Ripple effects and long-run convergence in the Dutch housing
market
Name: Tjerk Koster (10618279)
Supervisor: Martijn Dröes
Date: 27-06-18
Document: Master’s Thesis Finance and Real Estate Finance
Faculty: Economics & Business (University of Amsterdam)
Statement of originality
This document is written by Student Tjerk Koster who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Abstract
This study examines the presence of long-run convergence and ripple effects in the Dutch housing market. Existing studies do not provide a consistent answer whether ripple effects and long-run convergence exist, in housing markets. Besides, most of the earlier performed studies focussed on the UK housing market. Two beta-convergence tests are performed to test for long-run convergence, using data from 1995-2017. Furthermore, a vector autoregressive model (VAR), combined with Granger causality tests is used to test for relationships between municipalities, using data from 1990-2017 on three different sub-samples. These three different municipal sub-samples are used to test for countrywide and regional ripple effects, within the Dutch housing market. The outcomes of the different tests suggest that long-run convergence in relative house prices and countrywide and regional ripple effects are present, in the Dutch housing market. Besides, this study finds evidence that ripple effects occur due to “spatial patterns in house price determinants”, as mentioned by Meen (1999).
Table of Content
1. Introduction ... 5
2. Literature review ... 9
2.1 Current Dutch housing market ... 9
2.2 Investment in the Dutch housing market ... 10
2.3 Convergence in house prices ... 11
2.4 Ripple effect ... 12
2.5 Ripple effect and Convergence previous research ... 12
2.5.1 Ripple effect and convergence (UK) ... 12
2.5.2 Ripple effect and convergence (outside UK) ... 14
2.6 Reasons for a ripple effect ... 15
3. Research hypotheses & Data description ... 17
3.1 Hypotheses ... 17
3.2 Data ... 18
3.2.1 NVM data ... 18
3.2.2 Other data sources ... 20
4. Methodology... 20
4.1 Beta-Convergence (Proportions) ... 21
4.2 Beta-Convergence (Pooled OLS)... 23
4.3 Vector autoregressive model (VAR) ... 24
4.4 Robustness test (VAR)... 26
5. Results ... 26
5.1 Beta-Convergence (Proportions) ... 26
5.2 Beta-Convergence (Pooled OLS)... 27
5.3 Results provinces (VAR) ... 28
5.4 Results Amsterdam region (VAR) ... 33
5.5 Results Eindhoven region (VAR) ... 37
4.6 Robustness tests VAR models ... 41
5.6 Potential reasons for the ripple effect ... 42
6. Conclusion & Discussion ... 43
Reference list ... 48
1.
Introduction
In 2014, after the recent financial crisis, the Dutch housing market started to recover (NRC, 2015). However, the house price increase, recently observed in the Dutch housing market, is not evenly divided over the Netherlands with certain municipalities like Amsterdam and Utrecht, exhibiting a more significant price increase than others (Savills, 2017). This raises the question whether house price development follows a specific spatial pattern over time. As housing is the largest financial assets of households, understanding the predictors and the movements of house prices is extremely relevant (CBS, 2009). This leads to the following research questions: Is there long-run convergence and
are there ripple effects present in the Dutch housing market? What causes these ripple effects and where do they start?
Most of the previous research on ripple effects is performed on the housing market of the United Kingdom. In the UK, house prices tend to follow a spatial pattern over time (Meen, 1999). This spatial pattern is referred to as the ‘ripple effect’. House price developments seem to start in certain regions before spreading out to other regional housing markets through distance and time, according to Meen (1999). This implies that house prices developments can be significantly different between regions in the short-term but in the long-run normal relative prices would restore towards the earlier equilibrium.
However, the outcomes of the studies on ripple effects are not found to be consistent, as can be observed in Section 2.5. Several studies found evidence of countrywide ripple effects originating in the more developed regions within countries (Giussani & Hadjimatheou, 1991; MacDonald & Taylor, 1993; Nanda & Yeh, 2014). However, others found evidence for segmented ripple effects opposed to countrywide effects, while a different group of studies found no evidence for ripple effects (Rosenthal, 1986; Drake, 1995; Abbott & De Vita, 2013; Ashworth & Parker, 1997). In addition, economic theory suggests that ripple effects should not exists in housing markets (Canarella, Miller & Pollard, 2011). As houses are non-moveable assets, prices should be determined by supply and demand in the individual regional markets, according to Canarella et al. (2011).
Ripple effects in housing markets have been discussed and studied for multiple decades. In addition to the already extensive amount of literature on the ripple effect, this study adds to the current literature in multiple ways. First, this study will observe if there are ripple effects present in the Dutch housing market. As most of the previous studies are performed on the UK housing market the current literature on the Dutch housing market is still limited. As the Netherlands is a relatively small country with a high density, the housing price development could be different compared to other studied countries.
Furthermore, the data this study employs is more extensive than data used in many of the other studies. The current literature mainly looks at the ripple effect between metropolitan statistical areas (MSA’s) within countries. As the Netherlands is a rather small country, it could be compared with a metropolitan statistical area (MSA), within countries like the UK. Consequently, the results of this study can be used to give an interpretation of ripple effects within MSA’s, rather than between them. Besides, this study employs data including more than 300 Dutch municipalities. Therefore, this study includes more observations and these observations are on a lower level (municipal) than most previous studies. This makes that, the current study reviews the ripple effect in another dimension and includes a larger sample size than most of the previous studies.
Besides, the reasons for the ripple effect are still unclear and subject to debate. Potential reasons are mentioned in the literature but rarely tested. This thesis will, with the use of three sub-samples and different convergence tests, try to identify a possible justification for the presence of a ripple effect in the Dutch housing market. We focus on: Amsterdam and provincial capitals, the direct Amsterdam region and Eindhoven including surrounding municipalities.
Although, the Dutch housing market differs from the UK and other earlier studied markets, we expect to find evidence for a ripple effect. Especially internal migration, as mentioned by Meen (1999), is expected to play a more significant role, as the Netherlands is relatively small and distances between the regions are modest. Therefore, the Netherlands is expected to exhibit a relative large amount of non-job-related migration.
Next to this, the Netherlands has only one major economic centre, the Randstad, including Amsterdam. “Spatial patterns of house price determinants”, as mentioned by Meen (1999), suggests that this region benefits from economic growth first. Therefore, housing prices in the Randstad are expected to move first, followed by house prices in other regions. Economic downturns are also expected to have a more significant impact on the more developed regions, creating the opportunity of prices to converge back to the initial ratio. Therefore, these two potential reasons for the ripple effect, mentioned by Meen (1999), suggest that a ripple effect could be expected in the Netherlands.
