change in house prices can have serious consequences on their financial position.

Many housing price determinants have been already defined in previous studies. Macro-economic factors are one group that influences housing prices. One can think of interest rates, income and employment rates. Other determinants are geographical factors like population growth or government policies such as subsidies and regulations. This is discussed further in the literature review section.

This thesis focuses on environmental externalities, activities in the neighbourhood that influence home prices. They can affect house values in positive way by providing good water quality (Poor et al., 2007) or open space (Luttik, 2000). On the other hand, negative factors such as pollution, noise (Lusht, 2001) and the threat of natural disasters decrease home values. Natural disasters that already have been studied are floods (Bin et al., 2008), wildfires (Mueller et al, 2007) and earthquakes (Beron et al.,1997). This thesis studies one specific type of externality which is induced earthquakes.

There are two types of earthquakes, namely natural and induced earthquakes. Natural earthquakes are those caused by natural processes of the earth. On the contrary, induced earthquakes are caused by human activity. The latter can be caused, for example, by reservoirs (Gupta, 1992), fluid injection (Zoback, 1997), geothermal systems (Majer, 2006) or fluid extraction (Segall, 1989).

One type of fluid extraction is the extraction of gas from the soil. Globally, gas is one of the most important energy sources. Worldwide, more than 20% of electricity is generated by using gas as an energy source (International GAS Union, 2010). Most gas producing countries export some of their gas. In 2012, 3364 billion cubic metres of natural gas was produced (BP, 2013); 1033 billion cubic metres of this total output was exported. The revenues from gas export are important for many of the producing countries.

The Netherlands is the third largest producer of gas in Europe. In 2012, the income from gas production was 11.8 billion euro, which is 5% of the total income of the Dutch

6

Government and 1.8% of the total GDP. The biggest gas field, “Het Groningerveld,” is located in the northeast part of the Netherlands. Recently there has been much unrest among residents who live near this gas field. In this area gas is extracted from a gas field located at a depth of three kilometres. The gas extraction began in 1986, and since that time earthquakes have been observed in this area. Over time the number of earthquakes and their intensity have increased (Figure 1).

Figure 1. Number of observed earthquakes and magnitude of earthquakes in Groningen gas field

Source: KNMI

As a result of the earthquakes, houses have been damaged, people feel unsafe, and the liveability is decreasing (Klaassen, 2013).

The situation around “Het Groningerveld” is not an individual case. The extraction of shale gas is a relatively new phenomenon that can cause earthquakes (Das & Zoback, 2011). In Blackpool, England, seismic activity has already been attributed to the fracturing of shale, which is a part of the shale gas extraction process (Green et al., 2012). In Europe, there are several shale gas reserves in urban areas (Appendix 1). As there are plans to extract gas from these resources, it is important to know the impact of earthquakes on the housing market. The goal of this research is to identify this effect. The results of this research might contribute to better investment decisions in earthquake inducing activities. If it appears that induced

7

earthquakes negatively influence housing markets, these consequences can be taken into account in the investment decision-making process.

The influence of induced earthquakes on housing markets is researched on the basis of two sub-questions:

1. What is the influence on house prices?

2. What is the influence on the number of houses sold?

Different methods have been used to determine the influence of induced earthquakes on house prices and the number of sales. In order to answer the first question, a hedonic pricing model extended with time variables was used. As well as a difference-in-difference model was applied to identify the effect on the number of houses sold. As a sample for this research, “Het Groningerveld” area was used. Data was provided by the Dutch Brokers Association (NVM).

The paper is organized as follows: in this first section, the topic of this thesis is introduced and the research question and sub question are formulated. The second section discusses the consisting theory related to this topic and the two hypotheses based on this theory. The following section describes the empirical framework and the used models. Furthermore, the studied area is extensively discussed and explained in this part. Described in section 4 is the empirical analysis that has been conducted. In the first part, the data and the data shaping is described. Thereafter, statistical analyses are conducted on all variables separately. In the second part, the regression results are displayed and explained. In final section, conclusions are drawn and the applications of this thesis discussed. Additionally, within this section the research has been considered and recommendations for further research are given.

8

2. Literature Review

Literature with reference to thesis topic is discussed using a top-down approach in this section. First, the supply-side drivers of housing markets are discussed. Thereafter, determinants of the demand-side have been considered. The following paragraph elaborates more on certain demand-side determinants and environmental externalities. The final part narrows down environmental externalities to earthquakes, which are the most crucial for this thesis.

Many studies have been conducted on the determinants of house prices. This section presents some of this research to position the topic of this thesis against existing literature. Due to the fact that houses are sold in an open market, their prices are determined by supply and demand. A determinant that is hard to classify are housing characteristics. Several authors found characteristics such as size, residence type, quality and age to be significant factors. As this is very straightforward, it will not be discussed any further.

Hereafter, different determinants of both the demand side and supply side are discussed.

2.1 Supply Side Determinants

The first factor that influences prices of both existing and new housing is construction costs (Abraham & Hendershott, 1996). Every two years, the American National Association of Home Builders conducts a survey analysing construction costs of houses. The 2013 survey shows that 59% of the house price consists of construction costs (Heather, 2014). This means that a change in construction overhead has a significant influence on the total costs of a house. Malpezzi (1996) conducted research on the determinants related to regulation. He used U.S. rental prices and home values for all metropolitan areas with more than 50,000 people. The author took into account rent control, zoning, and other regulatory indices. Malpezzi concluded that regulations increase both rental prices and home values. Furthermore, different studies concerning specific types of regulations, such as land use restrictions, have been carried out. (Pollakowski & Wachter, 1990 and Ihlanfeldt, 2007). It is crucial to investigate its impact, as land is a very important component of housing prices. Therefore, its value plays an essential role in the residential market’s fluctuations. Results of discussed research show that land use regulations drive up land values, hence, increases home prices. Nevertheless, land regulations are not the only determinants of a house its value. Transportation cost is another determinant. It can be easily explained by the bid-rent curve function of Alonso (Alonso,

9

1960). This model presents a very simplified version; however, it clearly shows how transportation costs influence land values. The bid-rent curve reflects the maximum amount a potential buyer is willing to pay for certain real estate, depending on its location. It is a function of the distance to a central point. The central point being the location where the transportation cost is equal to zero. Therefore, land value in those points is the highest. In other words, potential buyers are willing to pay more for land that is closer to the central point to reduce transportation costs. This effect explains why land usually is more expensive in cities than in the countryside.

2.2 Demand Side Determinants

Many studies have been conducted on the other group of determinants which consist of demand side components. Reichert (1990) investigated the impact of different important economic and geographical factors in the U.S. For this research, quarterly data over 12 years was used. His results show some interesting outcomes. First, the author found that a 1% growth in population is associated with a 1.09% increase in real house prices. For an increase in employment rate of 1%, an increase of 0.91% was found in real housing prices. Furthermore, the author concluded that an increase of mortgage interest rate from 5% to 6% (i.e., a relative increase of 20%) causes a decline of housing prices by 3.8%. Lastly, the author found the greatest price effect for permanent income. A 1% growth causes an increase of 3.78% for real housing prices.

