Determinants of regional house price dynamics: the role of
supply constraints and sales activity
by
Sander Burgers (10166505)
Master Thesis MSc. Business Economics
Dual track: Finance and Real Estate Finance University of Amsterdam (UvA) Thesis supervisor: dhr. dr. M.I. Dröes
1st of July, 2017
2 Statement of Originality
This document is written by student Sander Burgers 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 paper analyzes the role of supply constraints and sales activity on regional house price dynamics. A novelty of this study is that it analyzes the relative importance of these factors using a large panel dataset which contains data of 388 Dutch municipalities over 29 years. An error correction model is conducted to estimate both long- and short-term relationships between house prices and determinants. To observe the impact of these factors, the models are estimated on several subsamples based on different degrees of supply constraints and sales activity. The results show that supply constraints and sales activity significantly affect regional house price dynamics. The heterogeneity in regional house price dynamics can therefore partly be explained. Moreover, this study finds that these factors amplify each other. As a consequence, the housing market in regions which are highly supply constrained and have high sales activity are more prone to bubbles.
4
Content
1. Introduction ... 6
2. Literature review ... 8
2.1 House price determinants ... 8
2.2 Heterogeneity in house price dynamics ... 10
2.3 Determinants of house price dynamics ... 12
2.3.1 Supply constraints ... 12
2.3.2 Sales activity ... 13
2.4 Hypothesis ... 14
3. Data ... 16
3.1 House prices ... 16
3.2 House price determinants ... 17
3.3 Determinants of house price dynamics ... 21
3.4 Data overview... 24
3.5 Validity checks ... 25
3.5.1 Unit root tests ... 25
3.5.2 Multicollinearity ... 28
4. Methodology ... 29
4.1 Model specification ... 29
4.2 Model estimation ... 30
5. Results ... 31
5.1 Long-term equilibrium model ... 32
5.1.1 Determination of the long-term model ... 32
5.1.2 Preferred long-term model ... 33
5.1.3 Long-term model on regions with different supply constraints ... 36
5.2 Short-term dynamic model ... 38
5.2.1 Determination of the short-term model ... 38
5.2.2 Preferred short-term model ... 39
5.2.3 The role of supply constraints on short-term dynamics ... 40
5.2.4 The role of sales activity on short-term dynamics ... 42
5.2.5 The role of supply constraints and sales activity on short-term dynamics... 44
5.3 Summary of results ... 47
6. Conclusion and discussion ... 49
5
References ... 53
Appendix A: Data Sources ... 58
Appendix B: Figures ... 59
6
1. Introduction
House prices in the Netherlands dropped sharply as a consequence on the Financial Crisis. On a national level, the average decline from 2008 to 2013 is around 20% (CBS, 2014). However, large regional differences in this decline can be observed. In Amsterdam, the capital and largest city, the decline in real house prices was around 22% while in Bunnik, a rural town in the center of the Netherlands, the decline was around 39%. This raised the question what drives regional house prices and their heterogeneous movement. Several studies towards heterogeneity in Dutch house price dynamics have been conducted. Research of Brounen and Huij (2004) showed that heterogeneity exists in terms of sensitivity to economic shocks. They find this heterogeneity along the dimension of geographical region, the degree of urbanization and the type of house. De Vries and Boelhouwer (2005) also find regional heterogeneity in Dutch house prices. They find that the three major Dutch cities are more sensitive to increases in the housing stock. Moreover, studies by Galati et al. (2011, 2013) show that heterogeneity exists along the dimension of geographical region and degree of urbanization. In 2011 they find that the most urbanized areas have negative serial correlation and the slowest mean reversion speed compared to other regions. Two years later, they find that urban areas show no persistence in house prices and the quickest mean reversion. Although they find heterogeneity, their results are ambiguous and therefore regional house price dynamics in the Netherlands need to be clarified.
Research towards foreign housing markets also demonstrates regional heterogeneity. Abram and Hendershott (1996) find that the coastal areas have larger persistence in house prices and quicker mean reversion. These results are confirmed by Capozza et al. (2002, 2004). In their study, they report an exploratory list of determinants of house price dynamics. These include supply constraints, information costs and market expectations. Their findings suggest that supply constraints and information costs are the factors that explain the heterogeneous reactions of house prices to economic shocks. They argue that regions with more supply constraints should exhibit more house price persistence and slower mean reversion. In contrast, house prices in regions with high sales activity have lower information costs and should exhibit less house price persistence and quicker mean reversion. These theories, however, do not provide a clear hypothesis when it comes to urban areas, which often have high supply constraints and high sales activity. Therefore, this thesis focusses on these two
7 factors and attempts to answer the following question: to what extent do supply constraints and sales activity explain the heterogeneity in regional house price dynamics?
This study makes use of actual transactional house price data of the Dutch Association of Realtors and Property Consultants (NVM) of 388 Dutch municipalities over a period of approximately thirty years. Data of house price determinants are obtained from Statistics Netherlands (CBS) and De Nederlandsche Bank (DNB). An error correction model on panel data is estimated to analyze the long- and short-term effects on house price determinants and short-term dynamics of house prices of various regions. The regions are then classified into different groups based on the degree of supply constraints and the degree of sales activity. This classification allows for comparing the influence of these factors on the regional housing markets.
The main novelty of this study is that it will analyze the relative importance of supply constraints and sales activity in determining the house price dynamics. To the best of my knowledge, this has not been done before in existing literature. Another novel aspect is that this study is based on a large number of regions in the Netherlands, which will provide at least two benefits. Firstly, using a large number of cross-sectional observations increases the statistical power tests. Besides, using a large number of regions allows for classifying the sample into several subsamples while maintaining a large number of regions within each subsample.
Understanding regional house price dynamics is essential for policy makers because resources may be misallocated in case prices diverge from equilibrium in certain regions. This might influence residential mobility, especially in the Netherlands, where supply responsiveness of house is low (Caldera and Johansson, 2013). Besides, it may affect the local employment and wage dynamics (Saks, 2008). Mian and Sufi (2010) have also emphasized the critical importance of an efficient housing market for a well-functioning economy.
This study will continue as follows. Firstly, the relevant literature on house price determinants and heterogeneity in house price dynamics will be discussed. In the end of that section, the hypotheses will be formulated. The third section describes the variables used in this paper and the explains the methods used to create the panel data set. In the fourth section, the methodology of the model specification is explicated. Subsequently, the results will be presented and analyzed to test the hypotheses. The conclusions are drawn in the final section.
8
2. Literature review
In order to answer the question, a wide range of house price literature is consulted. This section provides an overview of this literature by setting out the most relevant literature with respect to determinants of house prices, house price models and the determinants of house price dynamics. This section is organized as follows. The first subsection examines the existing literature on the fundamental determinants of house prices. The second subsection gives an overview of the house price models used in practice. The third subsection explains the main theories and findings of the causes of the heterogeneity of house price dynamics. The fourth subsection focusses on supply constraints and sales activity as determinants explaining the regional heterogeneity of house price dynamics.
