THE INFLUENCE OF HOUSING WEALTH ON PRIVATE CONSUMPTION IN THE NETHERLANDS
January 11, 2018
Noisa van Malkenhorst, 1450956 Master Thesis
Leiden University, Governance and Global Affairs Master Public Administration: Economics & Governance
Supervisor: Max van Lent
Acknowledgements: I would like to thank Max van Lent for his help during this research. It was very beneficial to have a supervisor, who always took time for discussion and kept me motivated when things were not going my way. Particularly his guidance on the empirical analysis was of vital importance to the quality of this thesis.
Abstract
This research explores the relationship between housing wealth and private consumption in the Netherlands between 2004 and 2017. In the literature, there are three hypotheses proposed for the co-movement between housing wealth and private consumption: increased housing values lead to increased consumption, as individual’s permanent income increases (wealth effect); housing values and consumption are both influenced by the same macroeconomic prospects (common causality effect); growing housing values lead to less credit constraints, as it increases collateral for homeowners to borrow against (collateral effect). The DHS panel (micro) data on households is used, to identify the influence of the different factors. In this paper, the common causality effect is controlled for and a marginal propensity to consume (MPC) out of wealth of 0.02% for the collateral effect is found and a 0.18% MPC for the wealth effect is found when housing wealth changes by 1%.
Keywords: Housing wealth effect, Collateral effect, Common causality, Micro-data, Consumption, Netherlands, housing prices, DHS panel, WOZ
Table of Contents
Abstract ... 2
1 The influence of Housing Wealth on Private Consumption in the Netherlands ... 5
2 Theory ... 10
2.1 Life Cycle-Permanent Income Hypothesis ... 10
2.2 Literature Review on Macroeconomic Research ... 12
2.3 Literature Review on Microeconomic Research ... 14
2.4 Theoretical Framework ... 16
3 Methodology ... 19
3.1 Data description DNB Household Survey (DHS) ... 19
3.2 Consumption Construction ... 23
3.3 Empirical Model ... 24
4 Results ... 36
5 Conclusion, Discussion and Recommendations ... 42
References ... 48
Appendix A: Assets used for the Consumption variable calculation ... 54
Appendix B: Debts used for the Consumption variable calculation (N=7,666) ... 55
Appendix C: Table OLS regression people that stayed in the same house (N=7,666) ... 56
Appendix E: Table FE model decreasing and increasing WOZ values (N= 9,828) ... 58
Appendix F: Table OLS regression young, middle-aged and old (N=7,666) ... 59
Appendix G: Table OLS regression 50 years old division (N=7,666) ... 60
1 The influence of Housing Wealth on Private Consumption in the Netherlands This paper is inspired by the rapid growing housing values in the Netherlands of the last years, following on the sharp decreases after the financial crisis of 2008. Figure 1 shows that in the years prior to the crisis, consumption and average housing values rapidly increased in the Netherlands. This went on till the start of the financial crisis in 2008. Starting in 2009, there is a clear downward trend noticeable for both housing value and consumption. In 2014, the pattern turned around and shows an upward movement from here.
Figure 1: Average housing value based on WOZ values reference dates (CBS, 2018a) and consumption expenditures (CBS, 2018b) in the Netherlands
In the Western world, the housing value is often the largest household asset (Wang-Li, Hook & Chin, 2015). In the academic world, consensus exists about the existence of a positive relationship between housing value fluctuations and consumption. However, the exact cause for this link is still controversial. Knowing the cause for this relationship can be beneficial for
policymakers who want to influence consumption, as they can make better informed decisions. For instance, understanding this relationship can be useful to prevent a next crisis, as falling housing values lead to even more decreased economic activity which may result in a recession. Furthermore, knowing the exact link is also important for academics trying to understand the choices individuals make regarding their consumption expenditures. The hypothesized link is a product of the Life Cycle-Permanent Income Hypothesis (LC-PIH) academic stream. This stream contains two models stating that individuals want to smoothen their earnings. The Life Cycle Hypothesis assumes individuals plan this period based on an average lifetime and want to consume all of their wealth in their lifetime (Modigliani, 1954). Under this model, the assumption is that if housing values increase, homeowners possess more wealth which they smoothly want to consume over their life. Older people increase their consumption more than younger individuals, as they presumably have a shorter period left to consume it in. Secondly, the Permanent Income Hypothesis assumes the period individuals want to smoothen their income is infinite (Friedman, 1957). This led to the assumption that old and young people adjust their consumption the same way, if housing values fluctuate. Recent academic papers have identified three different channels through which housing values potentially affect consumption of
individuals. The first is the wealth effect, which is based on the logic that if housing values decrease, individuals feel poorer than before (their perception of their own wealth goes down) and will reduce spending (e.g. Campbell & Cocco, 2007; Case, Quigley & Shiller, 2005; Carroll, Otsuka & Slacalek, 2011; Muellbauer & Murphy, 1990). Others scholars only find an effect for decreasing housing values. Engelhardt (1996) explains this by emphasising that in in a situation of increasing values, it is hard to turn this wealth into cash and subsequently into consumption.