The influence of tax subsidies and inflation was examined by Poterba (1984). For this study, asset-market model was used to estimate the impact of tax policy and inflation on the relative price of owner-occupied housing. The author conducted research on the housing boom of the 70s in the U.S. At that time, the prices of owner-occupied dwellings increased by 30%. Different inflation shocks were simulated by two different tax rates. Results show that in both cases the relative house price increases more than the inflation. For example, an inflation shock from 0% to 2% and a tax rate of 25% would increase housing prices with 8.3%. The same simulation can be used to identify effects of policy changes. Mortgage interest deductibility is one of the cases of such policies where impact has been already examined. Estimates suggest that removing mortgage interest deductibility causes a decline in house prices of 26%.

10

2.3 Environmental Externalities

“Environmental externalities have been defined by OECD as the economic concept of uncompensated environmental effects of production and consumption that affect consumer utility and enterprise cost outside the market mechanism.” (OECD, 2003). As this definition applies to housing markets, it means that environmental externalities are activities that occur and influence the housing price outside the direct control of the homeowners. Discussed in this section are existing studies on these factors.

It has been proven that pollution is a negative externality to housing values. Smith & Huang (1993) conducted research analysing the influence of air pollution on home values. They used a hedonic pricing model to analyse 37 previous studies from the past 25 years. Every study they examined included at least one air pollution measure. In all studies a significant result was found for the air pollution measure.

Water pollution is another factor that has been examined by different authors. Michael et al. (1996), Leggett & Bockstael (2000) and Poor et al. (2007) have already investigated ambient water quality externalities. All off these researches found a positive causal relation between water quality and home prices. Additional authors looked at the effect of landfills on home values ((Reichert, et al. (1991), Nelson et al. (1992)). The results show a decrease in home values between 3% and 7%, depending on distance from the landfill and the price of a dwelling. The price of expensive dwellings suffered more from neighbourhood landfills than cheaper dwellings.

In addition to water pollution, traffic noise is another externality that has negative impact on property value. Theebe (2004) studied the effect of planes, trains and automobiles on residential property values in the Netherlands. He used data from the western region of the Netherlands, which is an area with one of highest population density among developed countries in the world. One of Europe’s main airports is located in this area, however, this is not the only cause of noise nuisance. Due to the fact that the population density is high, the average use of roads and tracks is also higher compared to other European countries. The author used a hedonic pricing model including precise noise measures. The results showed that traffic noise effected housing prices if it exceeded 65dB. The estimators indicated a decline in residential property values of up to 12%.

The next group of externalities is related to natural disasters which is directly connected to the core topic of this thesis. Different types of natural disasters have been studied such as wildfires and flooding. Mueller et al. (2007) studied the effect of wildfire risk

11

on housing prices. They analysed the long-term and short-term effects of forest fires in California. Using a hedonic pricing model they found that housing prices in wildfire areas decreases by 10% when they were exposed to a wildfire for the first time. The second fire would decrease the prices by 23%.

Another natural disaster that has been studied is flooding. Most researches used capitalized insurance premiums and compared those to the reduction of house value in flood risk areas (Speyrer & Wade (1991), Bin, et al. (2008)). Generally, a decrease in housing value of 4% to 12% was found. Bin and Polasky (2004) used a different approach. They looked at the effect on residential properties in North Carolina (U.S.) before and after the flooding caused by hurricane Floyd in 1999. In the area researched, 6,000 houses were flooded and 50,000 people were affected by this catastrophe. A hedonic pricing model was used to identify the effect of flooding on the residential values. Authors found an average decrease of home values of 5.7% in the flood plain.

Another common natural disaster is earthquakes. This phenomenon is extensively discussed in the next sub-section.

2.3.1 Externalities from Earthquakes

As previously discussed, earthquakes can be classified in two groups: natural and induced earthquakes. Due to the fact that there is no existing literature about the influences of induced earthquakes on housing markets, studies about natural earthquakes will be reviewed. Thereafter, similarities and differences between natural and induced earthquakes and the theoretical consequences for the housing market will be explained. Lastly, a hypothesis will be formulated based on the literature.

Beron et al., (1997) analysed the effects of the Loma Prieta earthquake using home sales data from the San Francisco Bay area. They expected a causal relation between the premium of earthquake risk and the probability of property loss from an earthquake based on the general self-protection model of Ehrlich and Becker (1972). This model shows that people are willing to pay for self-insurance (a reduction in the size of a loss) as long as marginal benefits are higher than the marginal costs. In addition to this, Brookshire et al. (1985) showed that self-protection (a reduction in the probability of a loss (Ehrlich & Becker, 1972)) for earthquake hazards is facilitated by housing markets. A way of self-protection against losses is to buy a house in an area where the chance of loss from hazards is low. People are willing to pay more for a house in those areas than in a region where the loss probability is

12

high. In the main article Beron et al., used a hedonic pricing model to estimate the price of earthquake risk. This price estimates the additional amount that people are willing to spend to live in a non-earthquake area; in other words, the cost of self-protection. They found that the price of earthquake risk decreased after an earthquake took place. This indicates that homeowners overestimated the earthquake hazard before the earthquake occurred and that the costs of self-protection were too high.

Willis and Asgary (1997) explained the impact of earthquake risk reduction measures on housing markets with data from Tehran real estate agents. This data was collected using a survey method. It allowed authors to use a customer value model to investigate price differences between houses which were improved with earthquake risk reduction measures and houses without these kinds of improvements. They concluded that the value of resistant houses is 35% higher than non-resistant houses when corrected for land value. They also found that if more information is available concerning earthquake risk, people become more sensitive to earthquake risk reduction measures.

Naoi, et al., (2000) conducted research on property values before and after a major earthquake in Japan. A hedonic pricing model, was used including an objective earthquake occurrence measure, to determine housing and rental prices. They began by running a regression without distinguishing between prices before and after massive earthquakes. The estimator of earthquake occurrence probability shows an insignificant estimator decrease in rent and a small negative effect on housing prices. In the second regression, a distinction was made between the pre- and post-earthquake period by adding a dummy “earthquake occurrence probability × post-earthquake.” Regression results showed a significant decrease in average rent prices of 16% and in average housing prices of 13% when annual earthquake probability increases with 0.2%. It is remarkable that the earthquake occurrence probability variable in the second regression became insignificant, where the “earthquake occurrence probability × post-earthquake” is significant. The authors interpreted this to mean “households are initially unaware of, or at least underestimate, the earthquake risk in the pre-quake period.”

Francke & Lee (2013) conducted a research on an earthquake area in Groningen and the impact it had on the housing market. They did not find any significant deviations in the housing market caused by earthquakes. However, a disadvantage is that the models used are not publicly available and therefore the results cannot be verified. This means that this research cannot be part of the academic debate.

13

As mentioned before, the causes of natural earthquakes differ from causes of induced earthquakes. This does not necessarily mean different types of earthquakes affect the housing markets differently. The differences and similarities of consequences resulting from the different types of earthquakes, especially on the housing market of “Het Groningerveld” area are discussed below.