2.1 House price determinants
There exists some consensus about the fundamental factors that cause house prices in the long run. The vast majority of studies agree that the level of income and the (mortgage) interest rate are significant determinants of house prices.1 In most of these cases, both variables are denoted in real terms, i.e. corrected for inflation. Some other variables that possibly determine the house prices in the long run include: population (Francke, Vujic and Vos (2009), employment opportunities (Visser, Van Dam en Hooimeijer, 2008), loan-to-value (LTV) ratio (Timmermans, 2012), household wealth (Verbruggen et al. 2005; Francke et al. 2009), the level of construction costs (Van der Windt, 2014), housing stock (Francke, Vujic and Vos, 2009) and stock prices (Kakes and Van Den End, 2002). At least some evidence has been found for these variables to affect house prices in the Netherlands. However, there is not really a consensus about the best possible combination of these variables to explain house price variation.
Francke et al. (2009) show that real disposable household income, real long-run interest rate, the housing stock, the level of construction costs, the number of households, and the real wealth of household explain the fluctuations in house prices quite well. Kranendonk and Verbruggen (2008) explain house price movements from 1980 until 2007 by the real aggregate disposable labor income, the real long-run interest rate, a real wealth indicator and total
1(e.g. Galati et al. 2011, 2013; Hort, 1998; Van der Windt, 2014; Capozza et al., 2002; Malpezzi,
9 housing stock. Verbruggen et al. (2005) have estimated a model to explain the variation in real house prices from 1980-2003. They find that the total real disposable loan income, household wealth net of stocks, the real interest rate and the housing stock are able to explain 96% of its variation.
Dröes and Van de Minne (2017) include the impact of Gross Domestic Product (GDP) per capita, the opportunity costs of capital, unemployment, the share of the working-age population, construction costs, population growth and the housing stock to explain house price fluctuations in the long-run. Where the opportunity costs of capital is a combination of interest rates and user costs. Timmermans (2012) include the real GDP, the real mortgage rate, the real construction cost, a proxy for housing stock and the LTV to explain house price fluctuations.
Visser et al. (2008) have analyzed regional characteristics and their effect on house prices. They use hedonic price models and data of the residential environment to conclude that the accessibility to employment opportunities is an essential factor when homebuyers consider their residential location. According to De Vries and Boelhouwer (2005), the need for housing in the Netherlands depends primarily on local and regional factors.
In the macroeconomic model of DNB (2011) house prices are primarily driven by the amount of mortgage credit. Mortgage credit, in turn, is determined by disposable income and the mortgage interest rate net of the tax rate in the highest income bracket. In the short-run, dynamics of houses prices are affected by the mortgage rate and the unemployment. Despite that this is a rather unconventional way of explaining house price dynamics, it works remarkably well.
In foreign countries, comparable determinants explain house price variation. Abraham and Hendershott (1996), for example, explain changes in house prices of metropolitan areas in the United States (US) by changes in real income, real construction costs and the real after-tax interest rate. Capozza et al. (2002) analyze the heterogeneity of house prices across six metropolitan cities in the US. The determinants which have a significant effect on the real house prices are real median income, population size, real construction costs, the 5-year population change, user cost and land supply index. The user costs is a combination of the mortgage rate, the property tax rate, the income tax rate and the inflation rate.
Hort (1998) has found that the variation in real house prices in Sweden is caused by variation in income, user costs of housing capital and construction costs. Whereby user costs
10 of housing capital consist of mortgage costs net of tax, property tax and depreciation, maintenance and repair expenditures. She also tests whether the net lending ratio, as a measure of the availability of credit, has an effect, but it appears to be insignificant. Grimes et al. (2004) have estimated long-run relationship between real house prices and determinants in New-Zealand on a regional level. They find that the regional economic activity, housing stock, the user costs of capital and changes in regulations on the financial systems significantly affect house prices.
2.2 Heterogeneity in house price dynamics
Heterogeneity in the dynamics of house prices can be explained in two distinctive ways. The first case is to explain heterogeneity through the difference of impact of the house price determinants across segments (De Vries and Boelhouwer, 2005; Brounen and Huij, 2004). This way, the variables affecting the house price have a different effect across segments. The other case is through the different features of house prices cycles. Here, the heterogeneity could be observed through differences in persistence and mean-reversion. Studies which focus on this kind of heterogeneity are conducted by Capozza et al. (2002,2004), Galati et al. (2013) and Gao et al. (2009), to name a few.
De Vries and Boelhouwer (2005) focus on the first case of heterogeneity and find that the supply of houses influences house prices in a different magnitude across regions. They compared the four largest cities (Amsterdam, Rotterdam, The Hague and Utrecht) in the Netherlands with other regions and found that the house prices declined in response to an increase in the supply in the four largest cities, while in the other regions this relationship was not significant.
Moreover, Brounen and Huij (2004) assess the sensitivity of house prices to economic fundamentals across regions in the Netherlands. They find that the differences in regional price developments are not only caused by differences in the regional economies, but also by differences in their sensitivity to changes in economic fundamentals. The degree of urbanization seems to play a major role in explaining this heterogeneity. They suggest that the population density of the region and the flexibility of the regional labor market could be the underlying determining factors.
Galati et al. (2011) focus on the second case of heterogeneity in house price dynamics. They find heterogeneity in house prices dynamics along the dimension of geographical region,
11 the degree of urbanization, funding conditions and income expectations. They find in general a negative coefficient for the degree of serial correlation, which suggests that households do not value their house in a persistently adaptive way. Regarding the heterogeneity, they show that the areas with the highest degree of urbanization, i.e the biggest cities, have the lowest (most negative) serial correlation and the lowest speed of convergence to the equilibrium level. In a subsequent paper, Galati et al. (2013) also find that house prices have a negative serial correlation coefficient and that they are mean-reverting. However, house prices in highly urbanized areas have a higher speed of convergence and show more market efficiency. This finding with respect to the speed of convergence is in contrast with their previous study in 2011.
Van der Windt (2014) has analyzed the presence of heterogeneity in house price dynamics across price segments in the Netherland. He finds that interest rates have a higher impact on house prices in the higher price segments and that house prices in the lowest price segments adjust quicker to the long-run equilibrium price.
Empirical evidence of heterogeneity in housing markets is found in foreign countries too. Himmelberg et al. (2005), for example, have analyzed house prices in the United States from 1980 and 2004 and find heterogeneity in terms of sensitivity to economic shocks. They find, that interest rates have a larger impact on house prices in regions where the housing supply is relatively inelastic. Abraham and Hendershott (1996) have also found heterogeneity in house prices dynamics in the US. They find that house prices in coastal metropolitan areas have larger autoregressive coefficients compared to inland metropolitan areas. Coefficients on mean reversion appear to be insignificant in inland cities but are about 10% a year in coastal areas.