This study employs data received from the NVM and some smaller data sources. The NVM data consists of the annual square meter prices for all municipalities in the Netherlands between 1990 and 2017 and enables us to determine yearly house price developments. These yearly percental house price changes could be used to test how house price shocks spread over the country and the regional housing market. This could then be used to test for the presence of ripple effects and where the effects start. The smaller data sources are used to retrieve countrywide factors like the gross domestic product (GDP) and the consumer price index (CPI).
To test for long-run convergence two different models are employed. First a beta-convergence model, based on Cook (2012), is performed. All municipalities are distributed over four different groups. These groups are based on the initial square meter prices and average growth over the entire sample period (1990-2017). This provides a rather simple but innovative way to test for long-run convergence. The second convergence model, will perform a pooled OLS regression with dependent variable house price change. This regression will include the initial square meter value and multiple control variables. This model can be used to test whether divergence and convergence in house prices occur, based on the initial value, and what causes house price change. In order to test for the actual ripple effect and interrelations between municipalities the vector autoregressive model (VAR) is used. The VAR model, combined with Granger causality tests, will be used on the three sub-samples, described earlier. The subsamples enable us to test for a countrywide ripple effect and ripple effects within the smaller regions. Furthermore, these VAR model outcomes can be used to observe which municipalities are leading the effect.
This study finds evidence for long-run convergence (or constancy) of relative Dutch house prices. Relative house prices are not found to diverge over the total sample period 1990-2017. However, evidence for divergence is found in market upswings and for convergence during downswings. This provides evidence for long-run convergence of relative house prices and the short-term possibility for relative house prices to diverge. Another convergence model, including multiple control variables, does not find evidence for divergence of relative prices over the entire sample period based on initial square meter prices. Therefore, house price change seems to be independent of initial square meter price at the start of the sample period.
The vector autoregressive model (VAR), used in this study, finds multiple lead-lag relationships in the different studied regions. Amsterdam is found to be leading the countrywide effect, forecasting house price development in six of the total of twelve provincial capitals. The relationships found also indicate an outward trend with cities within the Randstad forecasting house prices further away. However, contradictory to the ripple effect, some lead-lag relationships are found to lead inwards, with municipalities closer to Amsterdam reacting to shocks originating in Amsterdam later than some municipalities further away.
Furthermore, in the Amsterdam region, Amsterdam and surrounding municipalities like Diemen and Amstelveen, show to be leading. Amsterdam cannot be forecasted by other regions possessed forecasting power over 9 of the 19 surrounding municipalities. The general trend also moves outwards from Amsterdam, although some municipalities, like Haarlem, seem to react faster.
The results for Eindhoven show a comparable trend. Eindhoven seems to lead the surrounding housing market as well. Lagged values of other municipalities on Eindhoven are not found to be significant and Eindhoven predicts multiple future price changes in other adjacent municipalities. Also, in this sub-sample, some of the surrounding municipalities react more rapidly than others.
The outcomes of the models used in this study show strong support and evidence for “spatial pattern of determinants of house prices change”, mentioned by Meen (1999), as a possible reason for the ripple effect. Economic development and housing market characteristics are believed to explain the ripple effect found in the Dutch market. Cities and regions with more developed economies and attractive housing markets are found to lead the effect in all three sub-samples. Besides, evidence for the two necessary characteristics, mentioned by Meen (1999), is found.
Understanding the dynamics in house price developments proves to be highly relevant. Houses tend to be the largest financial assets of households and have significant impact on the economy (CBS, 2009; Bank of England, 2017). Developments in house prices are linked with consumer confidence, construction and economic growth (Bank of England, 2017). When house prices increase, homeowners become more confident and are able to borrow more due to increased collateral value, according to the Bank of England (2017). However, the opposite occurs when they decrease (Bank of England, 2017).
As economic theory would oppose the ripple effect, understanding the movements and predictors of price change is extremely relevant for policymakers and households (Canarella et al., 2011). Policymakers could have the possibility to predict in which regions prices are expected to increase in the near future. This could give them the possibility to act on these forecasts, by for instance: granting building permits in certain regions to limit price growth.
Besides, the ability to predict house price developments could prove to be useful for investors. Lead-lag relationships could be used for investment allocation purposes, as previous house price developments could indicate current or future price movement in other municipalities. This information could help investors decide in which markets investment would be profitable.
Knowledge about spatial patterns in house prices can be useful for developers as well. They will be able to forecast which regions are expected to exhibit significant price growth. This information may benefit the timing of developments and enables the building process to start earlier. This may result in lower building costs as land and development costs will still be relatively cheap. Consequently, after the expected price increase occurs, the developer will benefit from the increased square meter prices.
2.
Literature review
This literature review starts with a description of the current state of the Dutch residential market and the extent of residential investments. Afterwards, both housing market convergence and the ripple effect will be introduced. Furthermore, previous research on long-run convergence and ripple effects in other housing markets will be discussed.
2.1
Current Dutch housing market
The current Dutch housing market is characterized by substantial price increase and tightness in, primarily, the Randstad area (Rabobank, 2018). After the recent financial crisis, the Dutch housing market recovered and exhibits significant house price increase since 2014 (CBS, 2015). The largest growth in house prices can be observed in the Randstad area which includes major cities like Amsterdam, Rotterdam, The Hague and Utrecht (Savills, 2017). Nonetheless, recently other provinces exhibit a significant increase in transaction volume and transaction prices as well (Rabobank, 2018).
The Dutch house price increase, in 2017, was one of the highest in Europe, with only Portugal and Ireland exhibiting higher growth (NRC, 2018). The current house price development, can be explained by several different factors. First, after the financial crisis the economy recovered, with increasing consumer confidence, income and overall economic growth (ING, 2018). On top of this, the low interest rates make housing relatively affordable and attractive for households (Rabobank, 2018). Furthermore, housing construction in the Netherlands stagnated in the financial crisis, causing housing shortage that still endures in the current housing market (NRC, 2018). Although, the construction of residential properties is currently growing, it still has not increased enough to close the shortage caused during the recent financial crisis (Rabobank, 2018; ING, 2018). Therefore, tightness remains in the Dutch housing market and is expected to maintain the upcoming years (Rabobank, 2018; ING, 2018).
Consequently, transaction prices started to rise in other Dutch regions, outside the Randstad, as well (Rabobank, 2018). This affects especially provinces close to the Randstad area, like Flevoland and Gelderland (Rabobank, 2018). According to the Rabobank (2018) these provinces also display a higher increase in transaction volume compared to the Randstad area. This could be explained by the tightness in the Randstad’s housing market combined with the relative high supply in other provinces (Rabobank, 2018; NVM, 2017a). In the major cities like Amsterdam, Utrecht and Rotterdam the number of available houses for prices under EUR 200.000 decreased
heavily (NVM, 2017a). This decrease went up to 67 percent in Amsterdam, 40 percent in Utrecht and 30 percent in Rotterdam, according to the NVM (2017a).