One difference between natural and induced earthquakes is the predictability of the occurrence. Scientists are still unable to predict the point in time and magnitude of natural earthquakes. On the other hand, induced earthquakes are easier to predict. As can be seen in the case of “Het Groningerveld” in Figure 1. The more gas that is extracted from the gas field, the more earthquakes occur, and the magnitude is higher. This is also shown by the TNO (2014) research institute. If production keeps up with market demand, the seismic threat will be earthquakes with a magnitude of 4 to 5 on Richter Scale in the next ten years. This difference has consequences for the information availability to homeowners. Uncertainty for people living in the induced earthquake area in Groningen is less than for people in natural earthquake areas. This means that homeowners in Groningen are better informed and therefore, they are able to make better decisions regarding earthquake consequences.

Another difference is that it is not possible to insure against induced earthquakes, where it is possible to insure against natural earthquakes depending on country. Homeowners in countries with high natural earthquake risk, including Turkey (TCIP, 2012), Japan (GIROJ, 2013) and the Unites States (CEA, 2014), have the option to purchase insurance for seismic hazards or are obligated to do so. People living in induced earthquake regions do not have this option and should look into options to secure themselves against earthquake damage. As a result, people may look for self-protection by moving to an area outside the earthquake region. Furthermore, for people living outside the earthquake area, the earthquakes will have a dissuasive effect. Both will cause a decline in demand for the houses in the earthquake area.

One could also argue, that homeowners can be compensated by the NAM, the company which owns the rights to gas extraction from “Het Groningerveld.” However, NAM will not reimburse for all earthquake consequences. The fear that people may have or the decreased liveability will not be recompensed by the NAM. Furthermore, it takes a lot of effort to get the compensation. Homeowners have to go through a long and complicated procedure and they must prove that the damage was caused by an earthquake, which might

14

not be easy. In summary, the possibility of reimbursement will reduce the effect of earthquakes on housing markets, but will not eliminate it.

Based on the theory above, expectations can be formed for both sub-questions. For the first question, the influence on house prices, one would anticipate a decrease in house prices. Similarly for the second question, the influence on the number of houses sold, one would anticipate a decrease in number of sales. From these expectations, the following statistical hypothesis has been formulated for each sub-question.

H1. The induced earthquakes in “Het Groningerveld” do not lead to a decrease in house values.

H2. The induced earthquakes in “Het Groningerveld” do not lead to a decrease in number of sales.

15

3. Methodology

This section describes the used methodology. Two different approaches were used to answer both research questions.

For testing the first hypothesis, two commonly used methods have been taken into account. These are repeated sales model and hedonic pricing model. The repeated sales model compares sale prices of certain properties in different points in time. The main advantage of this model is that no other characteristics have to be included, as the exact same properties are used. The main drawback is that the model is inefficient because not all sold properties are included, only the properties that are sold more than once within a certain time frame. This also brings forward a possible lack of data and limited number of transactions to be analysed. This is why this model was not chosen to test the hypothesis.

Another commonly used framework is the hedonic pricing model, which explains a home price based on its individual characteristics. Compared to a repeated sales model, hedonic pricing models are efficient as no transactions are omitted in advance. Furthermore, the model can be easily adapted by adding or removing particular features. The main disadvantage is that much data is necessary to estimate a thorough model. However, for this particular research a sufficient amount of information is available.

For the influences of induced earthquakes on housing prices the following model is specified (Equation 1): 𝑙 n(𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛 𝑝𝑟𝑖𝑐𝑒)_𝑖 = 𝛽0+ 𝛽1∗ (𝐴𝑟𝑒𝑎 𝑥 𝑞𝑢𝑎𝑟𝑡𝑒𝑟 𝑑𝑢𝑚𝑚𝑦)𝑖 + 𝛿1∗ 𝐿𝑜𝑡 𝑠𝑖𝑧𝑒𝑖+ 𝛿2∗ 𝑉𝑜𝑙𝑢𝑚𝑒𝑖 + 𝛿₃∗ 𝐸𝑥𝑡𝑒𝑟𝑖𝑜𝑟 𝑚𝑎𝑖𝑛. 𝑑𝑢𝑚𝑚𝑦_𝑖 + 𝛿₄∗ 𝑅𝑜𝑜𝑚𝑠_𝑖 + 𝛿₅∗ 𝑇𝑦𝑝𝑒 𝑑𝑢𝑚𝑚𝑦 _𝑖 + 𝛿₆∗ 𝑃𝑎𝑟𝑘. 𝑑𝑢𝑚𝑚𝑦_𝑖 + 𝛿₆∗ 𝐶𝑜𝑛𝑠𝑡𝑟𝑦𝑟 𝑑𝑢𝑚𝑚𝑦_𝑖 + 𝛿7 ∗ 𝑀𝑜𝑛𝑢𝑚𝑒𝑛𝑡 𝑑𝑢𝑚𝑚𝑦𝑖 + 𝛿8∗ 𝐺𝑜𝑜𝑑 𝑔𝑎𝑟𝑑𝑒𝑛 𝑑𝑢𝑚𝑚𝑦𝑖 + 𝜑₁∗ 𝑄𝑢𝑎𝑟𝑡𝑒𝑟 𝑑𝑢𝑚𝑚𝑦_𝑖+ 𝜃₁∗ 𝑍𝑖𝑝 𝑐𝑜𝑑𝑒 𝑑𝑢𝑚𝑚𝑦 _𝑖+ 𝜀_𝑖

The dependent variable, “ln(transaction price)” is the natural logarithm of the house price. The interaction term “Area x quarter dummy” is the variable that shows whether the house is in the earthquake area and in which quarter it was sold. The coefficient of these interaction terms are of the main interest. It shows the relative price change in a specific quarter, for

16

houses located in the earthquake area. The criteria deciding about both groups are explained in part 3.2. Variables with 𝛿 as coefficient are house characteristics. Variable “Quarter dummy” indicates the quarter in which the house was sold in the selected sample period. “Zip code dummy” indicates in which zip code area the house is located. In the case of a zip code that falls in both the earthquake area and control area, the postcode is separated based on those two areas. Table 1. contains all variables used and their descriptions.