Capozza et al. (2002, 2004) state that variation in the fluctuations of real house prices across areas is not only caused by variation in local economies but also by different responses of house prices to economic shocks. They find that house prices in regions with high real income growth have more persistence and a higher adjustment speed to the equilibrium price. Besides, they suggest that high construction costs in an area raise serial correlation and lower the adjustment speed. Gao et al. (2009) uses autoregressive mean reversion (ARMR) model to analyze the price dynamics of cyclical versus non-cyclical markets. They use the standard deviation between actual and fundamental house prices over time to classify housing markets in cyclical and non-cyclical markets. Cyclical (or volatile) markets show different dynamics than
12 non-cyclical (or tame) markets in the US. They found that cyclical markets have larger autoregressive coefficients than non-cyclical markets. In addition, upward periods more persistence than downward periods. This shows that that house prices are more likely to overshoot the fundamental price in booming periods while in busts markets, they will show downward rigidity.
2.3 Determinants of house price dynamics
In the previous section, it has been shown that several studies showed the presence of heterogeneity in house price dynamics along different dimensions. Regional heterogeneity has been found in the Netherlands by Galati et al. (2011; 2013) and Brounen and Huij (2004). Brounen and Huij (2004) suggest that population density and/or flexibility of the local labor market can possibly explain this heterogeneity. Capozza et al. (2002) argue that the heterogeneity in house price dynamics can be explained by sales activity, by supply factors and by heterogeneous expectations. This study will focus on the role of the first two factors and will, therefore, be explained in more detail below.
2.3.1 Supply constraints
According to DiPasquale and Wheaton (1994), new supply of houses will put downward pressure on house prices. The speed of which this new supply is available is however crucial. Adding houses to the stock generally occurs slowly, because building a housing is a very time-consuming process. The sooner newly constructed houses are supplied to the market, the sooner downward pressure is exerted on the prices. This so-called responsiveness of housing supply is a crucial factor in the functioning of the housing market (Caldera and Johansson, 2013). Several other studies have supported this.
As stated before, Capozza et al. (2002, 2004) find that areas with high real construction costs (due to stricter regulation, for example) show price dynamics which are characterized by higher serial correlation and lower mean reversion. They argue that new construction is a way to exploit the inefficient prices. When this is hampered due to high construction costs and/or supply constraints, they are no forces to lower the prices. As a consequence, house prices would show more persistence. The results of their studies provide evidence for their hypothesis.
Grimes and Aitken (2006) found that regions in New-Zealand with high supply responsiveness have relatively small price rises following demand shocks. Besides, the time
13 that prices are above the long-run equilibrium level is shorter when the supply in the regions is more responsive. Evidence of a significant impact of responsiveness of housing supply on the dynamics of house prices has also been found by Glaeser et al. (2008). They found that places with more elastic housing supply have shorter and fewer bubbles, along with smaller prices increases. However, the consequences may be more severe in case of an eventual burst because those places will overbuild more as a result of the sharp price increase.
Furthermore, Gyourko (2009) find that price adjustments after a demand shock translate themselves more in the price of a house rather than in expanding the housing stock in supply-constrained areas. Himmelberg et al. (2005) confirm this, stating that house prices are more sensitive to interest rate changes in regions where housing supply is relatively inelastic. Additional evidence is recently found by Hilber and Vermeulen (2016) in the English housing market. They confirm their prediction that supply constraints have a significant impact on the house price-earnings elasticity.
Andrews (2010) have analyzed real house prices in OECD countries and finds that the variability of house prices is higher in countries where the supply of housing is price inelastic. In rigid supply areas, demand shocks are more capitalized into prices than to be translated into a higher quantity of houses. Similar results have been found by Andrews et al. (2011).
2.3.2 Sales activity
Information costs are high in real estate markets. Transactions occur sporadically and the asset is strongly heterogeneous. This heterogeneity of housing leads to search and transactions costs, which makes it difficult for homeowners or potential buyers to assess the current value of a property (Capozza et al, 2002). Potential buyers or sellers, therefore, must look to transactions distant in time or location of similar properties (Quan and Quigly, 1991). Ideally, the appraiser uses a weighted average of transactions. Higher sales activity can therefore positively affect the information dissemination because information is easier obtainable when houses sell more often. Clapp et al. (1995) confirm this, by stating that information costs will be lower the greater number of transactions per unit area. Van Dijk and Francke (2017) also find a link between sales activity in the housing market and information dissemination. They state that price discovery is faster in regions where the rate of sale is high. However, there are geographical differences in these dynamics. Liquidity temporarily
14 increases in urban areas as a result of more market tightness, but not in rural areas. Transactions remained low in the rural regions and therefore price discovery is slower.
The hypothesis of Capozza et al. (2002, 2004) continues by reasoning that homeowners can incorporate the latest market information more easily when there is more price discovery and therefore set a price which better reflects the equilibrium price. As a result, the price should adjust more quickly to the equilibrium level and respond with more magnitude to an economic shock according to Capozza et al. (2002, 2004). They find evidence which confirms this hypothesis. However, it should be noted that population is used in their study as a proxy for the amount of transactions of houses in a specific period, which is by far perfect measure. Grimes et al. (2004) have analyzed whether sales activity influences the efficiency of the housing market through improved information dissemination. They constructed a panel dataset containing housing prices and house price determinants covering all regions of New Zealand over 88 quarters, ranging from 1981 – 2002. They use the ratio of house transactions to the housing stock in each region to proxy sales activity. They hypothesize that higher sales activity improves information dissemination and leads to a higher speed of adjustment of house prices to the equilibrium level. Their findings, however, do not confirm their hypothesis. Instead, their results reveal that a higher rate of sales leads to a slower adjustment to the equilibrium price. Their finding is therefore in contrast with the reasoning of Capozza et al. (2002, 2004). They interpret this finding by associating high sales activity with some fad-element. They reason that in a booming market when the price is already above equilibrium, the high number of transactions may delay the adjustment back to the equilibrium prices. In a depressed market, when prices are below equilibrium, the higher number of transactions does drive prices upward towards equilibrium.
2.4 Hypothesis
Based on the theory reviewed above, several hypotheses are derived in an attempt to answer the research question.
Hypothesis 1: Changes in income have a larger impact on house prices over the long-run in regions that are more supply constrained. According to the existing literature, supply constraints negatively affect the supply responsiveness of houses. When the housing stock cannot expand rapidly enough in response to a demand shock, this shock will be capitalized
15 into higher house prices. Therefore, an income shock is expected to have a larger impact on house prices in regions where the supply constraints are relatively high.
Hypothesis 2: House prices in regions that have high supply constraints have higher persistence and adjust slower to the long-run equilibrium house price. This hypothesis is derived from the rationale that the effect of an exogenous shock on house price will have a more prolonged effect if the housing stock cannot expand rapidly enough. If the construction of houses cannot be added quickly enough in response to demand shocks, house price would not experience downward pressure and therefore the potential overshooting will not be quickly reversed. Therefore, the persistence is expected to be higher in regions which are more supply constraint and moreover the speed of reversion would be lower.