However, if values decrease, individuals can easily decide to consume less (Engelhardt, 1996). Secondly, in the literature the collateral or credit constraint effect is known, which entails that an increase in housing wealth may lead to the relaxation of credit constraints for homeowners, as their property increases in value they have more collateral which allows them to borrow more money (e.g. Aoki, Proudman & Vlieghe, 2004; Browning, Gørtz & Leth-Petersen, 2013;
Campbell & Cocco, 2007; Disney, Bridges & Gathergood, 2010; Iacoviello, 2004; Leth-Petersen, 2010). This group of homeowners is primarily interested in borrowing as much money as
possible. For this group the increased housing value is only an extra possibility to borrow more money, but not a motive as it is under the wealth effect. The third and last hypothesized
explanation for the co-movement between housing value and consumption, are macroeconomic factors that influence both housing value and consumption independently, e.g. productivity growth (Attanasio, Blow & Hamilton, 2009; Attanasio & Weber, 1994; King, 1990; Pagano, 1990).
For the Netherlands specifically, until the writing of this thesis, little research has been done on estimating the effect of fluctuation in housing wealth on private consumption. In 2017 the Bureau for Economic Policy Analysis (CPB) found that a 1 euro decrease in housing value led to an increase in savings between 0 and 5 euro cent in the period 2008-2013 (Bijlsma & Mocking, 2017). Abdulaziz (2018) has done a cross country analysis from sixteen industrialized countries including the Netherlands. This research found evidence for the wealth effect of 0.1% if housing prices increase by 1% (Abdulaziz, 2018). The Dutch National Bank (DNB) performed an analysis using macro data on the wealth effect and found a positive link between the two (DNB, 2018). Rabobank, one of the large Dutch banks, estimated a wealth effect of 4 cents, per 1
euro housing wealth increase (Lennartz & Spiegelaar, 2018). In the latter research, only data from Rabobank clients was used. In this thesis, data from the DNB Household Survey (DHS) between 2004 and 2017 is analysed to verify these results beyond the clients of the Rabobank, as it is a representative sample of the Dutch population (CentERdata, n.d.-a). Besides the wealth effect, also the credit constraints effect is tested. For all the different analyses in this thesis, macroeconomic prospect influences will be controlled for in a statistical way, which is explained in depth in the methodology chapter. Based on all of the above, the main research question is formulated, which can be answered through the five sub-questions.
How do house value fluctuations influence private consumption expenditures in the Netherlands?
I. To what extent does a relationship between housing value fluctuations and consumption expenditures exist in the Netherlands?
II. To what extent is the relationship between housing wealth and consumption caused by credit constraint individuals (i.e. collateral effect)?
III. To what extent is the relationship between housing wealth and consumption caused by the fluctuation in an individual’s permanent income (i.e. wealth effect)?
IV. Do individuals respond differently to house value increases than to decreases?
V. Do older homeowners adjust their consumption expenditures more as a result of housing value fluctuations, than young homeowners?