17

Table 1. An overview of the variables included in equation 1

Transaction price: The transaction price of the dwelling

Area * quarter dummy: Area=1 when dwelling is located in earthquake area quarter is the period 1998q1 - 2014q1

(2014q1 is left out in regression)

Lot size: The lot size of the property in square meters Volume: The volume of the house in cubic meters

Exterior maint. dummy: - Poor - Moderate to poor - Moderate (left out in regression) - Moderate to reasonable - Reasonable - Reasonable to good or not filled

- Good - Good to excellent - Excellent

Rooms: The number of rooms

Type dummy: - Simple (left out in regression) - Single family

- Town house - Residential farm - Bungalow - Manor house

Park. dummy: - No parking space (left out in regression) - Parking lot

- Carport and no garage - Garage and no carport - Garage and carport - Garage for multiple cars Constryr dummy: - Construction year 1500-1905

- Construction year 1906-1930 - Construction year 1931-1944 - Construction year 1945-1959 - Construction year 1960-1970 - Construction year 1971-1980 - Construction year 1981-1990 - Construction year 1991-2000

- Construction year after 2000 (left out in regression) Monument dummy: dummy variable with value 1 if the dwelling has a monumental

status; 0 otherwise

Good garden dummy: dummy variable with value 1 if the dwelling has garden with good maintained garden; 0 otherwise

Quarter dummy: 1998q1 - 2013q4 (2014q1 is removed from the regression) Zip code dummy: 910, 912, 913, 914, 915, 929, 930, 931, 932, 933, 934, 935,

947, 948, 949, 956, 960, 961, 962, 963, 964, 965, 966, 967, 968, 969, 976, 977, 978, 979, 980, 981, 982, 983, 984, 985, 986, 988, 989, 990, 991, 992, 993, 994, 995, 996, 997, 998, 999, 961E, 994E (994E is left out from the regression)

18

In order two answer the second research question, a difference-in-difference (diff-in-diff) model was used. Diff-in-diff models are suitable for identifying a causal effect while correcting for non-random effects by automatic correction based on pre-treatment characteristics. To estimate unbiased estimators the common trend assumption must hold. This means that the number of sales in the treatment and control group would have followed the same trend if no earthquakes would have occurred. Since it is not possible to test whether this is the case, the number of sales in the pre-earthquake period were compared to test if both regions follow the same trend over time. When this is the case, one should be aware that no variables other than earthquakes affect one of the groups separately. Otherwise, control variables should be added to correct for these effects. Section 4.2 shows that it is not unlikely that the common trend assumption holds for the sample used.

The following model has been specified (Equation 2) for the influence of induced earthquakes on the number of sales:

𝑙𝑛(𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑙𝑒𝑠_𝑖𝑡)

= 𝛾₀+ 𝛾₁ ∗ (𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑎𝑟𝑒𝑎𝑖∗ 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑝𝑒𝑟𝑖𝑜𝑑_𝑡)

+ 𝛾₂∗ 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑎𝑟𝑒𝑎 𝑑𝑢𝑚𝑚𝑦_𝑖+ 𝛾₃ 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑝𝑒𝑟𝑖𝑜𝑑 𝑑𝑢𝑚𝑚𝑦_𝑡+ 𝜖𝑖

The dependent variable “ln(number of salesit)” is the natural logarithm of the number

of sales in area i in period t. 𝛾₁ is the diff-in-diff coefficient 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑎𝑟𝑒𝑎_𝑖∗ 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑝𝑒𝑟𝑖𝑜𝑑𝑡. The estimator of this coefficient is of main interest. It shows the

percentage change of sales in the earthquake area in the selected treatment period with respect to the control area. Coefficient 𝛾₀ is a constant that shows the general difference between number of sales in the earthquake area and control area. “Treatment area dummy” is a dummy variable which is 1 for houses located in the earthquake area and 0 if otherwise. “Treatment period dummy” is a dummy variable for a certain period which is 1 if the quarter is in this period and 0 if otherwise. Its coefficient 𝛾₃ shows the percentage change in number of sales in this period.

19

3.2 Treatment and Control Group

In this section the determination of the control group and treatment group is described. For this research the same treatment and control group were used as per study of Francke & Lee (2013).

3.2.1 Treatment Group

The treatment group should cover the area that is exposed to the induced earthquakes. Therefore, this group is also called “earthquake area”. Francke & Lee (2013) used as treatment group the area that is formed by “Het Groningerveld” and the neighbouring area where damages occur due to earthquakes. They had to choose a classification to clearly delineate this area. In the Netherlands, multiple classifications of areas are available. The country is divided into 12 provinces, which consist of 40 COROP1 areas, including 403 municipalities and 4,878 zip code areas. Francke and Lee had to find balance between number of observations and the size of the treatment area. Insufficient number of observations in a certain area would bias estimation results. In contrast, creation of too big areas will include houses that are not affected by earthquakes and will skew estimation results. The authors decided to use the municipalities to construct both treatment and control groups.

Francke & Lee had to determine which municipalities would be marked as earthquake municipalities. They decided to include municipalities which were exposed to earthquakes with a magnitude of 2.4 or higher on the Richter Scale. These municipalities are Appingedam, Ten Boer, Delfzijl, Loppersum, Slochteren, and Eemsmond. The municipalities of Bedum and Winsum were also added to the earthquake area because the effects of earthquakes were observed there as well. In Figure 2. the treatment municipalities are marked in red. The control municipalities are marked in green and the excluded municipalities are marked in blue. More information concerning the control and excluded municipalities can be read in the next section.

1

A COROP area is a regional area within the Netherlands and is an amalgamation of municipalities. The COROP grouping is often used by research institutions to display statistics.

20

Figure 2. Map of the research area

21

3.2.2 Control Group

Selecting a good control group is really important to obtain unbiased results. Ideally, the control group is exactly the same as the treatment group except of the factor that is being tested. If this is not the case, factors other than the factor of interest can influence the regression results in a negative way.

For this research it means that the control area cannot in anyway be influenced by the induced earthquakes at all. However, the control and earthquake area should be as similar as possible. For this reason, Francke & Lee chose the municipalities adjacent to the earthquake area. Thereafter, they compared those municipalities with those of the treatment group to see whether those municipalities were comparable to the control group.

The numbers used for the comparison come from statistics Netherlands (CBS) and the land register (Kadaster).

In the control group, figures show that some municipalities strongly deviate on specific characteristics. First, Groningen and Assen are deviates because of their urbanization (appendix 2). Those municipalities are “very strongly urbanized,” which can greatly influence housing prices. Additionally, the municipality Haren differs because of significantly higher incomes. Finally, in Aa en Hunze and Midden Drenthe earthquakes occurred. However, this municipality is relatively far from “Het Groningerveld”. These municipalities are not well comparable to the rest of the municipalities in the Northeast of the Netherlands, so they are not included in the control group. All other municipalities in the control group do not systematically deviate from other municipalities in Northeast Netherlands.