Hypothesis 3: House prices in regions that have high sales activity have less persistence and adjust faster to the long-run equilibrium house price. According to the information-based theory of Capozza et al., (2002, 2004) it is expected that a higher degree of sales activity leads to more price discovery. As a result, it is easier to assess the current value of a property in these regions. In case the house price is not in line with equilibrium level, the price should revert faster to the equilibrium price in these regions. This is consistent with more market efficiency. For this reason, it is therefore expected that house prices would exhibit a significantly lower degree of serial correlation. In an efficient market, house price should be determined more by current market information instead of information out of the previous period(s).
Hypothesis 4: House prices in regions that have high supply constraints and have low sales activity have more persistence and adjust slower to the long-run equilibrium house price. This hypothesis is derived from a combination of hypothesis 2 and 3. The same holds for the next hypothesis.
Hypothesis 5: House prices in regions that have low supply constraints and have high sales activity have less persistence and adjust faster to the long-run equilibrium house. This hypothesis is derived from a combination of hypothesis 2 and 5. These hypotheses will be tested and discussed in the section with results.
16
3. Data
In order to answer the main question and test the hypotheses, data on house prices and several house price determinants are collected to compile a panel dataset covering all 388 municipalities over a period from 1988 until 2016. This section describes the variables and the required techniques to construct them. Moreover, their relevance and expected effect on house prices are briefly discussed.
3.1 House prices
The median house price per cubic meter is obtained from the Dutch Association of Realtors and Property Consultants (in Dutch: Nederlandse Vereniging van Makelaars; NVM). The data is obtained for 388 different municipalities, which is according to the municipality classification of the CBS per 1-1-2017. The dataset covers a period from 1985 until 2016. The median price per cubic meter is calculated based on the transacted properties within the municipality in a specific year. Prices per cubic meter have the advantage over prices per dwelling to obtain an index that is at least, to some extent, of constant quality. The median house price per cubic meter is used in order to reflect the house price development in each municipality. Using the median price per cubic meter is preferred over the average price per cubic meter because it is less sensitive to outliers. Using the median price of the transactions assume that the underlying mix of properties and characteristics is on average constant over time. This assumption is, however, questionable because a sample selection bias might occur. Despite this, I am still conformable that the median price per cubic meter functions well as a proxy for the actual price developments. For the reason that some municipalities have a very low amount of transactions, some municipalities are not included in the analysis. A more detailed description of the decision criteria to include a municipality is found in Appendix A. Eventually, house prices of 316 municipalities covering 1988-2016 (29 years) are remaining. The house price variable will be corrected for inflation is therefore denoted in real terms. Data of the Consumer Price Index (CPI) of the CBS are collected. The CPI series are only available on a national level and ranges from 1963 until 2016. Based on the CPI, inflation rates are computed.
17 3.2 House price determinants
Data of the average disposable household income per municipality per year is obtained from Statistics Netherlands (in Dutch: Centraal Bureau voor de Statistiek; CBS). To create the variables, several databases covering different time periods have been used.
For the period of 1946 until 1994, a dataset of regional income distribution is used. This dataset does not contain the absolute level of disposable income but instead contains indices per municipality relative to the average national disposable income of the Netherlands. The municipality classification is per 1-1-1997. The dataset is not continuous and therefore does not contain the required indices of the time period used in this study but solely contains the relative indices of 1984, 1989 and 1994. These indices are used to calculate the absolute average disposable household income based on the national average disposable income. The average national disposable household income of the years 1989 and 1994 is directly obtained from CBS (2014). The average disposable income of 1984 is computed by means of linear interpolation of the income of 1981 and 1985. Subsequently, the average disposable income per municipality is calculated for the years 1984, 1989 and 1994 based on the nationwide absolute number and the municipality specific index numbers. The incomes of the missing years (1985, 1986, 1987, 1988, 1990, 1991, 1992, 1993, 1995, 1996 and 1997) are calculated by linear interpolation based on the income of 1985, 1989, 1994 and 1998.
For the period of 1998 until 2005, separate annual databases are used containing the average disposable household income per municipality. For the period of 2006 until 2014 the average disposable income per household is obtained from one single database. The classification of municipalities is per 1-1-2015, which consists of 393 unique municipalities in the Netherlands. For the years 2015 and 2016, data of the average disposable household income is not available. Therefore, yearly growth rates of the GDP per municipality are used to estimate the average disposable household income based on the level of each municipality in 2014.
Since many municipalities in the Netherlands have merged over the last three decades, the classification of municipalities changed yearly. At 1-1-1985, 741 municipalities existed in the Netherlands, while at 1-1-2017 388 municipalities exist covering the same area. To match the income to the house prices of the 388 municipalities per 1-1-2017, an overview of the mergers on municipalities is consulted from CBS. In order to obtain a complete series for the regions which have been merged over the sample period, a weighted average of the
18 household income is taken to reflect the household income of the regions according to the municipality classification at 1-1-2017. This weights of the weighted average are based on the population of the respective municipalities.
covering the same region over time, the household income of a year equals a weighted average of the incomes of the former municipalities of which the concerning municipality consisted of.2 These weights, are based on the population of the respective municipalities.
After merging these databases, 28 municipalities do not contain a full series of data. Most of these, miss around two data points, in that case, the missing observation is replaced by linear interpolation. Four municipalities miss more than three data points in a row and are therefore excluded from the analysis.
The disposable household income is expected to have a substantial positive effect on house price since is it a major factor in obtaining mortgage credit, which is in turn, a primary driver of house prices (DNB, 2011). According to Capozza et al. (2004) income affects house price because the demand for housing services is income elastic. They argue that residents consume more housing space in response to an increase of income and therefore results in higher median prices per housing unit. This positive relationship between income and house prices is confirmed in many other studies (e.g. Grimes et al., 2004; Adams and Füss, 2010; Hort, 1998; Verbruggen et al.2005).
Another variable which is expected to significantly influence the development of house prices is the mortgage interest rate. This variable is solely available on a national level and is obtained from DNB. The variable is composed of two series ranging from 1980 until 2016. The first series ranges from 1980 until 2006 and contains the average of the 5-year mortgage interest rate issued to a mortgage. From 2003 until 2016, the mortgage rate is calculated by taking the average of the interest rate paid on newly issued mortgages with a maturity over 1 year and up to 5 years and mortgages with a maturity over 5 years up to 10 years.
2 To clarify, the municipalities of ‘s-Hertogenbosch is taken as an example to illustrate this complexity. In 1993,
Geffen and Nuland merged to become Maasdonk. In 1995, Rosmalen became part of ‘s-Hertogenbosch. In 2015, Maasdonk was dissolved, as a result Geffen became part of Oss and Nuland became part of ‘s-Hertogenbosch. The income series of ‘s-Hertogenbosch from 1985 until 2016 therefore consists of several parts. From 1985 until 1993 a weighted average of Nuland, Rosmalen and ‘s-Hertogenbosch is. From 1993-1995 a weighted average of Maasdonk, Rosmalen and ‘s-Hertogenbosch is computed. From 1995 until 2015, a weighted average of Maasdonk and ‘s-Hertogenbosch is computed. From 2015 onwards, ‘s-Hertogenbosch did not merge and thus the areas remained the same.