In the continuation of this thesis, the theoretical and the methodological methods are discussed. However, according to Checkel (2005), it is also recommendable to talk about the
philosophical or meta-theoretical level. The theoretical and methodical level almost always have a place in today’s papers, as it is needed to understand and possibly replicate the research. By contrast, the philosophical level does not necessarily have to be discussed to understand the research. However, decisions in the other two levels are influenced by the philosophical level in the background and are therefore recommendable to explain (Checkel, 2005). Among
contemporary economists (logical) positivism (LP) is the dominant philosophical theory. LP refers to the belief of researchers that there is a single reality which can be studied value free by researchers based on objective facts (Carson et al., 2001). This has led to the dominant approach in today’s microeconomics: the Neoclassical Economics (NE). This metatheory assumes that income-constrained individuals and credit-constrained firms try to maximise their own satisfaction (utility) based on fully informed evaluations of utility (Berg & Gigerenzer, 2010). However, individuals do not always make rational and fully informed decisions. Behavioural Economic scholars use methods from NE, but learn from psychological research, to take irrational behaviour into account (Berg & Gigerenzer, 2010). To uncover the objective truth, scholars often apply (economic) models on large empirical datasets to unveil behaviour patterns of people. This thesis follows this scientific stream to come to conclusions about the reality.
The remainder of this thesis is set out as follows. In the next chapter, the basic foundation behind the wealth hypothesis is explained, followed by an overview of the research that has been done to uncover the link between housing wealth and private consumption. This culminates into a theoretical framework which is used to outlay the relevant concepts in this thesis. The third chapter explains how this paper tries to uncover this link on a practical level. In this chapter is especially highlighted what identification methods are used and which variables and formulas are
needed for these methods. Moreover, a short description of the DHS dataset is given to provide the reader with some background information on the data. The fourth chapter presents and discusses the results of the analysis. The last chapter describes the conclusions of this study and how these relate to prior research. Furthermore, recommendations for further research and government policy are given.
2
Theory
This chapter, first discusses the work of Modigliani (1954) and the work of Friedman (1957), as it built the theoretical foundation of the link between housing wealth and consumption and this still is the basis for most studies on this topic today. Both theories are based on two different, but very similar concepts (i.e. Life Cycle and Permanent Income Hypothesis). The following section discusses previous research on the effect of fluctuations in housing wealth on the level of private consumption. In the 1980s, the first studies were published on this topic mostly based on a macro-level analysis which uses aggregated time series data. Research with a micro level approach entered the stage in the 2000s. The latter is often based on household surveys which provide the ability to account for heterogeneity between households. In macro-economic research, this is often a possible confounding factor as it is very difficult to precisely control for heterogeneity if time series data is used for the analysis. The discussion about previous research on this hypothesised relationship, will culminate into the theoretical framework which is used for the analysis.
2.1 Life Cycle-Permanent Income Hypothesis
In 1954 Modigliani contrives the concept: Life Cycle Hypothesis (LCH). This model presumes that individuals adjust their level of spending on the expectation of their future income
(Modigliani, 1954). Young people will lend money to be able to spend more than their current income, as they are expected to earn more in the future. As these individuals become older, their salary will increase and they are able to pay off their debt. When people have no more debts they save more for the time when they are retired and at that point they will consume their savings (Modigliani, 1954). In 1957 Friedman describes the concept of Permanent Income Hypothesis (PIH), which is very similar to the LCH from Modigliani (Friedman, 1957). This model assumes that individuals spend money based on their expected long term average income, also called: ‘Permanent Income’. People save money if their current income is higher than their permanent income and dis-save if their current income is lower (Friedman, 1957).
The LCH model takes into account a period of a person’s lifetime, meaning that no inheritance is left behind for his descendants. The implications for this are displayed in figure 2. At some point, the individuals (who own a house) have to sell it anyway, to maintain their level of consumption. It may seem at first as if the people with a higher amount of financial assets do not have to sell their houses as a result of their high value of financial assets. However, this group is also used to a high level of consumption and therefore needs to sell their houses to maintain this consumption level. If housing values go up, older people increase their spending more than younger people, as older people have a shorter expected lifetime left.
Under the PIH model, the period which is taken into account is infinite (Friedman, 1957). This leads to the practical expectation that individuals do not consume all their wealth in their own lifetime, but also leave an inheritance. For the wealth effect, this holds the implication that young and old people adjust their consumption behaviour the same way to wealth fluctuations, as the timespan is assumed to be infinite.
Figure 2: Life-Cycle Hypothesis for Wealth. Adapted from: Life-Cycle Hypothesis, T. Pettinger, 2017b, retrieved from: https://www.economicshelp.org/blog/27080/concepts/life-cycle-hypothesis, Copyright 2017 by
Economicshelp.org. Adapted with permission.