22

In conclusion, it can be stated that the chosen control group is fairly comparable to the treatment group. The main differences between control and treatment are in WOZ value (lower in earthquake area) and in demographics (decline in population in treatment group).” (Francke & Lee, 2013)

Nr. Factor Eartquake Control group Excluded Research area North Netherlands 1 ΔP -0.15% 0.05% 0.73% 0.30% 0.34% 2 ΔHH 0.45% 0.61% 1.07% 0.79% 0.83% 3 %OH 64.0% 64.8% 50.2% 59.3% 60.0% OH 31,048 83,047 74,210 256,029 446,002 4 ΔOH 0.75% 1.01% 1.86% 1.30% 1.38% 5 %V 3.3% 3.6% 5.1% 4.0% 3.8% 6 WOZ (x €1000) 168 185 190 187 190 7 I (x €1000) 22.2 22.00 22.6 22.1 22.1 8 ΔI 3.5% 3.4% 3.6% 3.5% 3.4% 9 %U 2.7% 3.0% 2.3% 2.8% 2.9% 10 %SAP 3.8% 4.2% 5.7% 4.6% 4.4% 11 Urb ++ 0.0% 0.0% 33.3% 11.3% 8.4% Urb + 0.0% 5.4% 15.2% 8.3% 9.6% Urb +/- 8.6% 16.1% 17.2% 14.3% 15.9% Urb - 33.0% 27.3% 15.5% 23.3% 22.5% Urb -- 58.4% 51.2% 18.7% 42.8% 43.6%

ΔP Population growth, on average per year from 1995 and 2013

ΔHH Growth of number of households, on average per year between 1995 and 2012 %OH Percentage of owner occupied houses in 2010

ΔOH Growth of number of owner occupied houses, on average per year between 2006 and 2010

%S Average number of sales per year (between 1993Q1 and 2013Q) as percentage of number of sales in 2010 WOZ Average WOZ -value as of January 2012 with the price level of January 2011

I ΔI %U %SAP Urb

Source: Francke & Lee (2013), adjusted. Average standardised income in 2010

Growth of standardised income, average per year between 1999 and 2010

Unemployment rate indicator, calculated from number of unemployment payments as percentage from population between 20 and 65, per 2011

Social assistance payments as percentage of number of households, per 2009 Degree of urbanization

23

4. Empirical Analysis

This section describes the origin of the data followed by the procedure of cleaning the dataset. Thereafter, statistical analyses have been conducted and the results are displayed in section 4.2. In the last subsection, the regression results are presented and analysed.

4.1. Data

For this study, data from the Nederlandse Vereniging van Makelaars (NVM) has been used. The NVM is an association of brokers and real estate specialists with around 3,900 affiliated brokers. The NVM sustains a database with information about houses sold and rented in the Netherlands. It contains sale prices, dates, house characteristics, and other variables. The database contains around 70% of the total home sales transactions in the Netherlands. Therefore, it is a good reflection of the total housing market.

From the NVM database, a raw sample was taken containing 66,241 observations from 1998 until the first quarter of 2014. Not all observations are relevant and some data is not or incorrectly filled in by the members of the NVM, hence some adjustments to the dataset had to be made. Furthermore, new variables have been created based on existing ones. This process is described below.

The first step in adjusting the dataset was removing observations that did not meet certain criteria. All observations that are not in one of the municipalities subject to this study were removed. Then, objects that are not dwellings, are fully or partly rented, or were sold at auction, have been deleted. Observations with missing data on dwelling type, transaction date, and transaction price have been rejected as well. Dwellings with a lot smaller than 10m2 and larger than 900,000 m2 have also been removed. Apartments and flats are dropped as well because many observations had abnormal values for lot size and volume. Removing this type of dwelling will not bias the estimation results as less than 0.3% of the total sold dwellings were flats or apartments. Objects with transactions before their construction period have been deleted. Finally, observations with a transaction prices less than € 10,000 and more than €6,000,000 have been dropped.

After cleaning the dataset, several new variables were created. Based on municipality ID, two dummies for treatment and control group have been created. Four dummies for quarters were made basing on the transaction date. The dataset contains information about the construction period. Based on this variable, nine dummies for the construction period have been generated. The dataset includes variables that show whether a dwelling has a parking

24

facility as well as its type. From this variable, six dummies were created which show the type. Another variable that is covered in the dataset is dwelling type. Six different dummies were created to indicate the dwelling classification.

4.2 Statistical Data Analysis

The results of statistical analysis of the dataset are presented in this section. The analysis was performed after the dataset was cleaned as described in section 4.1.

Figure 3. shows the average transaction prices from 1998 to 2013. It appears that the average price in the control area is higher than in the earthquake area. However, the prices in this figure are not corrected for housing characteristics like quality, type, and size. The peak of average transaction price in 2007 and the downward trend afterwards can be clearly observed on the graph below.

Figure 3. Average transaction price in control area and earthquake area

In Table 3. the number of transactions is displayed by year. A distinction is made between control and earthquake area. Because the earthquake area is smaller than the control area, the number of transactions in the earthquake area is also smaller than in the control area. What can also be derived from the table is decline in number of transactions since 2008. It was caused by the financial crisis in 2007 followed by the worst economic decline in Dutch history from Q2 2008 till Q3 2009 (CBS, 2010) and difficult years afterwards. Unemployment

0 50000 100000 150000 200000 250000 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13

Average transaction prices

Control area Earthquake area

25 0 500 1000 1500 2000 2500 3000 3500 4000

Number of sales

rates increased from 3.8% in 2007 to 8.3% in 20132. This uncertain economic situation caused a decrease in consumer confidence which affected housing markets in a negative way.

Table 3. Number of transactions in dataset

Number of transactions in dataset

Total Earthquake area Control area

transactions in 1998 2809 623 2186 transactions in 1999 2883 706 2177 transactions in 2000 2629 643 1986 transactions in 2001 2724 701 2023 transactions in 2002 2774 769 2005 transactions in 2003 2794 706 2088 transactions in 2004 2945 773 2172 transactions in 2005 3264 958 2306 transactions in 2006 3569 960 2609 transactions in 2007 3475 1044 2431 transactions in 2008 2964 850 2114 transactions in 2009 2257 674 1583 transactions in 2010 2285 636 1649 transactions in 2011 2060 587 1473 transactions in 2012 1932 550 1382 transactions in 2013 1832 473 1359

Total number of transactions 43196 11653 31543

Figure 4. Shows the number transactions displayed in the table above in graphical form.

2

http://statline.cbs.nl/StatWeb/publication/?DM=SLNL&PA=80479NED&D1=12-13&D2=0&D3=0&D4=64,77,90,103,116,129,142&HDR=T,G2,G1&STB=G3&VW=T Figure 4. Number of transactions in dataset

26

Next, the common trend assumption will be verified on the basis of Figure 5. Due to the fact that the earthquakes started to become serious in 2003, the five years before have been used. Before 2003 small earthquakes occurred but the effects were not felt or almost not felt by people and buildings are not damaged by such small earthquakes. From the graph one can see that it is not unlikely that both areas follow the same trend over time. Based on this observation and the comparison of economic and demographic factors in part 3.2 it is assumed that the common trend assumption discussed in part 3.1 is satisfied. Thus, the diff-in-diff model can be used to study the effect of earthquakes on the number of house sold.

Figure 5. Number of sales in earthquake area and control area before earthquakes occurred

The following graphs and tables present statistics about the housing characteristics. The NVM does not provide definitions of the different characteristics to brokers. Brokers fill in the database according to their own insights. This can lead to discrepancies caused by subjectivity.