19 Subsequently, these series are combined to create a national mortgage rate series. The mortgage rate is expected to have a negative impact on house prices because it is a major determinant of the maximum amount of mortgage credit a household can obtain (DNB, 2011). Besides that, the cost of borrowing capital to purchase a house is a major component in the user cost of a dwelling. Therefore it is expected that a lower mortgage rate makes buying a housing relatively cheaper and therefore more attractive. Thus, prices will tend to increase. Several studies have confirmed this negative effect on house prices in the Netherlands (e.g. Francke et al., 2009; OESO, 2004; Verbruggen et al., 2005; Van der Windt 2014).
Data of unemployment is collected from CBS and is available on municipality level from 2003 until 2016.3 Before 2003, unemployment rate data are not available on municipality level but solely on a province level. Therefore, the unemployment rates on province level are collected and their yearly change is used to complete the unemployment rate series on a municipality level. Unemployment is expected to have a negative effect on regional house prices. Studies of Visser et al. (2008) and De Bruyne and Van Hove (2013) have shown the importance of the accessibility to employment opportunities in explaining regional house price variation in respectively the Netherlands and Belgium. This has also been confirmed by Miller (1982) and Kauko (2003) who state that the accessibility of jobs is a dominant force in the resident’s location choice and therefore expected to have a significant effect on house prices. Apart from the fact that unemployment affects house prices, there is no consensus whether it affects house prices in the long-run or in the short-run, or both. Studies which include unemployment in the long run relationship are those by De Wit et al. (2013) and Dröes and Van de Minne, (2017). Others, like Zandi and Chen (2006), Verbruggen et al. (2005) and Adams and Füss (2010), include changes in unemployment in the short-run model only.4
Data of the total level of the housing stock is obtained from two datasets of the CBS and is available per municipality on a yearly basis. The first dataset contains the housing stock from 1947 until 2011 and the second dataset from 2012 until 2016. The level of the housing stock is originally per 31st of December, but in this thesis the average housing stock of the previous year and the current year is used. The housing stock levels in the datasets do not match with
3 The unemployment rate is defined as the number of working people over the labor population. Working people
is defined as people (aged 15 until 65) who have a paid job for more than 12 hours a week. Labor population comprises people who work at least 12 hours a week or are already agreed to work at least 12 hours a week or declared that they are willing and able to work and actively looking for work at least 12 hours a week.
20 the classification per 1-1-2017. Therefore, the mergers of municipalities must be taken into account. This is done in the same way as with the income variables. Although, the simple sum of the housing stock is taken instead of a weighted average. Missing observations of municipalities are linear interpolated and/or extrapolated. The most intuitive theory of the impact of housing stock on house prices states that housing stock functions as the supply of houses and therefore is expected to have a negative effect. This negative effect has been confirmed in the Dutch housing market in several studies (Verbruggen et al., 2005; De Vries and Boelhouwer, 2005; Timmermans, 2012). However, the literature on the international housing market suggests that the effect may be small (Abraham and Hendershott, 1996; Hort, 1998; Malpezzi, 1999). The estimated effect of the housing stock on regional house prices will be tested. The expected effect of housing stock is negative.
Population data are obtained from CBS and is available on municipality level on an annual basis. The dataset covers the period from 1960 until 2015. The data originally represents the number of inhabitants at the 31st of December of the specific year. In this study, the average of the current year and previous year is taken to reflect the average population that lived in the regions at the concerning year. This dataset is also corrected for the mergers of municipalities over the years to match the municipality classification of 1-1-2017. It is expected that the population positively affects house price of that region, based on the theories and empirical findings of several studies. Capozza and Helsley (1989) argue that cities with a higher number of inhabitants have higher rents and prices to preserve the intraurban locational equilibrium. Empirical evidence is found by Capozza et al. (2004) who include population in the long-term house price equation and find a positive relationship between metropolitan area size and house prices in the US. Moreover, Dröes and Francke (2016) find that population growth was one of the main drivers of house prices.
Another possible determinant of house prices is the construction cost index. This index reflects the development of the total costs for newly constructed dwellings. The index is provided from 1914 until 2016. To create the index in real terms, it is a corrected for inflation. Several studies (e.g. Dröes and Van de Minne, 2017; Capozza, et al.2004) have included a real construction cost index in the model and find positive effects. Therefore, the real construction cost index is also considered to include into the model.
Data of the LTV-series is composed based on information from the Dutch Household Survey (DHS), which is a yearly survey of household finance. In the DHS of 2014, the
21 respondents are asked when the house is bought and what the value of their mortgage was relative to the amount they paid for their house at that time. Based on the information of 1,057 respondents, the LTV-series for first-time homebuyers from 1982 to 2012 is derived. A more detailed description can be found in Appendix A1. This series has previously been used in Timmermans (2012) and Verbruggen et al. (2015).
The LTV ratio reflects the willingness of banks to provide a mortgage to first-time home-buyers. The higher this ratio, the more capital homebuyers can spend on their house and therefore it is expected that house prices will increase. In the 1990s the average LTV for first-time buyers increased sharply, see Appendix B, Figure A1. This can partly be explained by the fact that in 1995 the second income of the household started to be taken into account when determining the mortgage amount. Since 2010, a downward trend commenced, reflecting the policy of the government of lowering the maximum LTV to 100% in 2018, which is expected to have put downward pressure on prices. Empirical evidence of this positive relationship has also been found by Verbruggen et al. (2015) and Timmermans (2012) who studied the impact of lowering the LTV cap to 90%. Verbruggen et al. (2015) conclude that restricting the maximum LTV to 90% would lead to a long-run decline in house prices of around 10%.
3.3 Determinants of house price dynamics
To create a proxy for supply constraints, a share developed variable is constructed based on the method of Hilber and Vermeulen (2016) is used. The share developed is the fraction of land that is already developed over land where development is possible. In order to get this measure, data is obtained from Van Dijk et al. (2016). They use geographical data based on satellite pictures made in 2004 to determine each type of land use. These type of land uses are then into three classes: non-developable land, developable yet undeveloped land and developed land. In the paper of Hilber and Vermeulen (2016), the share developed variable is operationalized in the following way. Land uses which are defined as ‘suburban/rural developed’ and ‘urban development’ are classified as developed area. ‘Grass heath’, ‘mown/grazed turf’, ‘meadow/verge/semi-natural swards’, ‘bracken’, ‘dense shrub heath’, ‘scrub/orchard’, ‘deciduous woodland’, ‘coniferous/evergreen woodland’, ‘tilled land’, ‘inland bare ground’ and ‘open shrub heath’ are classified as developable area. ‘Sea/estuary’, ‘inland water’, ‘coastal bare ground’, ‘saltmarsh’, ‘ruderal weed’ and ‘felled forest’ are classified as
22 non-developable area. Then, to compute the share developed for each municipality, the fraction of developed area over the total developable area is calculated5.