The LCH takes into account a finite period, the individual’s expected lifetime, leaving the option of a bequest at the end. From this point of view, older people would change their
consumption more if housing values fluctuate, as they have less time to consume this additional wealth. The next two sections discuss first the macroeconomic and then the microeconomic research on housing wealth effect.
2.2 Literature Review on Macroeconomic Research
In 1980, Elliott was the first researcher who studied the effect of non-financial wealth (housing wealth is a form of non-financial wealth) on consumption. This study did not find a significant relationship. However, this study broadly defined “non-financial wealth” compared to the narrow definition of housing wealth, which excludes other durable goods. This could have
influenced the outcome. Following the study of Eliott (1980), more researchers started to study this link. In the beginning, almost all of them used macro data to investigate this relationship (e.g. Muellbauer & Murphy; 1990; Poterba, 2000). Girouard and Blondal (2001) did not find significant results across countries (for some individual countries, they did find significant results). Ludwig and Slok (2004) were one of the first researchers to actually prove a consistent positive relationship between housing wealth, as well as financial wealth, on consumption by using data from sixteen OECD countries. Dvornak and Kohler (2007) did a similar kind of research for Australia, using state-level data, leading to a similar outcome. Mian, Rao and Sufi (2013) found a marginal propensity to consume (MPC)1 out of housing wealth of 5-7 cents per 1
dollar increase for the United States (U.S.). Aoki et al. (2004) and Aron and Muellbauer (2006) included the level of credit liberalization in their analysis for the United Kingdom (UK), both studies concluded that a higher level of liberalization leads to a higher MPC out of housing wealth. Iacoviello (2004) confirms this result for the U.S. The study of Case et al. (2005) estimated, using data on current income, stock and housing market wealth in the U.S., that a 1 per cent increase in wealth leads to 0.11 per cent increase in consumption. In 2013, Case,
Quigley and Shiller used new data to conduct the same estimation and ended up with an estimate of 0.08 per cent. Dong, Hio and Jia (2017) find evidence for the wealth effect in China,
depending on the price to income ratio. The latter is a more profound, or specific measure, of the collateral effect. Dong et al. (2017) find that a housing price-to-income ratio above 5.0882 leads to a much higher housing wealth influence on consumption than a ratio below this boundary.
1 Marginal Propensity in this case means the portion of increased consumption if housing wealth increases. This relationship is the same the other way around: the portion of decreased consumption if housing wealth decreases (Haavelmo, 1947).
This is explained by stating that in regimes with a ratio above 5.0882, people have a relatively expensive house compared to their income, which is probably paid for with a mortgage. Since people in such a regime have such a high mortgage, lending additional money is difficult for them (Dong et al., 2017). Abdulaziz (2018) and the DNB (2018) both found a positive wealth effect for housing on consumption. Having said that, for all research on this topic that uses time-series data, the biggest methodological problem is endogeneity. The factors that influence housing wealth can also influence consumption directly, mainly through overall macroeconomic prospects (Calomiris, Longhofer & Miles, 2009). The latter is a confounding factor, as the outcomes of the studies which use time series data, are also influenced by the correlation between income shocks and price fluctuations and not purely by the effect of housing wealth on consumption. This was found to be the main factor, driving significant results for research done for the late 1980s in the United Kingdom (King, 1990; Pagano, 1990)
2.3 Literature Review on Microeconomic Research
Attanasio & Weber (1994) were among the first researchers using microdata to analyse the link between housing wealth and consumption. They wanted to check the results from previous studies that had used macro data, in particular, that of Muellbauer and Murphy (1990) and King (1990) which found economic conditions to be the driving factor behind the co-movement between housing wealth and consumption. Attanasio & Weber (1994) confirmed this concluding using microdata. In contrast to the earlier studies, however, also evidence for the wealth effect was found. Furthermore, Attanasio & Weber (1994) found a significantly higher wealth effect for older than younger homeowners. Skinner (1996) found a small, but significant positive wealth effect, using microdata from the U.S. For younger individuals however, he found
a much higher wealth effect than for older people, since the latter are more cautious spending wealth gains (Skinner, 1996). In the same year, Engelhardt (1996) published a study that