Figure 6. displays the breakdown by house type in the earthquake and control area. It shows that the proportion between the different house types is similar in both areas. Because the researched area is non-urban, (see appendix 2), the percentage of residential farms and manor houses is relatively high; 4% and 3% respectively in both the earthquake and control area. 0 200 400 600 800 1000 1200 1400 1998 1999 2000 2001 2002 N u m b e r o f sal e s Year

Number of sales

27

Figure 6. Composition of earthquake area and control area by house type

Another variable used in one of the models is the construction year. Figure 7. shows the relative amount of houses that were built in a specific period. Again, a distinction has been made between earthquake and control area. The percentage if houses built between 1960 and 1989 is remarkably high. This can be explained by the “baby boom” after World War II. In this period there was a strong population growth in The Netherlands including the Northern provinces, see appendix 4. This caused an increase in housing demand. After 2001, a relatively low amount of houses were built. In this period the population stabilized in the research area. In some municipalities the population even declined. See appendix 3 for an overview of the population since 2000.

28

Figure 7. Construction period by area

Table 4. shows the average and standard deviation of lot sizes, volume, and number of rooms by area. It can be observed from this table that the average lot size in the earthquake area is 622 square meters with a standard deviation of 955.16. For the control area the average is 654.18 with a standard deviation of 1,059.76. It is remarkable for both areas that the standard deviation is larger than the average. This is caused by dwellings with large lots. These homes cause the distribution of this variable to have a long right tail. An additional reason is the fact that the sample is highly diversified. Another variable that can be seen in the table is volume. The average volume in the earthquake area is 407.92 cubic meters with a standard deviation of 182.61 cubic meters. For the control area it is 405.51 cubic meters with a standard deviation of 192.78. Lastly, the average number of rooms is 4.68 with a standard deviation of 1.15 in the earthquake area and 4.73 with a standard deviation of 1.17 in the control area.

Table 4. Descriptive statistics of lot size, volume and number or rooms

Earthquake area Control area Complete area

Average Std. Deviation Average Std. Deviation Average Std. Deviation Lot size 622.08 955.16 654.18 1,059.76 645.52 1,032.68 Volume 407.92 182.61 405.51 192.78 406.16 190.09 Number of rooms 4.68 1.15 4.73 1.17 4.71 1.17 0 5 10 15 20 25 30

Construction period

29

Figure 8. displays information about parking facilities. Both areas are very similar. The figure shows that 85% of the houses have a garage, which is typical for non-urbanized areas.

Figure 8. Composition of earthquake area and control area by parking facility

Table 5. shows the statistical analyses of the remaining housing characteristics. The first variable is the dummy for monument. In the earthquake area 0.79% of the dwellings are monuments. In the control area this is 0.47%.

35% 3% 3% 50% 2% 7%

Earthquake area

Carport and no garage Garage and no carport Garage and carport Garage for multiple cars

34% 3% 3% 49% 3% 8%

Control area

Carport and no garage Garage and no carport Garage and carport Garage for multiple cars

30

Furthermore, garden quality is part of the hedonic regression. Statistical analysis show that in the earthquake area, 35.71% of the houses sold have a good garden. In the control area 31.78% of the houses sold are identified with having a good garden.

Table 5. Statistical analyses of monument and garden quality variables Earthquake area Control area

Yes No Yes No

Monument 0.79% 99.21% 0.47% 99.53% Good Garden 35.71% 64.29% 31.78% 68.22%

31

4.3 Empirical Analysis

It is the attempt of this thesis to determine the influence of induced earthquakes on housing markets. To answer this question, two sub questions were formulated. The previous section described the models and the data that were used to find an answer to these questions. The results of the two models are displayed and interpreted in this section.

4.3.1 Results of the Hedonic Pricing Model

This part analyses the influence of induced earthquakes on housing prices by using the hedonic regression model described before. The model is used to test the following null hypothesis:

H0: Induced earthquakes in “Het Groningerveld” do not lead to a decrease in house values.

According to economic theory one would expect that induced earthquakes influence house values. If that is the case, the null hypothesis should be rejected. The regression analyses are based on 42,240 dwellings sold from 1998 to 2014. The regression table consists of coefficients, significance levels, and robust standard errors. All of these figures are the result of the regression from the logarithmic transaction prices on different determinants of value. The model contains a constant and the coefficients of main interest are 64 “earthquake*quarter” dummies. In addition to the “earthquake*quarter” dummies the results show 139 control variables. The results contain both numeric and dummy variables. The following numeric variables are used in the regression, the first two of which are in logarithmic form:

- Lot size - Volume

- Number of rooms

The other variables are dummy variables: - Earthquake*transaction quarter - Exterior maintenance

- House type - Parking

32 - Monument

- Maintained garden - Zip code

- Transaction quarter

The above variables were previously described in section 3. To avoid the dummy variable trap (multicollinearity) one term of every dummy variable group is left out of the regression. Once it is removed, it becomes the reference value. For the variables “earthquake* transaction quarter” and “transaction quarter,” “2014Q1” is left out. For “exterior maintenance” “poor – moderate poor – moderate” is removed. Another dummy variable is “house type,” and “simple” is excluded from the regression. For the variable “parking,” “no parking space” is left out from the model and therefore the reference. “Construction year” is another binary variable and “after 2000” is removed from the regression. The last dummy variable not included in the regression is the zip code “994E,” which is the zip code located in the earthquake area.

The regression results for both models are presented in tables. The result of the hedonic pricing model is presented in Table 6. The asterisks next to the coefficients (***, ** and *) are indicators for statistical significance: a level of 1%, 5%, and 10% respectively. When a variable is at a significant level of 1% it means that, in repeated samples, the conclusion will be correct in 99% of the cases.

The obtained results show that the null hypothesis cannot be rejected in this model. The regression results in Table 6. show no significant coefficients for the estimators “Eartq Area*Quarter” in the period 2003-2013 (the period where houses in the earthquake area were exposed to earthquakes). The estimators of the interaction terms in the period 1998-2002 are generally insignificant. This is consistent with the expectations because no earthquakes occurred in that period. As explained in section 3.2 the earthquake and control area should be the same. If this is the case, the interaction terms should not show any significant results. However, “Earthq Area*2000Q1” is the only exception. The estimator shows a significant decline in house prices in the earthquake area with respect to the control area. The reasoning for this has not been found. This significant decline is likely caused by a temporary effect.

The control variables have the estimators as expected. All housing characteristics have a significant estimator for a confidence level of 99%. A single increase in volume and lot size

33

increase the value of a house (by 0.21% per m2 and 0.16% per m3 respectively). The condition of exterior maintenance increases the value of home if it is well maintained. Parking is another significant control variable. The results show that better parking facilities have a positive effect on a home’s value. The construction period is also a factor that influences house values. For example, construction before 2000 has a negative effect on the value of a house. Newer homes are usually of better quality, have better insulation, and require less maintenance. This is displayed in the prices. The results show a steady decrease in price from 2000 backwards until the construction period 1931-1944. Houses built in this period have a higher value, which can be explained by the vintage effect. The vintage effect is described by Goodman & Thibodeau (1995) as the effect that “occurs when some unmeasured housing characteristic is correlated with the year that a dwelling was built.” Here, it is probably the building style that people like and are willing to pay for. If a dwelling is a monument it has a positive influence on the house value. This also applies to houses with a good quality garden.