Figure 1. Share developed per municipality
Notes: share developed is the fraction of developed area over the developable area. The data is based on the
Land Cover Map of The Netherlands in 2004 and is obtained from Van Dijk et al. (2016).
5 In formula: 𝑆ℎ𝑎𝑟𝑒 𝑑𝑒𝑣𝑒𝑙𝑜𝑝𝑒𝑑 = 𝐷𝑒𝑣𝑒𝑙𝑜𝑝𝑒𝑑 𝑎𝑟𝑒𝑎 𝑖𝑛 𝑠𝑞𝑢𝑎𝑟𝑒 𝑚𝑒𝑡𝑒𝑟𝑠
23 The share developed for each municipality is illustrated in Figure 1. This variable proxies the supply constraints in each region and will be used to analyze its impact on regional house price dynamics. Since geographical constraints are found to be strongly and positively correlated to regulatory barriers (Saiz, 2010), this proxy also captures some of the regulatory constraints.
24 As a proxy for the sales activity, the average rate of sale is used. The rate of sale is defined as the total amount of transactions in a year divided by the number of houses available for sale in that year. The number of houses available for sale in a given year consists of the number of houses that were available for sale in the beginning of the year plus the number of houses that has been put on the market during the year. These data are provided by the NVM covering a period from 1985 until 2016. The rate of sale is municipality specific and varies over time. The average rate of sales is the average per municipality over the sample period. This variable indicates the turnover of houses in a region and is found to positively affects price discovery (Van Dijk and Francke, 2017). The average rate of sale for each municipality is illustrated in Figure 2.
3.4 Data overview
Table 1 contains an overview of the variables used in this paper. Real house prices, real disposable household income, the real construction cost index, population and housing stock are expressed in natural logarithms. Table 2 shows the descriptive statistics of the variables.
Table 1. Variables
Variable Description
h Log real median house price per cubic meter
y Log real average household disposable income
i Mortgage interest rate
𝑖𝑟 Real mortgage interest rate
u Unemployment rate
hs Housing stock
p Log population
c Log real construction cost index
ltv Loan-to-Value (LTV) ratio
sd Share developed
ros Average rate of sale
Notes: Variables in real terms are computed by multiplying the variable by (𝑝𝑟𝑖𝑐𝑒 𝑖𝑛𝑑𝑒𝑥100 ), where the price
index is equal to 100 in the year 2000. The log real median house price per cubic meter, log real average
household disposable income, unemployment rate, housing stock and log population are on municipality level and time-varying. The variables share developed and average rate of sale are on municipality level but are not time-varying. The mortgage interest rate, the real mortgage interest rate, the LTV ratio and the log real constucion cost index are time-varying but on national level.
25 3.5 Validity checks
This subsection discusses the two main conditions a variable should meet in order to be validly used in the error correction model. Firstly, the variables have to be integrated of order one and secondly, the explanatory variables should not suffer from multicollinearity.
3.5.1 Unit root tests
Concerning the first condition, the house prices and the house price determinants (the explanatory variables) have to be integrated of order one, shortly I(1)6. This is a requirement to find a cointegrating relationship between house prices and its determinants. A cointegration relationship exists if a combination of these series is stationary, shortly I(0). If this is the case, the combination of series show the same pattern over time and will fluctuate around an equilibrium level, which is a crucial assumption when using an error correction framework.
Table 2. Descriptive statistics
Variable Notation Obs. Mean Std. Dev. Min Max
h 9164 6.20 0.36 4.87 7.28 y 9164 3.50 0.13 3.08 4.14 i % 29 5.79 1.89 2.45 9.81 𝑖𝑟 % 29 3.73 1.64 1.38 7.10 u % 9164 4.97 1.60 1.43 13.38 hs 9164 9.50 0.76 7.71 12.96 p 9164 10.41 0.72 8.82 13.63 c 29 4.57 0.10 4.44 4.74 ltv ratio 29 1.00 0.04 0.92 1.04 sd ratio 316 0.24 0.16 0.04 0.77 ros ratio 316 0.58 0.06 0.41 0.74
Note: The amount of observations (Obs.) of the mortgage interest rate, the real mortgage interest rate,
the construction cost index, share developed and the average rate of sale indicate the unique amount of observations.
6 A series is called I(d) or integrated of order d, when the d’th-difference is stationary and when it contains d unit
26 In the context of this thesis, a cointegrating relationship between the dependent variable, the house prices, and the independent variables, the house price determinants, is crucial. Because it implies that house prices and its determinants show a constant relationship over time. If that is not the case the series might have a tendency to diverge from each other. Hence that, to meet the cointegration assumption, the time series variables should be I(1) and the linear-combination should be I(0). To test this, both panel unit root tests and individual unit root tests are performed. With regards to the variables which are in panel form, panel unit root test suggested by Van Dijk et. al (2011) and Dröes and Francke (2016) are performed.
According to Van Dijk et. al (2011), the most popular panel unit root tests in the literature are those of Levin et al. (2002) and Im et al. (2003), which are therefore considered. These tests, however, are only consistent when the series within the panel are uncorrelated. In the panel data used in this paper, it does not seem plausible that regional house prices are not correlated because of two reasons. Municipalities comprise a relatively small area and therefore the amenities are often easily accessible by inhabitants of the adjacent municipalities. Moreover, despite that regional housing markets have their own dynamics, the national trend still has a strong influence on local prices (Goodman, 1998; Berg, 2002). Therefore, it is very likely that house prices are correlated across entities.
To formally test whether these regions affect each other, a Cross-sectional Dependence (CD) test of Pesaran (2004) is conducted. The results are given in Appendix C, Table A3. As can be seen from the CD test statistic, the null hypothesis of no cross-sectional dependence should be rejected, indicating that the median house prices of municipalities are cross-sectionally dependent. Hence that, the LLC (2002) and IPS (2003) are inconsistent and therefore not applicable. Instead, a Cross-sectionally augmented IPS (CIPS) test (Pesaran, 2007) is conducted. This test takes cross-sectional dependence into account and can also be adequately used when the number of cross-sectional units (N) is larger than the time periods (T) (Van Dijk, et al. 2011). The CIPS is based on the IPS test but adds the cross-sectional averages of lagged levels to the augmented Dickey-Fuller (ADF) tests to correct for the cross-section correlation. This test is called the cross-cross-sectional augmented Dickey-Fuller (CADF) test. The CIPS statistic is computed by taking the simple average of the individual CADF tests. The results are shown in Appendix C, Table A4.
Next to the CIPS-tests, Fisher tests are conducted as is done by Dröes and Francke (2016). The Fisher test allows to perform augmented Dickey-Fuller tests or Phillips-Perron tests for
27 each time series individually and subsequently combines p-values from these tests to produce an overall test. The null hypothesis states that all-time series in the panel data contain a unit root. The alternative hypothesis states that at least one time series is stationarity.
Both tests are comparable and combine the several independent tests. Besides, both tests allow for heterogeneous alternative hypothesis in which some series are non-stationary and some are not. The crucial difference is however that the Fisher test combines the significance levels while the CIPS test combines the test statistics. Since the errors in this panel data set appear to be cross-sectionally dependent, it seems more appropriate to use the CIPS test. Nevertheless, Fisher tests are conducted too.