Therefore, in order to find affordable housing, households have to search in regions outside the Randstad, causing transactions and demand to increase in these sub-markets (NVM, 2017a; Savills, 2017). According to CBS (2017), especially young families with new-born children leave the cities relatively often. This could be explained by the high transaction prices combined with lower possible mortgage loan-to-value in the Netherlands (NRC, 2018). This movement or internal migration explains the increase in transaction volume in the regions outside Zuid-Holland, Noord-Holland and Utrecht. (Rabobank, 2018). This internal migration leads to higher transaction volumes and prices in other provinces (Rabobank, 2018; Savills, 2017). Nonetheless, this price increase primarily takes place in the major cities within these provinces outside the Randstad (Savills, 2017).
2.2
Investment in the Dutch housing market
Investment in the Dutch housing market is increasing in recent years as well. Both private and institutional investors display increasing interest in the Dutch housing market (ABN AMRO, 2018; NVM, 2017b; Capital Value, 2018). The Ministry of the Interior and Kingdom Relations (2014), reports that the Dutch housing market is interesting for investors, possessing strong fundamentals. According to ABN AMRO (2018), the increase in residential investments, contributes to the house price increase currently observed in the Dutch housing market. In 2016, residential investments were the second largest investment category in the Netherlands, totalling EUR 3.2 billion (JLL, 2017).
According to ABN AMRO (2018), the increase in private investors can be explained by the low interest rates and the beneficial fiscal treatment of residential properties. The number of private investors that possess between three and fifty houses has doubled in the Netherlands over the past decade (NVM, 2017b). These private investors are, according to the NVM (2017b), mainly active in cities like Amsterdam, Groningen and Maastricht.
Besides, institutional investors are increasingly interested in the Dutch housing market as well (Capital Value, 2018; ABN AMRO 2018). Foreign and domestic pension funds are attracted by the low risk-return profile in this market (Capital Value, 2018). According to Capital Value (2018), foreign investors invested EUR 950 million in 2017. Moreover, in 2018, these investors have an expected EUR 2.5 billion available to invest in Dutch housing (Capital Value, 2018). Currently mainly German, North-American and English institutional investors are active in the
Dutch housing market (Capital Value, 2018). On top of this, according to Capital Value (2018), Asian investors are increasingly interested in the Dutch housing market as well.
2.3
Long-run convergence in house prices
Long-run convergence in house prices is the tendency for relative regional house prices to exhibit a constant equilibrium price ratio, compared to the countries average price level in the long-run (Holmes & Grimes, 2008). This should be considered as converging towards earlier equilibrium levels after short-term divergence or convergence in house prices has taken place. Therefore, in the short-run regional house prices are able to deviate substantially (Meen, 1999). However, in the long-run interregional house prices move proportionally, staying around a stationary ratio between the regional and the national house price, according to Meen (1999).
Different hypotheses on the possibility of long-run convergence were created. The divergence hypothesis states that interregional relative price differences increase over time (Abbott & De Vita, 2013). Another line of studies, mentioned by Abbott and De Vita (2013), emphasize that there is long-run convergence in the interregional housing markets (convergence hypothesis). The cyclical gap hypothesis, states that price gaps increase when the market is in an upswing but decrease when the market experiences a downswing (Abbott & De Vita, 2013).
However, the outcomes of these studies do not prove to be consistent. In the United Kingdom, Holmes and Grimes (2008) discovered a long-run stability in relative house prices between all studied regions. Cook (2005) found evidence for long-run convergence in the UK housing market, as an extensive amount of cointegration relationships were found between different regions. However, MacDonald and Taylor (1993), found evidence for long-run convergence but the multiple cointegrating relationships that were found are accompanied with some evidence for segmentation, within the UK housing market.
Furthermore, Cook (2012), found evidence of long-run convergence, using a beta-convergence model, although beta-convergence did not take place over the entire sample period. Therefore, the initial findings supported the divergence hypotheses instead of the convergence hypothesis (Cook, 2012). However, due to the possibility of cyclical movements, different sub-periods were tested by Cook (2012) as well. This resulted in finding evidence of convergence in market downturns with housing price differences narrowing. These findings support the cyclical gap hypothesis, as mentioned by Abbott and De Vita (2013).
2.4
Ripple effect
The ripple effect refers to a spatial pattern in house price development within a country or region (Meen, 1999). House prices seem to be moving first in certain regions and are followed by others, according to Meen (1999). This implies that house price movements can significantly differ between regions in the short-term. However, in the long-term the house prices will restore toward normal relative prices between regions. This is in line with the long-run convergence described in Section 2.3. Meen (1999) mentions two necessary characteristics of the ripple effect: (1): House prices have to move similarly in the long-term, creating long-run convergence towards the earlier equilibrium, in the interregional housing market. (2) Short-run divergence in interregional house prices exists.
However, as Canarella et al. (2011) state, economic theory should deny the existence of ripple effects in housing markets. As housing is a non-moveable asset, house prices in different regions should reflect the different supply and demand for housing in the individual regions (Canarella et al., 2011). The presence of ripple effects would contradict this theory and suggest that the house prices are not only determined by demand and supply but follow a distinct spatial pattern over time and distance (Canarella et al., 2011).
2.5
Ripple effect and Convergence previous research
Most of the constructed research on ripple effects and long-run convergence in house prices has been performed on the United Kingdom. The following section will also include outcomes of convergence, as long-run convergence is a necessary characteristic of the ripple effect. However, the results of different studies do not seem to be consistent. Although some studies found a ripple effect, others found the effect to be limited or not present. Besides, a majority of different models were used to test for the effect. The current paragraph will focus on the outcomes of prominent researches on the ripple effect in and outside the UK. Both sub-paragraphs will start with earlier research on the ripple effect and convergence, ending with studies performed in recent years.
2.5.1 Ripple effect and convergence (UK)
In the UK, Giussani and Hadjimatheou (1991) found that relative house price differences between Northern and Southern regions increased in the short-run when a strong house price hike was present in the market. However, as the activity in the housing market decreased, the relative prices
returned to the earlier equilibrium (Giussani & Hadjimatheou, 1991). These findings demonstrate evidence for the presence of a ripple effect. For this research, a Granger causality test combined with cross correlation matrices was used. These findings are supported by the findings of Meen (1999). He found evidence of specific spatial pattern that support the presence of a ripple effect. Both Meen (1999) and Giussani and Hadjimatheou (1991) found evidence that the Greater London area and the South-East overall are leading the transaction price changes in other regions, within the UK.