The model also includes 50 dummies for zip code and 64 dummies for quarter. The dummies for zip code are included to correct for the local character of the housing market. Quarter dummies are included to correct for value changes over time. For sake of readability the estimators of the zip code and quarter dummies are not included in the table below. They can be found in appendix 5.

The regression results show an R-squared of 0.830. The model explains 83% of the variance in the dependent variable.

Robust Robust

Variables coef S.E. Variables coef S.E.

Earthq Area*1998q1 -0.039 0.037 Earthq Area*2009q3 0.006 0.036 Earthq Area*1998q2 -0.025 0.036 Earthq Area*2009q4 0.01 0.034 Earthq Area*1998q3 -0.037 0.036 Earthq Area*2010q1 -0.024 0.035 Earthq Area*1998q4 -0.042 0.035 Earthq Area*2010q2 0.003 0.034 Earthq Area*1999q1 -0.009 0.036 Earthq Area*2010q3 0.004 0.035 Earthq Area*1999q2 -0.005 0.035 Earthq Area*2010q4 0.002 0.035 Earthq Area*1999q3 0.011 0.035 Earthq Area*2011q1 -0.006 0.037 Earthq Area*1999q4 -0.011 0.036 Earthq Area*2011q2 0.009 0.037 Earthq Area*2000q1 -0.015 0.035 Earthq Area*2011q3 0.003 0.037 Earthq Area*2000q2 -0.009 0.035 Earthq Area*2011q4 0.008 0.038 Earthq Area*2000q3 -0.031 0.039 Earthq Area*2012q1 -0.024 0.039 Earthq Area*2000q4 -0.073 ** 0.035 Earthq Area*2012q2 0.03 0.038 Earthq Area*2001q1 -0.018 0.035 Earthq Area*2012q3 0.058 0.036 Earthq Area*2001q2 -0.036 0.034 Earthq Area*2012q4 0.016 0.036 Earthq Area*2001q3 -0.038 0.035 Earthq Area*2013q1 0.051 0.042 Earthq Area*2001q4 -0.018 0.035 Earthq Area*2013q2 0.013 0.038 Earthq Area*2002q1 -0.006 0.036 Earthq Area*2013q3 -0.019 0.038 Earthq Area*2002q2 -0.049 0.034 Earthq Area*2013q4 0.051 0.037 Earthq Area*2002q3 -0.008 0.035 ln(lot size) 0.214 *** 0.003 Earthq Area*2002q4 -0.047 0.034 ln(volume) 0.156 *** 0.008 Earthq Area*2003q1 -0.018 0.034 Ext. maint. Moderate-good/not fulfilled 0.187 *** 0.01 Earthq Area*2003q2 -0.016 0.034 Ext. maint. Good-Excellent 0.335 *** 0.01 Earthq Area*2003q3 -0.019 0.034 Number of rooms 0.037 *** 0.001 Earthq Area*2003q4 -0.006 0.034 Single family 0.128 *** 0.005 Earthq Area*2004q1 -0.020 0.034 Town house 0.283 *** 0.006 Earthq Area*2004q2 0.026 0.034 Residential farm 0.228 *** 0.009 Earthq Area*2004q3 0.005 0.034 Bungalow 0.308 *** 0.008 Earthq Area*2004q4 -0.008 0.036 Manor house 0.398 *** 0.009 Earthq Area*2005q1 0.023 0.036 Parking lot 0.048 *** 0.006 Earthq Area*2005q2 0.017 0.035 Carport and no garage 0.081 *** 0.005 Earthq Area*2005q3 0.006 0.033 Garage and no carport 0.107 *** 0.003 Earthq Area*2005q4 -0.004 0.033 Garage and carport 0.13 *** 0.006 Earthq Area*2006q1 0.028 0.035 Garage for multiple cars 0.125 *** 0.005 Earthq Area*2006q2 0.014 0.033 Construction year 1500-1905 -0.278 *** 0.009 Earthq Area*2006q3 -0.003 0.034 Construction year 1906-1930 -0.284 *** 0.008 Earthq Area*2006q4 0.013 0.034 Construction year 1931-1944 -0.252 *** 0.008 Earthq Area*2007q1 -0.002 0.035 Construction year 1945-1959 -0.317 *** 0.008 Earthq Area*2007q2 0.007 0.033 Construction year 1960-1970 -0.313 *** 0.008 Earthq Area*2007q3 -0.020 0.036 Construction year 1971-1980 -0.261 *** 0.007 Earthq Area*2007q4 0.018 0.033 Construction year 1981-1990 -0.203 *** 0.008 Earthq Area*2008q1 0.002 0.033 Construction year 1991-2000 -0.084 *** 0.008 Earthq Area*2008q2 0.012 0.034 Monument 0.109 *** 0.021 Earthq Area*2008q3 -0.013 0.034 Good garden 0.078 *** 0.002 Earthq Area*2008q4 -0.010 0.035 Constant 8.913 *** 0.050 Earthq Area*2009q1 0.009 0.035 Observations 42,240

Earthq Area*2009q2 -0.007 0.035 R-squared 0.830

***= significant at 1% level **= significant at 5% level *= significant at 10% level Table 6. Results of the hedonic pricing model

35

4.3.2 Results of the Difference-in-Difference Model

This section analyses the influence of induced earthquakes on the number of sales by using the diff-diff model described in section 3.4. The model is used to test the following null hypothesis:

H0: The induced earthquakes in “Het Groningerveld” do not lead to a decrease in number of sales.

According to the theory described in Section 2, one would expect that induced earthquakes influence the number of houses sold. If that is the case, the null hypothesis should be rejected based on the regression results.

The regression analyses are based on 60 quarters - the first quarter of 1999 to the fourth quarter of 2013. The regression results displayed in Table 7. consist of coefficients, significance levels, and robust standard errors. The figures in Table 7. are the result of the regression from the logarithmic transaction prices on the “Area” dummy, “period 2013” dummy, “Area * Period 2013” interaction term, and a constant. The dummy “area” has value of 1 if the sales took place in the earthquake area, and 0 if the sales took place in the control area. The “period 2013” dummy has value 1 if the sales took place in the year 2013, 0 otherwise. “Area * Period 2013” is the interaction term of “Area” and “period 2013”. Thus, this interaction term will only have value 1 if the number of sales took place in the earthquake area in 2013.

The period 2013 was used because, as can be seen in Figure 1, 2013 was the year in which the most earthquakes occurred. It was also the period where the earthquakes had the highest magnitude. If one could expect an effect, it would be in this year. Longer periods are researched in the robustness checks.

The null hypothesis cannot be rejected in this model. The regression results in Table 7. show no significant coefficients for the estimator of interest “Area * Period 2013”.

The variable “area” has a significant coefficient of -0.978, which means that 97.8% fewer houses were sold in the earthquake area compared to the control area. This can be explained by the fact that the earthquake area is much smaller than the control area. The coefficient of the other variable, “period 2013,” displays a value of -0.406, which implies that 40.6% fewer houses were sold in 2013 compared to the entire period before. Both estimators and the constant show a significant result for a significant level of 1%.