The results, including a more elaborate interpretation of these tests can be found in Appendix C, Table A5. Both the levels and first differences of the variables are tested against mean stationarity and trend stationarity. The results indicate that the disposable household income, the housing stock and log population are integrated of order one.
The results of house prices and unemployment rates of the various tests are somewhat contradictory. With respect to house prices, both panel unit root tests indicate that at least one of the series is stationary, but when separate unit root tests are performed on individual series, it seems that the series are integrated of order two, shortly I(2). Some other studies also find ambiguous results, but decide to treat (regional) house prices as non-stationary containing a single unit root, because of the lack of power of the panel unit root tests or economic theory (e.g; Taylor and MacDonald, 1993; Dröes and Francke, 2016; Grimes et al., 2004; Van der Windt, 2014). There are also studies that find that Dutch house prices are I(1) (Francke, et al., 2009; Hofman, 2005; Timmermans, 2012).Therefore, house prices are treated as I(1) is this paper too. It is expected that the results will still be interpretable.
Concerning the unemployment rates, panel unit root tests indicate that at least one of the series is stationary. When separate unit root tests are performed on individual series, the order of integration cannot be determined with certainty and is possibly I(1) or I(0). Theories have often stated the existence of a natural rate of employment, which would imply that the unemployment rate is stationarity around this level because shocks would only have a temporary effect and converge to this natural level. And since it is a ratio, it is bounded between zero and hundred. However, if the convergence is too slow, unemployment may be persistent and thus non-stationary. Since both theories sound valid, the ambiguity remains. Like Grimes et al. (2004), the series is treated as if it is I(1).
28 To test for unit roots in univariate series Augmented Dickey-fuller (ADF) tests are conducted. The results are shown in Appendix C, Table A6. It can be concluded that the mortgage interest rate, the real mortgage interest rate and the LTV ratio are integrated of order one. The real construction cost index is integrated of order two.
It can be concluded that most variables satisfy the conditions to be validly used in a panel error correction model. The real house price variable and the construction cost index violate the conditions of an error correction model. However, it is expected that the results will still be interpretable. Table 3 provides a summary of the results of both the panel unit roots tests and univariate unit root tests.
3.5.2 Multicollinearity
Concerning the second condition, the possibility of perfect multicollinearity is analyzed. To avoid multicollinearity in the regressions, correlations between the variables are examined. Table A1 in Appendix C shows the correlation coefficients for the variables in levels and Table A2 in Appendix C shows the correlation coefficients for the variables in first differences. The results conclude that there is no perfect multicollinearity. For this reason, problems because of perfect multicollinearity will not occur.
Table 3. Summary of results of unit root tests
CIPS Fisher ADF
Variable Intercept only With intercept and trend Intercept only With intercept and trend Intercept only With intercept and trend
h I(1)/I(0) I(1)/I(0) I(1)/I(0) I(1)/I(0) I(2)/I(1) I(2)/I(1)
y I(1) I(1) I(1) I(1) - -
u I(1)/I(0) I(1)/I(0) I(1)/I(0) I(1)/I(0) I(0) I(1)
hs I(1) I(1) I(1) I(1) - -
p I(1) I(1) I(1) I(1) - -
i n.a. n.a. n.a. n.a. I(1) I(0)
𝑖𝑟 n.a. n.a. n.a. n.a. I(1) I(1)
c n.a. n.a. n.a. n.a. I(2) I(2)
ltv n.a. n.a. n.a. n.a. I(1) I(1)
Notes: ADF tests on the individual series of the panel variables house price and
unemployment are performed to provide more insights concerning the order of integration. The ADF tests are run separately on every region individually.
29
4. Methodology
4.1 Model specification
In order to estimate long-run relationships and short-run dynamics, the error correction model (ECM) is used, conducted on panel data (Engle and Granger, 1987). This type of model is used by many other researchers with time-series data to explain house price fluctuations. 7 The ECM allows for estimating a long-run relationship between house price and fundamentals and a short-run dynamic model. The model is defined as:
ℎ𝑖𝑡∗ = 𝛽1𝑥1𝑖𝑡+ ⋯ + 𝛽𝑘𝑥𝑘𝑖𝑡+ 𝛽𝑘+1𝑍𝑡+ 𝛽𝑘+2𝐷𝑖 (1) ∆ℎ𝑖𝑡 = 𝛼∆ℎ𝑖𝑡−1+ 𝜕(ℎ𝑖𝑡−1 − ℎ𝑖𝑡−1∗ ) + 𝛾1∆𝑥1𝑖𝑡+ ⋯ + 𝛾𝑘∆𝑥𝑘𝑖𝑡+ 𝜀𝑖𝑡 (2) Where ℎ𝑡𝑖 is the real median house price per cubic meter of municipality i at time t and ℎ𝑖𝑡∗ is the unobserved long-run equilibrium house price. The 𝑥𝑘𝑖𝑡 variables denote the house price determinants that explain the house prices and are time-varying and varying across regions. These variables include the average real disposable household income, the population, the unemployment rate and the level of housing stock. In model (1), the 𝑍𝑡 denote the fundamentals that are time-varying, but not do vary across regions. The (real) mortgage rate, the real construction cost index and the LTV-ratio are examples of these. The 𝐷𝑖 denote municipal fixed effects. These variables will be specified as dummies. These municipal fixed effects are fixed over time but vary across municipalites. These are included in the model to capture unobserved municipal heterogeneity. The first equation (1) is labeled as the long-term equilibrium model. In other papers (e.g. Galati et al., 2011; Gao et al., 2009) this is referred to as the fundamental house price model.
In model (2), the short-term dynamic model, several components can be recognized. The first component consists of the variable ∆ℎ𝑖𝑡−1, which denotes the lagged value of the change in the real house price and the coefficient 𝛼. The coefficient 𝛼 estimates the degree of serial correlation. This coefficient reflects the persistence of house prices. It is also stated that it indicates speculative behavior (Abraham and Hendershott, 1996) or capture market ineffiency (Francke et al., 2009). The coefficient also shows the features of the cycles, specifically its persistence and amplitude (Capozza et al., 2004).
7 (e.g. Hort, 1998; Capozza et al., 2002; Grimes et al. 2004; Malpezzi, 1999; Verbruggen et al. 2005; Francke, et
30 The second component is the error correction term. The error correction term is defined as the lagged difference in actual house price and the equilibrium house price, i.e. 𝑒𝑐𝑡𝑖𝑡−1 = (ℎ𝑖𝑡−1 − ℎ𝑖𝑡−1∗ ). In case this difference is positive, i.e. (ℎ𝑖𝑡−1 − ℎ𝑖𝑡−1∗ ) > 0 , house prices were overvalued in municipality i in period t-1. Conversely, if the error correction term is negative, i.e. (ℎ𝑖𝑡−1 − ℎ𝑖𝑡−1∗ ) < 0, house prices were undervalued. The error correction term is included in the model in its lagged form, because it is expected that mispricing in the previous year affects the price movement in the current year. The coefficient 𝜕, indicates the speed of adjustment towards the long-run equilibrium house price and is also often referred to as the mean reversion coefficient. The coefficient is expected to be negative, since overvaluation should exert downward pressure on prices in the next period. Vice versa, undervaluation should have upward pressure on price in the subsequent period. If the coefficient increases, the frequency and amplitude of the cycle tends to increase (Capozza et al., 2004). The inverse of the coefficient indicates the amount of years needed to reach the equilibrium price again.