Moreover, studies by Macdonald and Taylor (1993) and Alexander and Barrow (1994) looked for a ripple effect originating in London. Macdonald and Taylor (1993) tested if house price shocks in London affected transaction prices in other cities or regions within the UK. They found co-integrating relationships between different regions and evidence that the UK house price movement begins in the Greater London area. Alexander and Barrow (1994) found evidence for a ripple effect originating in the South that causes house prices in the Midlands and North to move in the same direction. However, in contrast with the findings of Macdonald and Taylor, London seemed to act like an independent market compared to the other Southern regions.
However, there seems to be evidence for segmentation between Northern and Southern regions as well (Macdonald & Taylor, 1993; Drake, 1995). Drake (1995) found evidence of cointegration between house prices within the South. Nonetheless, a distinction between house prices determination between South-Eastern regions and Northern regions was also found, providing evidence for segmentation (Drake, 1995). This suggests the presence of a ripple effect but within different segmented housing markets (Macdonald & Taylor, 1993; Drake 1995).
Not all research supports the presence of a ripple effect in the UK housing market. Rosenthal (1986), found hardly any interregional relationships between house prices. Besides, the results did not present a spatial pattern across regions (Rosenthal, 1986). Furthermore, according to Rosenthal (1986), London and the South-East region did not display conclusive evidence to lead the other UK regions. Moreover, he did not find reliable evidence that neighbouring regions showed stronger co-integration than distant regions did. As a ripple effect would first affect neighbouring regions before spreading out to regions further away, this should have been expected if a ripple effect existed (Rosenthal, 1986). It should be noted, nonetheless, that this study only included a six-year sample.
Furthermore, Ashworth and Parker (1997) supported the findings of Rosenthal (1996). They did not find compelling evidence for the ripple effect and doubt its existence. Besides, they found a similar timing of lags, of different regions to shocks that started in the South-East. This is
in line with the findings of Rosenthal and denies the presence of a ripple effect, as more distant regions would be expected to react later than adjacent regions (Ashworth & Parker, 1997).
After the twentieth century, there have been multiple researches on the ripple effect in the UK housing market as well. Cook and Thomas (2003) used a non-parametric test along with business cycle dating procedures to test for the ripple effect. They found evidence that support the presence of a ripple effect. Furthermore, Cook (2005) found an extensive amount of cointegrating ties between regions, using threshold autoregressive methods. This large amount of cointegration between regions opposes the results of Drake (1995) and MacDonald and Taylor (1993), that found evidence of segmentation. Holmes and Grimes (2008) also found evidence for convergence in long-run relative prices between all the regions. Moreover, Tsai (2014), found evidence for an equilibrium relationship between regional housing markets and the countrywide market. Tsai (2014) also looked at a ripple effect with the variable transaction volume and expects that without including this variable, other studies underestimate the ripple effect in the housing market. To sum up, Cook (2005), Holmes and Grimes (2008), Cook and Thomas (2003) and Tsai (2014) all support the findings of Meen (1999), suggesting the existence of a ripple effect in the overall UK housing market.
However, Holmes (2007) found evidence of segmentation, thereby supporting the findings of Drake (1995) and MacDonald and Taylor (1993). According to Holmes (2007), three regions within the UK do not display long-run convergence with the overall UK housing market, therefore providing evidence for segmentation.
Moreover, a study by Abbott and De Vita (2013) found no evidence for a ripple effect. They did not find a long-run equilibrium relation between regional house prices. Only two combinations of regions showed some evidence of convergence (Abbott & De Vita, 2013). Therefore, they conclude that differences in house prices among regions are expected to endure in the long-run.
2.5.2 Ripple effect and convergence (outside UK)
Outside the UK several studies have investigated ripple effects in housing markets as well. These studies all took place in the twenty-first century after a large extent of studies had already taken place on the UK housing market. These studies looked at the ripple effect in the national housing markets of the United States, Singapore, Australia, New-Zealand, Ireland and Malaysia.
Some studies provided evidence for the existence of ripple effects and convergence in national housing markets. Luo, Liu and Picken (2007) tested for a house price pattern between the
eight Australian state capital cities, while Kuethe and Pede (2011) tested for the ripple effect in the western part of the United States. Both studies found that in the long-run house prices returned to their equilibrium price ratio (Luo et al., 2007; Kuethe & Pede, 2011). Besides, they found evidence for the existence of spatial patterns within the US and Australian housing markets. According to Luo et al. (2007), this shows a general trend in the long-run. Their study presents, by using a Johansen cointegration test, that there is convergence over the entire Australian housing market accompanied with a spatial diffusion pattern, starting in Sydney. Kuethe and Pede (2011) found, using the spatial VAR model, evidence for the presence of neighbouring spill over effects in the US market. Besides, they found convergence between California and adjacent states. Therefore, both support the findings of Meen (1999).
Stevenson (2004), studied convergence and the ripple effect in Ireland and North-Ireland combined. He found evidence for convergence in the Irish housing market, thereby supporting the findings of Meen (1999). Besides, Stevenson (2004), found evidence for diffusion starting in Dublin, comparable to the effect found in earlier studies in the UK. In addition, the North-Irish housing markets seems to be connected, more heavily, to the Irish housing market than to the UK housing market, finding evidence for a cross-country ripple effect (Stevenson, 2004).
Moreover, Nanda and Yeh (2014), studied the ripple effect in Tapai. Using spatial panel models, residential land prices were studied between 1992 and 2010 (Nanda & Yeh, 2014). They found that central regions seem to lead surrounding regions. However, high growth regions seem to display leading behaviour in the Tapai housing market as well (Nanda & Yeh, 2014). They found that area’s exhibiting high growth even advance city centres in price movements.
Lean and Smyth (2013) also found evidence for the presence of a ripple effect in Malaysia. However, they did find some evidence for segmentation in the Malaysian housing markets as they tested for different housing types. This study supports the earlier findings of segmentation in the housing market by Drake (1995) and MacDonald and Taylor (1993).
Contrary, to the previous studies, Shi, Young and Hargreaves (2009), found slightly different results for the New-Zealand housing market. There seems to be ripple effects within regions, but they found no evidence that the ripple effects spreads across different regional centres (Shi et al., 2009).
2.6
Reasons for a ripple effect
Economic theory opposes the existence of a ripple effect, as mentioned earlier (Canarella et al., 2011). However, many studies did find evidence of a ripple effect in UK and other housing markets.
Therefore, some potential reasons for the existence of a ripple effect are mentioned in the literature. Potential reasons could, according to Meen (1999) include: equity transfer, spatial arbitrage, spatial patterns in determinants of house prices and internal migration.