estimated the MPC wealth effect to be around 0.03, but only for decreasing housing prices. This finding is remarkable as it goes against the expectations under the, in behavioural economics, very dominant loss aversion theory which states that people feel more pain from losing (in this case consumption), than the pleasure of gaining (Kahneman & Tversky, 1979). Engelhardt provides the explanation, that if housing values increase it is hard for individuals to turn their extra wealth into liquid capital, which is needed to increase consumption expenditures. Campbell and Cocco (2007) found a small positive effect for young-, and a large effect for old
homeowners, on their level of consumption to fluctuations in housing wealth. This study also analysed the impact of changes in housing price on renters, but did not find a significant effect, confirming the wealth effect (Campbell & Cocco, 2007). Attanasio et al. (2009) found, in contrast to Campbell and Cocco (2007), a stronger effect for younger than for older households. However, they did not make a distinction between the tenure status of the different age groups and therefore their results are not directly comparable to the study from Campbell and Cocco (2007). The study of Attanasio et al. (2009) contradicts the wealth hypothesis that, if housing prices go up, young households should be less or sometimes even negatively affected by this, as they have to save more to be able to buy a house, leaving less room to consume. Attanasio et al. (2009) point out that younger households are usually more vulnerable to changes in economic prospects and that the factors influencing these prospects also affect consumption and housing prices together. Therefore, the researchers interpret their results as proof for the common
can be through a direct effect, as the wealth increases individual’s future income, Calcagno, Fornero and Rossi (2009) find this effect to be dominant for Italy, Gan (2010) for Hong Kong and Wong (2017) finds evidence for this effect in New Zealand. Carroll et al. (2011) estimate the wealth effect to be 2 cents increased consumption for every 1 $ increase in housing value in the U.S. This estimation is confirmed by Khalifa, Seck and Tobing (2013). For the Netherlands, Lennartz and Spiegelaar (2018) find that a 1 euro increase in housing wealth leads to 4 cents increase in consumption. The CPB found that for individuals in the Netherlands a decrease of 1 euro leads to decreasing in saving between 0 and 5 cents on the short term (Bijlsma & Mocking, 2017). On the other hand, wealth can influence consumption through the (indirect) collateral effect, as increased wealth provides the possibility to borrow more money, e.g. Disney et al. (2010) find this to be true in the United Kingdom. Browning, et al. (2013) find evidence for the collateral effect in Denmark. Leth-Petersen (2010) confirms this result for Denmark, finding an even stronger relationship for younger than older individuals. Pan (2015) finds the collateral effect to be dominant for China. Atalay, Whelan and Yates (2016) find evidence for the collateral effect in Australia. Furthermore, they find a positive relationship between age and the influence housing wealth has on the individual’s consumption (Atalay et al., 2016). The former result for Australia is confirmed by Windsor, Jääskelä, and Finlay (2015).
2.4 Theoretical Framework
In most recent academic work, three different explanations for the co-movement between housing wealth and private consumption expenditures have been analysed, i.e. the wealth effect, the collateral effect (or credit constraint effect) and the common causality effect. In the analysis of this paper, first the existence of a significant pattern between housing wealth (independent
variable) and consumption (dependent variable) will be analysed, while correcting for the influence of economic prospects. From the literature review, it becomes clear that economic prospects are often a confounding factor in this type of research, because it influences both housing wealth and consumption. This relationship is expected to be positive as economic growth usually is accompanied by increasing housing values and consumption. If a pattern between housing wealth is indeed detected, the influence of the collateral and wealth effect will be distinguished. In figure 3 this is made visible.
Figure 3: Conceptual Framework thesis
Based on the literature, the relationship between housing wealth and the wealth effect is expected to be positive. Increasing housing values will lead to the perception of increased wealth, which individuals want to turn into consumption, vice versa for decreasing housing values. For
individuals who are credit constrained, increasing housing values create the possibility to borrow more money as they have more collateral to borrow against. For decreasing housing values, the opposite effect is hypothesized, in other words: a positive relationship is expected for the collateral effect. This hypothesized relationship will be distinguished from the wealth effect, to be able to estimate the effect of the true wealth effect. The two effects hold different policy implications and are therefore important to isolate from each other.