The third component consists of the first differences of the house price determinants ∆𝑥𝑘𝑖𝑡 and the corresponding coefficients. These coefficients 𝛾1, . , 𝛾𝑘 estimate the short-run effects of the change in house price determinants on changes of house prices. It is assumed that these variables are weakly exogenous.
4.2 Model estimation
One of the main assumptions of this method is that house prices fluctuate around a long-term equilibrium house price, which can be explained by a set of economic variables. As stated in the literature review, many studies have shown that this is the case (e.g. Adams and Füss, 2010; Capozza et al., 2002; Verbruggen et al., 2005; Malpezzi, 1999; Taylor and MacDonald, 1993). This implies that the error correction term must be stationary. This is a crucial condition to validly use an error correction model. In case the error correction term is non-stationary and contains a unit root, there does not exist a long-term cointegrating relationship between the house price and the house price determinants. In that case, house prices and determinants might tend to diverge from each other. Unit root test will test for this stationarity and is discussed in the results section.
31 In case this condition is satisfied, the two-step procedure of Engle and Granger (1987) can be followed. The first equation (1) is estimated by Dynamic OLS (DOLS) with hetereroscedascitiy- and autocorrelation consistent (HAC, Newey and West (1987)) standards errors. The estimated coefficients of the DOLS estimator shows the long-run effects, which captures the accumulation of effects over time. Moreover, it takes the stickiness of house prices into account (Adams and Füss, 2010). The DOLS estimation in this paper employs one lead first difference, one current first difference and a lagged first difference8. The HAC standard errors correct for the possibility that the errors are uncorrelated across clusters but allow the regression errors to have an arbitrary correlation within an entity (Stock and Watson, 1993). DOLS is also conducted by other studies towards long-run house price relationships using panel data. These include Andrews (2010), Adams and Füss (2010) and Dröes and Francke (2016). DOLS performs better than most alternative estimation methods (Wagner and Hlouskova, 2009). Equation (2) is estimated by OLS, assuming heteroscedasticity-consistent (HC, White (1980)) standard errors.
To assess the impact of supply constraints and sales activity on house price dynamics, the regions will be classified into equally sized subsamples based on the proxies shared developed and the average rate of sale. In order to assess the effect of these factors simultaneously, four equally sized groups will be created based on both the level of supply constraints and sales activity. Subsequently, in order to test whether two coefficients of different groups are statistically different, a t-test of Clogg et al. (1995) is conducted:
𝑡 = θ1−θ2
√(𝑆𝐸θ1)2+(𝑆𝐸 θ2)2
(3)
Where θ𝑘 denotes the estimated coefficient of variable k and 𝑆𝐸θ𝑘denotes the standard error
of θ𝑘 .
5. Results
In this section, the results are presented. Several steps will be performed to test the hypotheses and answer the main question. Firstly, several combinations of house price
8 In this study, there is chosen for only one lead and lag because increasing the number results in a shorter sample
and thus less observations to estimate the long-run relationship. The estimation was also performed using two leads and lags, but this did not substantially change the estimated coefficient and therefore, these results are not reported.
32 determinants are included in the model to find the best fitted long-run equilibrium price. Then, the long-run model will be estimated for low, medium and high supply constrained areas. Subsequently, the run dynamic models are estimated. The most optimal short-run model will be applied to estimate the house price dynamics for the three groups, based on supply constraints and sales activity. In addition, the model will be estimated on four groups, based on both factors. Finally, the results of the models will be compared to test the hypotheses.
5.1 Long-term equilibrium model
Based on the literature discussed in section 2, several combinations of variables are explored to include into the long-term model. These variables comprise the real average disposable household income, the (real) mortgage rate, unemployment, population, the real construction cost index, the LTV-ratio and the housing stock. Although evidence of a significant effect has been found for any of these variable, there is no consensus about which combination of these variables fits best in explaining house price. Therefore, several combinations will be explored.
5.1.1 Determination of the long-term model
In Appendix C, Table A7 the different model estimations can be found. At first, a model (1) including the real average disposable household income, the nominal mortgage rate, population and municipal fixed effects is estimated. In model (2), the unemployment rate is added and gets the expected negative sign. This combination of variables results in a significant positive long-run effect of population. In model (3), the nominal mortgage rate is replaced by the real mortgage rate. The t-value increases (closer to zero) which indicates that the impact is less significant. When comparing the models, the (adjusted) R-squared and F-statistic are slightly higher and the RMSE slightly lower when using the nominal mortgage rate and indicating a better fit. For this reason, including the nominal mortgage rate is preferred. In model (4) the real construction cost index is added. The sign is positive indicating that higher real construction costs lead to higher house prices in the long-run. Subsequently, a model (5) excluding the unemployment rate and construction cost but including the LTV-ratio is estimated. The effect of the mortgage rate turns positive and significant, which is counterintuitive. Therefore, the unemployment rate is added back to the model, see model (6). The coefficient of the unemployment rate enters the model significantly and with the expected sign. However, the coefficient of the mortgage interest rate remains positive.
33 Another possible combination is model (7), which includes real average disposable household income, the mortgage rate, population, unemployment, the real construction cost index, the LTV-ratio, the housing stock and municipal fixed effects. In this combination, the mortgage rate turns positive. The coefficient of the housing stock is, however, insignificant and therefore it is removed from the model. This results in model (8).
5.1.2 Preferred long-term model
Considering the significance of the variables, the goodness of fit and the stationarity of residuals, model (8) provides the best long-term house price model. Model (8) includes the real average disposable household income (y), the mortgage interest rate (i), the unemployment rate (u), population (p), the real construction costs (c), the LTV ratio (ltv) and municipal fixed effects to explain variation in regional house prices over time. The cross-sectional average of house prices, the fitted equilibrium value and the residuals (error correction term) are plotted in Figure 3.
Figure 3. The actual house price, the long-run equilibrium house price and residuals
The coefficients should be interpreted as effects over the long run. The interpretation is as follows: for a given region, a 1% increase in the real disposable household income leads on average to a 0.71% increase in house prices; a 1%-point increase in the mortgage rate reduces house prices by 4.74%; a 1%-point increase in the unemployment rate reduces house prices by 3.22%; a 1% increase in population raises house prices by 0.11%; a 1% increase in real
-.15 -.1 -.05 0 .05 .1 .15 .2 .25 .3 .35 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 6.4 6.6 1985 1990 1995 2000 2005 2010 2015 2020 year