First, according to Meen (1999), people in the South-East of the UK (London region) have higher buying power when prices start to increase. This wealth effect could influence other regions house prices as people migrate from London towards these regions. Higher buying power forces prices to go up, as they compete for properties. Equity transfer is slightly correlated with internal migration, which will be described subsequently.
Furthermore, spatial arbitrage, could be a possible reason for the national diffusion of house prices (Meen, 1999). If the housing market would be efficient, arbitrage between regions would erase the relative price change in the long-run, according to Meen (1999). An example, given by him, is the presence of new information that affects house prices in a housing sub-market. If this information would spread, first to neighbouring markets, this could generate the same pattern as expected from a ripple effect. This creates, according to him, a “positive feedback effect” as market movements within regions could spill over to adjacent housing markets.
Thirdly, “spatial patterns of house price determinants” refers to the situation where regressors of house prices pursue identical patterns (Meen, 1999). This could take place when more economically developed regions benefit from economic growth earlier, compared to other regions. According to him this could, combined with inelastic housing supply and tightness within these developed regions, lead to a house price movement that appears to have a particular pattern. When the economy is growing and rents are declining this could lead to more significant price increases in these markets (Meen, 2001).
At last, internal migration could explain the ripple effect as well (Meen, 1999). This explanation is largely addressed in the ripple effect literature. Jones and Leishman (2006), found a relationship between internal migration and house price diffusion in Scotland. This study was based on local rather than regional housing markets, as migration linkages would be expected to be stronger between local housing markets (Jones & Leishman, 2006). They, found that regions that exhibited large migration linkages with Glasgow, revealed a higher lead-lag association. Besides, Alexander and Barrow (1994) refer to internal migration as a potential reason for the ripple effect their study found. They discovered that, in the southern part of the UK, the pattern of house price development matches the non-job-related migration. This is in alignment with the findings of Giussani and Hadjimatheou (1991) as well. They found a high correlation between interregional mobility and increases in house prices. Meen (1999) however, emphasizes that internal migration
within the UK is limited and that not all studies regarding the influence of internal migration on house price movements have reached the same conclusions.
Moreover, internal migration could explain the co-movement between house prices in the long-run, between different regions (Meen, 1999). This is in line with the current Dutch markets reports in section 2.1. When house prices in certain cities or regions become too high, households tend to move to different regions where the houses are still affordable (Savills, 2017; Rabobank, 2018). This can be seen in the current Dutch housing market, where households move to different regions to find affordable housing (Rabobank, 2018).
3.
Research hypotheses & Data description
The following section discusses both the hypotheses and the data that this study employs. First, the different research questions and the corresponding hypotheses will be discussed. The hypotheses will also be explained with the use of findings of earlier studies. Afterwards, the different datasets this study employs, to answer the hypotheses, will be introduced.
3.1
Hypotheses
The main focus of this study is to find evidence for a ripple effect in the Dutch housing market. However, as mentioned in the introduction, several side questions will be investigated as well. The following section will introduce the research questions and hypotheses of this study.
Main Questions & Hypotheses
Is there a ripple effect in the Netherlands?
H0: There is no ripple effect present in the Dutch housing market. H1: Ripple effects are present in the Dutch housing market.
Most of the studies mentioned in section 2.5 found evidence for the ripple effect. Therefore, based on these earlier researches and characteristics of the Dutch housing market, as described in the introduction, ripple effects are expected to exist.
Is there long-run convergence in house prices?
H0: There is no long-run convergence in the Dutch housing market. H1: There is long-run convergence in house prices in the Netherlands.
As most of the previous studies found evidence for long-run convergence we expect to find this in the Dutch market as well. However, based on the findings of Cook (2012) it is expected that prices converge during market downswings and diverge in upswings.
Does Amsterdam lead the ripple effect in the Dutch housing market? H0: Amsterdam is not leading the ripple effect in the Dutch housing market. H1: Amsterdam is leading the countrywide ripple effect.
Based on the extensive literature in the UK with London as the leading market, described in Section 2.5, Amsterdam is expected to lead the house price ripple effect in the Netherlands.
What is the reason of the ripple effect?
Do we observe a pattern in the results on provincial capitals and the region Amsterdam and Eindhoven? As the reason is still unclear, a good hypothesis is difficult to compose. However, we do expect that the economic performance of municipalities plays a role, as London leads the UK market and Meen (1999) mentioned it as possible reason. Therefore, it is expected that the ripple effect exists due to the economic development of regions. Regions that are the most economically developed, are expected to benefit from economic growth earlier.
3.2
Data
The following sections describes the data that is used for the different tests performed in this study. The first section will focus on the main dataset employed, that is provided by the NVM. The second section will focus on smaller and openly available datasets that were used to construct different control variables in the second beta-convergence test. Besides, both sections will elaborate on the main adjustments that were necessary in order to use the data.
3.2.1 NVM data
This study uses data on house price developments delivered by the NVM. This dataset contains yearly average square meter prices for all the municipalities in the Netherlands. The square meter price is preferred above average transaction prices as it controls for the potentially different sizes of the houses traded in various years and municipalities. Furthermore, the yearly data on square meter prices is available between 1990 and 2017. Therefore, this dataset enables us to measure the
28 years. Figure 1 shows the average square meter prices and the standard deviation for every year in the sample. This time period contains different up- and downswings of the Dutch housing market. This also enables us to test whether the effect is different between these periods.
First, this dataset empowers us to create price indices and yearly price developments for the separate Dutch municipalities. Therefore, it can be observed how house prices developed in these municipalities over the sample period. The timings of the house price shocks and the house price developments themselves can then be compared by techniques described in the methodology section. This data facilitates the opportunity to test for a ripple effect and long-run convergence of house prices in the Dutch residential market. As, the NVM data contains different house types, we can test the initial results for robustness.
However, the data used in this study does not enable us to create hedonic price models. Therefore, although we can control for housing type and size of the transacted houses, we are not able to control for various other housing characteristics. These characteristics could for instance be: lot-size, number of bathrooms and construction year. Hence, the square meter prices retrieved could be influenced by possible quality disparities of the houses sold within the studied time frame. Therefore, in this study, we are not able to control for the quality of the transacted houses.
Before the obtained NVM data can be used, it has to be adjusted in multiple ways. First, the yearly increases in square meter prices is generated. In order to correctly interpreted the VAR results, the variables have to be stationary. As the yearly house price development is found to be non-stationary the first difference of house price increase has to be generated. Besides, not all municipalities include the square meter prices for every year between 1990 and 2017. Furthermore, one municipality is unidentified. Municipalities that do not include the entire sample period or are not identified are dropped.
Figure 1: Descriptive statistic including: Yearly average square meter price of municipalities and the corresponding standard deviation.