36 Table 7. Results of diff-in-diff model of 2013 period

ln(Sales) Robust

Variables coef S.E.

Area * Period 2013 -0.078 0.216 Area -0.978 *** 0.039 Period 2013 -0.406 *** 0.152 Constant 6.195 *** 0.027 Observations 120 R-squared 0.849 ***= significant at 1% level **= significant at 5% level *= significant at 10% level

Although it is less obvious in the studied period above, other years are studied as thoroughly. This is done because, theoretically, one could expect a drop in number of sales in a certain year. Consequently followed by a drop house prices, which causes the number of sales to move back to the original level. Although, it is already concluded that there is no significant decline house prices caused by earthquakes, it is still interesting to investigate this possibility.

The regression was run for every single year from 2003 until 2012. The regression results do not show a significant results for the “Area * Period” estimators. This means that also here the null hypothesis cannot be rejected and that there is no evidence that the earthquakes influences the number of houses sold.

37

4.4 Robustness Checks 4.4.1 Hedonic Pricing Model

The hedonic pricing model was checked to see if the sample used for the regression is homogenous. Homogeneity means that the statistical properties of one part of the dataset is the same as every other part. If this is the case, the regression results are not influenced by a few extreme observations (outliers). Homogeneity makes it more likely that the sample is a good representation of the whole population.

As mentioned in part 4.1, the sample data used in this research comes from the NVM. Approximately 70% of all home sale transactions are included in the NVM’s database, meaning the sample used for this thesis is around 70% of the total population. To check whether the regression results are influenced by outliers, a random sample of 10,000 observations were obtained from the NVM dataset. Next, the regression of equation 1 was performed, but this time with the random sample data.

The results of this regression are generally the same as the results in the original model. The regression results shown in Table 8. display only slight changes. Some of the estimators of the control variables such as monument, zip codes, and quarters became less significant because 10,000 observations were used instead of the 42,240 that were used in the original model. When fewer observations are used the power of the model decreases. Since the results of regression with data from the random sample are not significantly different from those of the original model, it is likely that the dataset used is homogenous. For sake of readability the estimators of the zip code and quarter dummies are not included in the table below. They can be found in appendix 6.

Robust Robust

Variables coef S.E. Variables coef S.E.

Earthq Area*1998q1 -0.037 0.075 Earthq Area*2009q3 -0.123 0.077 Earthq Area*1998q2 -0.065 0.071 Earthq Area*2009q4 -0.009 0.068 Earthq Area*1998q3 -0.005 0.075 Earthq Area*2010q1 0.057 0.075 Earthq Area*1998q4 -0.064 0.072 Earthq Area*2010q2 -0.016 0.071 Earthq Area*1999q1 -0.076 0.076 Earthq Area*2010q3 -0.06 0.07 Earthq Area*1999q2 -0.027 0.07 Earthq Area*2010q4 -0.011 0.07 Earthq Area*1999q3 0.013 0.07 Earthq Area*2011q1 0.002 0.074 Earthq Area*1999q4 -0.024 0.073 Earthq Area*2011q2 -0.022 0.081 Earthq Area*2000q1 -0.036 0.072 Earthq Area*2011q3 0.01 0.072 Earthq Area*2000q2 -0.037 0.071 Earthq Area*2011q4 -0.019 0.079 Earthq Area*2000q3 -0.086 0.072 Earthq Area*2012q1 -0.047 0.084 Earthq Area*2000q4 -0.065 0.069 Earthq Area*2012q2 0.034 0.078 Earthq Area*2001q1 0.028 0.069 Earthq Area*2012q3 -0.027 0.074 Earthq Area*2001q2 -0.023 0.069 Earthq Area*2012q4 0.009 0.07 Earthq Area*2001q3 -0.062 0.071 Earthq Area*2013q1 0.074 0.09 Earthq Area*2001q4 -0.046 0.073 Earthq Area*2013q2 0.035 0.082 Earthq Area*2002q1 0.018 0.097 Earthq Area*2013q3 0.001 0.075 Earthq Area*2002q2 -0.035 0.068 Earthq Area*2013q4 0.025 0.077 Earthq Area*2002q3 -0.035 0.07 ln(lot size) 0.211 *** 0.005 Earthq Area*2002q4 -0.042 0.069 ln(volume) 0.177 *** 0.017 Earthq Area*2003q1 -0.067 0.067 Ext. maint. Moderate-good/not fulfilled 0.196 *** 0.021 Earthq Area*2003q2 -0.027 0.069 Ext. maint. Good-Excellent 0.345 *** 0.021 Earthq Area*2003q3 0.015 0.069 Number of rooms 0.034 *** 0.003 Earthq Area*2003q4 -0.04 0.07 Single family 0.135 *** 0.01 Earthq Area*2004q1 -0.061 0.072 Town house 0.294 *** 0.013 Earthq Area*2004q2 0.003 0.072 Residential farm 0.238 *** 0.02 Earthq Area*2004q3 0.041 0.071 Bungalow 0.345 *** 0.016 Earthq Area*2004q4 0.002 0.072 Manor house 0.419 *** 0.018 Earthq Area*2005q1 0.017 0.068 Parking lot 0.043 *** 0.013 Earthq Area*2005q2 0.035 0.081 Carport and no garage 0.103 *** 0.012 Earthq Area*2005q3 -0.02 0.066 Garage and no carport 0.11 *** 0.006 Earthq Area*2005q4 -0.013 0.067 Garage and carport 0.129 *** 0.012 Earthq Area*2006q1 -0.024 0.081 Garage for multiple cars 0.13 *** 0.01 Earthq Area*2006q2 0.002 0.067 Construction year 1500-1905 -0.274 *** 0.016 Earthq Area*2006q3 0.004 0.07 Construction year 1906-1930 -0.293 *** 0.014 Earthq Area*2006q4 0.008 0.068 Construction year 1931-1944 -0.254 *** 0.014 Earthq Area*2007q1 -0.025 0.077 Construction year 1945-1959 -0.313 *** 0.015 Earthq Area*2007q2 0.008 0.069 Construction year 1960-1970 -0.313 *** 0.013 Earthq Area*2007q3 -0.054 0.069 Construction year 1971-1980 -0.266 *** 0.012 Earthq Area*2007q4 0.002 0.067 Construction year 1981-1990 -0.209 *** 0.013 Earthq Area*2008q1 0.015 0.067 Construction year 1991-2000 -0.089 *** 0.013

Earthq Area*2008q2 0.005 0.069 Monument 0.085 * 0.044

Earthq Area*2008q3 0.012 0.067 Good garden 0.075 *** 0.004 Earthq Area*2008q4 -0.049 0.074 Constant 8.807 *** 0.104 Earthq Area*2009q1 -0.014 0.068 Observations 10,000

Earthq Area*2009q2 -0.068 0.076 R-squared 0.826

***= significant at 1% level **= significant at 5% level *= significant at 10% level Table 8. Results of the robustness check on the hedonic pricing model