3.2.2 Other data sources
Besides the NVM data on house prices, this study uses multiple other data sources for the second beta-convergence regression. This data is retrieved from open databases, like the Dutch Central Bureau of Statistics (CBS). However, some other smaller sources are used as well.
The open database of the CBS provides us with the yearly development of the Dutch gross national product (GDP) over the time period 1995-2016. This database also includes the provincial economic growth rates. Besides, the CBS data enables us to generate population growth rates over the years 1995-2016.
Other, independent variables like interest rates and consumer price index (CPI) are retrieved from other sources. Global-Rates, is used to find the yearly Libor interest rates (12 months) used in the regression. Furthermore, Inflation.eu is used to retrieve the Dutch CPI over the entire sample period.
The data gathered for this second regression, including control variables is not perfect, with some data and years missing. Therefore, the sample time period is slightly smaller than the NVM data sample period (1995-2016). Besides, municipalities with data missing are dropped, leaving a total of 288 municipalities in this regression.
4.
Methodology
The following section contains the methodology employed in this study. The upcoming paragraphs will describe the different models used to test the different hypotheses. This study will start with a beta-convergence test on proportions. This model will test for overall convergence and divergence of house prices in the Netherlands with all municipalities included (after dropping municipalities with missing values). Besides, another beta-convergence model, using a pooled OLS regression model, will be performed. Furthermore, the ripple effects will be tested by employing a VAR model, on the three different sub-samples. First, this study will check the relationship between Amsterdam and provincial capitals. Secondly, Amsterdam and surrounding municipalities will be tested. The third sub-sample contains Eindhoven and surrounding municipalities.
4.1
Beta-Convergence (Proportions)
This studies initial test on beta-convergence is based on the study of Steven Cook (2012). This model will be used to test whether house prices, in the Netherlands, converge in the long-run. As mentioned in section 2.4, long-run convergence does not imply convergence in the sense that housing prices differences narrow over time. Long-run convergence, as used in this study, is the tendency of house prices to return to earlier equilibrium ratio’s and therefore uses relative house prices. This model will also be used on different sub-sample of the time period (1990-2017) and test whether convergence and divergence may be dependent on the stage of the market cycle. For this study the annual house price data of all municipalities, after removing municipalities with missing values, are included.
Steven Cook (2012) provided two reasons why testing for these types of convergence has to be considered. First, a majority of other studies examined stochastic convergence earlier. Therefore, beta-convergence could offer other insights and enrichment of the evidence on the ripple effect (Cook, 2012). Besides, he accentuated that earlier studies on stochastic convergence possibly have to be re-evaluated, as he believed that they could be problematic.
Furthermore, beta-convergence is present in the housing market when the different series of house prices draw-together over time (Cook, 2012). This implies that housing markets with lower initial square meter prices will experience higher growth than sub-markets with high initial prices. The proportion of the low initial price group experiencing above average growth would be higher than the overall proportion of municipalities experiencing this high growth.
Comparable to the study of Cook (2012), this method will test both the entire sample period and different periods within the 28 years of data. This enables us to test whether the convergence and divergence occur in specific states of the market cycle. For instance, in the UK, there is evidence that London experiences a greater decrease in house prices in economic downturns (Cook, 2012). Therefore, it can be expected, that the housing market convergence would be more feasible in bearish markets, according to Cook (2012).
In his study, Cook (2012), uses the beta-convergence model proposed by Drennan and Lobo (1999). He, explains the model as two events A and B that might be inter-dependent. We denote the possibility that B occurs (prob[B]), as . The difference between and 𝜋𝑐 can then be compared and tested to observe whether A influences the occurrence of B.
𝜋𝑐 = 𝑝𝑟𝑜𝑏 (𝐵|𝐴) = 𝑝𝑟𝑜𝑏 (𝐵 ∩ 𝐴)
𝑝𝑟𝑜𝑏 (𝐴) (1)
To test the difference, according to Cook (2012), the required standard error has to be found. Afterwards we test for a significant difference with the standard normal distribution (Cook, 2012). 𝜎 = √𝜋(1 − 𝜋) 𝑛 (2) 𝑍 =𝜋𝑐− 𝜋 𝜎 (3)
As Cook (2012) proposes, the different municipalities have to be classified based on initial value and average growth rate. A1 implies a municipality with original square meter price below the average price of all municipalities in 1990. In contrast, A2 corresponds to a municipality with original value above this average value. Similarly, B1 implies a growth rate below the average of the entire sample, while B2 denotes an above average growth rate between 1990 and 2017. This leads to four different combinations where the Dutch municipalities, included in the sample, can be appointed to. These four combinations are described in Table 1. All these four classes can be tested by the Z-statistic described in equation 3 (Cook, 2012). This leads into the null hypothesis for the different classifications, in equations 4 and 5.
𝐻0: 𝑝𝑟𝑜𝑏 (𝐵𝑖|𝐴𝑗) = 𝑝𝑟𝑜𝑏 (𝐵𝑖) (4)
Or another way
𝐻0: 𝜋𝑐 = 𝜋 (5)
As Cook (2012) emphasized, the difference between 𝜋𝑐 and is important for the convergence of house prices. He, demonstrated the importance by considering {Bi Aj}. If convergence would occur, the probability of low growth given the above average original value will be larger compared to the unconditional probability of low growth (Cook, 2012). This would mean that high initial value leads to below average growth and therefore convergence of prices.
This test will be used to observe whether there is long-run convergence in the Dutch housing market, using equation 3. Therefore, it will be used to answer the second hypothesis, where we expect to find long-run constancy in price ratio’s and the short-term ability for diverge. The outcome of the convergence model will be tested for robustness by using the different time periods, as described earlier. As short-term divergence and convergence are expected to depend on the market cycle, using multiple time periods seems rational (Cook, 2012; Abbott & De Vita, 2013). This enables us to test whether the results vary when different time periods are used.
4.2
Beta-Convergence (Pooled OLS)
Besides the previous study based on Cook (2012), this study will also test for convergence in another way. The following model controls for multiple other variables instead of merely looking at growth and initial value. The housing market can be affected by many factors like for instance: GDP, interest rates, and the CPI (ING, 2018). Furthermore, local housing markets can be expected to be influenced by population growth and the growth of the local economy.
The following model will therefore test whether price change occurs due to the initial square meter price, countrywide effects or local effects. This model will test for long-run convergence in house price as well, using equation 6. In contrast with the previous model by Cook (2012), this could help explain why convergence or divergence take place, as this model also tests for the influence of the control variables on house price developments. Convergence would take place when housing markets with low prices would exhibit higher growth. Therefore, a significant negative coefficient for initial house prices would reveal convergence based on initial value. This model will look the following way:
∆𝜌𝑖𝑡 = 𝛽𝜌𝑖0+ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑉𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠𝑖𝑡+ 𝜀𝑖𝑡 (6)
Where the dependent variable is the house price change (in square meter) for municipality i in year t. Independent variable 𝜌𝑖0 is the initial square meter price of the municipality (1995). The control
variables will include yearly changes of GDP, interest rate change, CPI and the municipalities population. Besides, the economic growth of the province the municipality is located in, will be included. Furthermore, 𝜀𝑖𝑡 is the error term.
4.3
Vector autoregressive model (VAR)
This study uses the VAR model to test the ripple effect and to find the leading municipalities, using equation 9. As Amsterdam is believed to lead the ripple effect, the first regressions will be performed to test whether Amsterdam house price shocks help forecast future price shocks in provincial capitals. Besides, the interrelations between all other combinations of municipalities are tested as well. The two regional sub-samples will study the spatial patterns between house price shocks within different regions.
To test whether the house price development is a stationary process different ADF tests will be performed. In order to correctly interpreted the outcomes of the VAR model, stationarity of the time series is requested. To test for the stationarity of the house price shocks the Augmented Dickey Fuller (ADF) test will employ different lags. The following two variations of the ADF test will be performed (with different numbers of lags):
Intercept and no Trend
Δ𝑥
𝑡= 𝛼
0+ 𝛾𝑥
𝑡−1+ 𝜀
𝑡 (7)Intercept and Trend
Where Δ𝑥 is the percental house price change in year t. Furthermore 𝑥𝑡−1 is the percental house
price change in year t-1. Besides, 𝛼0 is the intercept, 𝛼2𝑡 is the trend and 𝜀𝑡 is the error term in
both regressions in year t.
If the percental development in house prices proves to be non-stationary, the tests will be performed on the first difference of the house price development of the municipalities. The ADF tests are performed with the use of AIC lags, two lags and no lags. These outcomes will be used to determine if the variables are indeed stationary.
After the tests for stationarity, the VAR model presented in equation 9 is used to test whether a municipality’s house price development could be forecasted by own lags and lags of other municipalities. As this research uses yearly indexed data a maximum lag of two years is introduced. The maximum lag of two years seems logical if we look at outcomes of other studies. The following VAR model is created:
Δ𝑥𝑡= 𝛼0+ ∑ 𝛼𝑖 𝑛 𝑖=1 Δ𝑥𝑡−𝑖 + ∑ 𝛽𝑖 𝑛 𝑖=1 Δ𝑦𝑡−𝑖+ 𝜀𝑡 (9)
where 𝛼0 is the constant, x and y are yearly percental house price developments, n is the number
of lags included and 𝜀𝑡 is the error term.
The outcomes of the VAR model will be tested with the Granger causality test. This test will be used to check whether the lagged values, found in the VAR model, are significantly different from zero. Therefore, this will observe whether price shocks in a particular municipality can help to forecast future house price movement in others. This study will test these relationships for all combinations of Amsterdam and the twelve provincial capitals. This enables us to study if Amsterdam is indeed leading the effect and if the effect is expected to be nationwide or segmented. Besides, the two regional sub-samples will test whether there are ripple effects within regions and which municipalities are leading these effects. The null hypothesis will be: lagged values of y do not Granger cause future values of x. The alternative hypothesis will be: lagged values of y do Granger cause future values of x.
The methods to test for the stationarity and ripple effects are based on the paper of Shi et al. (2009) that tested for a ripple effect in New-Zealand. Another paper that uses a similar methodology is the paper of Jones and Leishman (2006) that tested for a regional ripple effect around Glasgow.
4.4
Robustness test (VAR)
The data this study uses will enable us to test for robustness in multiple ways. As described in the data section earlier, the NVM data includes square meter prices for different housing types. Therefore, it can be tested if the relationships between municipalities differ between these various types of housing. This thesis will use square meter prices of rowhouses to test for robustness of the initial outcomes. This robustness test will be performed on all three of the sub-samples.
5.
Results
The following section will include the results of the regressions included in the methodology section. The results will be discussed and explained. Besides, possible reasons for the results will be discussed. Furthermore, the outcomes of the robustness tests will be compared and discussed as well.
5.1
Beta-Convergence (Proportions)
The first convergence model looks at the relationship between initial square meter value and the corresponding growth throughout the years. Table 2, shows for each group, based on the composition explained in the methodology, when convergence and divergence take place.
This method, based on Cook (2012), is performed on different periods and therefore already includes the robustness test. The results obtained with the use of equation 3 can be found in Table 3. This table presents the results for the entire period (1990-2017) and sub-periods. The time period 2003-2007 indicates a peak in the cycle, a time period where house prices developed positively. The third time period 2008-2013 indicates the financial crisis where average house prices were declining.
Table 2: When do convergence and divergence take place
Over the entire sample period, the results do suggest constancy or long-run convergence of relative house prices. No evidence was found for divergence of relative house prices when using the entire sample period. In the time period 2003-2007 there seems to be no relationship between initial square meter value and growth rate for all groups, as well. Although the coefficients point in the direction of convergence, they are not significant and almost equal to zero.
However, during the financial crisis, house prices, between municipalities, did seem to converge. Municipalities with initial square meter values below average showed smaller decreases in house prices compared to municipalities with above average square meter prices in 2008. These outcomes, presented in Table 3, are also significant at the 1% level. Nonetheless, during the period after the financial crisis, characterized by price increases, house prices appear to have diverged again. Municipalities with above average square meter prices in 2014 seem to exhibit higher growth, compared to municipalities with lower house prices. These differences between municipalities in the different groups are significant at the 1% level.
Table 3: Outcome convergence and divergence test. ***significant at 1%
5.2
Beta-Convergence (Pooled OLS)
This beta convergence test is based on equation 6 and includes multiple control variables. Table 4, includes the outcome for the time period 1995-2016. The initial square meter price, as can be observed in Table 4, does not have a significant impact on percental house price changes in municipalities.
Besides, Table 4 includes the outcomes for the control variables. Countrywide effects seem to have great influence on the Dutch house prices. The percentage change in gross domestic product seems to significantly affect house price change. If the GDP increases with one percent this would lead to an increase in house prices with 1.77 percent. Besides, house price change is affected by the interest rate in the markets. When the LIBOR twelve-month interest rate decreases by one percent, house prices are expected to increase by 1.19 percent. Furthermore, the consumer price index (CPI) significantly affects house prices. An increase in CPI of one percent increases
house prices with 1.94 percent. Besides, all three countrywide variables are significant at the 1% significance level.
Local and regional factors, do seem to have some influence as well. The economic growth of the province, where the municipality is located in, seems to significantly influence house price growth. Although this effect is smaller than the effect of the countrywide factors, it is significant at the 1% level. However, the growth of inhabitants in the municipality does not significantly affect house price change.
These results, suggest that divergence and convergence of relative house prices do not occur due to the
initial square meter values of the municipalities. Although, the coefficient is negative it does not prove to be significant. Countrywide factors like GDP, CPI and interest rates do seem to influence house price change heavily.
5.3
Results provinces (VAR)
The current paragraph will present the results for the ripple effect between provincial capitals and Amsterdam. As, percental house price
change was not found to be stationary the first difference was taken. Table 5, shows that the first difference of percental house price change is indeed stationary, for all included municipalities, in most of the different ADF tests performed. Therefore, non-stationarity can be rejected, and the variables are assumed to be stationary.
Table 4: Outcome pooled OLS regression for conversion. Regression 1 uses normal standard errors. Regression 2 uses Robust standard errors. ***significant at 1%
Table 5: ADF outcome for the provincial capitals. The different columns display different AFD tests with variance in lags and the inclusion of a trend. ***significant at 1% **significant at 5% *significant at 10%
Table 6, presents the Granger causality relationships between the different cities, included in the sample. The vertical axis represents the municipality which lagged value forecast future price change in the cities on the horizontal axis. Because the ripple effect encompasses the spread of house price shocks, only positive forecasting relationships are included in Table 6. Besides, Figure 2 gives the locations of the provincial capitals and the lead lag relationships originating in Amsterdam. Furthermore, Figure 3 displays the relationships found in Table 6. The order of house price movements is made by observing the interrelations between cities. Regions that are not affected by previous shocks in other regions and forecast multiple other municipalities are believed to lead the ripple effect.
Table 6, immediately displays that Amsterdam is unaffected by other municipalities within the sub-sample. Besides, Amsterdam can be used to forecast six out of the twelve provincial capitals. Utrecht is also unaffected by other house price shocks but can only be used to forecast Maastricht. Although unexpected, Assen is unaffected by shocks emanating from other municipalities and seems to lead shocks to The Hague and Groningen. Due to the use of yearly data the order of these three cities cannot be made. However, due to the amount of forecasting relations and the non-significant coefficients between the municipalities, Amsterdam is believed to lead the effect.
Furthermore, house price shocks seem to affect Groningen, Haarlem and ‘s-Hertogenbosch rapidly. Haarlem and ‘s-‘s-Hertogenbosch both are only affected by Amsterdam. Groningen is only affected by Amsterdam and Assen. Besides, Haarlem forecasts house price change in multiple other municipalities like The Hague and Zwolle. Due to the lead-lag relations, that can be observed in Figure 3, these municipalities are believed to move after Amsterdam, Utrecht and Assen.
The Hague and Maastricht can only be forecasted by municipalities that are placed in the first two categories. This also applies to Zwolle, although The Hague has some forecasting power on Zwolle as well. Besides, these three municipalities are able to forecast some of the remaining municipalities in the sub-sample. Therefore, these municipalities are believed to exhibit price shocks after the first two categories of cities. Furthermore, Arnhem is only affected by Haarlem and does not seem to be able to forecast other cities within the sample. Nonetheless, Arnhem is believed to move before the last category as it is only influenced by Haarlem.
Leeuwarden, Lelystad and Middelburg are found to be located in the last category. These municipalities are believed to react last to house price shocks. Leeuwarden forecasts Middelburg and Lelystad and is therefore expected to react before those cities. Both Lelystad and Middelburg
are affected by other municipalities, although, they are not able to forecast. Therefore, these two are believed to react to price changes last.
These outcomes provide some evidence for a ripple effect starting in Amsterdam, leading to later house price developments in other regional capitals. The effect seems to ripple out to other regions ending in Lelystad and Middelburg.
However, some of the relationships between municipalities, are perceived to be strange and unexpected. For instance: the leading position of Assen and the small amount of lead-lag relationships originating in Utrecht are both found to be unexpected. Nonetheless, it should be noted that the amount of observations is limited and intermunicipal relationships can change over time. For instance, strong relationships in the earlier years of the sample can change later. However, the overall results match with what would have been expected.
Figure 3: Var lead-lag relationships between municipalities within the sub-sample. These relationships are based on the findings that can be found in Table 6 and represent all significant lead-lag relationships between the different municipalities. Municipalities are located in the most efficient way using Gephi software.
Ta ble 6 : O utco m e o f th e G ra ng er ca usa lity te st on th e o utco m es of th e V AR m od el, in cl ud in g tw o la gs. Th e su b-sa m ple co nta in s A m ste rd am a nd a ll D utch p ro vi nci al ca pit als. Sig nifi ca nt re la tio nsh ip s fo un d st art in th e ve rti ca l m un ici pa liti es an d a ffe ct m un ici pa liti es on th e h ori zo nta l a xis . ***si gn ifi ca nt at 1% **si gn ifi ca nt at 5% *si gn ifi ca nt at 10 % le ve l.
5.4
Results Amsterdam region (VAR)
The current paragraph will present the results for the ripple effect including Amsterdam and surrounding municipalities. As, percental house price change was not found to be stationary the first difference is taken. Table 7, shows that the first difference of percental house price change is indeed stationary, for all included municipalities, in most of the different ADF tests performed. Therefore, non-stationarity can be
rejected, and the variables are assumed to be stationary.
Table 8, presents the relationships between the Amsterdam and surrounding municipalities. Just as the previous VAR model on provincial capitals, the vertical axis represents municipalities that forecast horizontal ones if the coefficient is found to be significant. As in the previous case, municipalities will be ordered based on the intermunicipal relationships found. Again, only positive relationships are presented in Table 8. Figure 4 shows how the price shocks develop within the Amsterdam
region. This shows that there are indeed lead lag relationships, but they are not even through time and space, as would be expected. Furthermore, Figure 5 shows the intermunicipal relationships based on the outcome presented in Table 8.
In Table 8, it can be observed that Amsterdam and the surrounding municipalities Amstelveen and Diemen, are not affected by earlier price shocks in other municipalities. Besides, the lagged values of these municipalities forecast the largest amount of future house price shocks within other municipalities. The same is observed for De Ronde Venen and Purmerend. Therefore, it looks like the greater Amsterdam leads the cycle with Purmerend and De Ronde Venen following relatively quickly or even immediately. Due to the use of yearly data it is not possible to find more detailed evidence for the order of these municipalities.
Furthermore, Velsen, Haarlem and Heemstede are only affected by the Amsterdam, Amstelveen and De Ronde Venen (1% and 5% significance levels). Besides, both Haarlem and Heemstede are able to forecast five other municipalities. Ouder-Amstel is also included in this
Table 7: ADF outcome for the region Amsterdam. The different columns display different AFD tests with variance in lags and the presence of a trend. ***significant at 1% **significant at 5% *significant at 10%.