From the literature review we know that the strength of the wealth effect can depend on increasing or decreasing housing values. Based on the research of Engelhardt (1996) the expectation is that decreasing housing values lead to a larger MPC than an increase. Furthermore, in this paper, the effect for different age groups is estimated, to see if older individuals are affected differently by fluctuations in housing wealth than younger people. Age has proven to hold significance in different studies (Atalay, et al., 2016; Attanasio & Weber, 1993; Campbell & Cocco, 2007; Skinner, 1996). Furthermore, under the LCH model, older people would increase their consumption more as younger individuals if their housing wealth increases. Contrary, under the PIH it is assumed young and old people would respond the same to housing values. In general, more often evidence if found for a positive, than a negative relationship between age and the wealth effect.
All the concepts discussed in this paragraph are specified in measurable indicators in the next chapter, together with the explanation of the overall research design.
3
Methodology
This chapter explains what the indicators are for the different concepts from the theoretical framework and what research design is used for the analysis of these indicators. Based on the literature review of the previous chapter, for this thesis is chosen to use micro over macro data. This type of data provides the ability to control for unobserved characteristics of individuals, without having to identify all the characteristics. In the Netherlands, two different panel surveys are held amongst different individuals, the Longitudinal Internet Studies for the Social sciences (LISS) and the DHS survey. The former does not provide enough observation points for our research, therefore the DHS panel is used. In the next paragraph the DHS panel data is discussed, together with the explanation of the independent (housing wealth) variable. Furthermore, the construction of the consumption (dependent) variable is explained. Lastly, the empirical model, which is used to answer the sub-questions, is elucidated.
3.1 Data description DNB Household Survey (DHS)
The DNB Household Survey (DHS) is a panel study which is active since 1993 and is conducted every year (CentERdata, n.d.-a). In this survey, the few questions asked about consumption are not directly useable for this study. However, this survey contains much information about assets, debts and income, enabling the ability to construct a consumption variable. Before this construction is discussed in depth, the indicator that is used to measure housing wealth is explained, followed by the characteristics of this dataset.
For housing wealth, the assessed value for taxation purposes (this is called the WOZ value, in Dutch: Wet Waardering Onroerende Zaken) is used. The WOZ value is being determined by municipalities on an annual basis for every real estate property within their jurisdiction,
according to the Special Act for Real Estate Assessment (Wet WOZ). This act prescribes how every municipality should asses the value of houses and commercial real estate (Amsterdam, n.d.). For every municipality the same method is prescribed, guarantying a consistent value between municipalities in the Netherlands. This act also specifies that an independent organization must supervise and monitor the quality of the valuation. The Netherlands
Council for Real Estate Management (in Dutch: Waarderingskamer) is the organization which has been given this task (Waarderingskamer, n.d.)
Since 2004, households are being asked, in the DHS, what the WOZ value of their real estate properties is. The first time the WOZ value was reported by respondents was in 2004. This is therefore the first wave of the DHS panel that is being analyzed. Although the WOZ value is the estimation of the market value of the real estate property on the value reference date January 1st prior to the fiscal year, in this study I use the WOZ value of the year in which it is declared to the property owners. This is most suitable for this study, since the goal is to capture the effect of individuals at the time they know their housing value. Preferably also renters would be included in the analysis. This would provide the opportunity to filter out the effect of economic prospects, as for renters there would be no hypothesized effect on their income and thus every
co-movement of housing value and consumption would be due to economic prospects. Since 2016, renters also receive their computed WOZ value from the municipality. However, in the DHS survey, renters did not declare their WOZ value. But even if renters had declared their value in the survey for 2016 and 2017, it would probably lead to too less observations to find any significant effect. Since the effects of economic prospects cannot be known trough renters, they are controlled for in a statistical way, which will be discussed in the last paragraph of this chapter.
As said before, this thesis is based on a dataset that is constructed from survey data. The disadvantage from survey data is the reliability of the answers, as respondents may not always answer every question accurately or honestly. It is almost impossible to correct for this, but since there are many observation points in total, we rely on the rule of large numbers to correct for this. Table 1 presents some descriptive statistics. What stands out is the high number of males
compared to females. However, this is not a sign for an overrepresentation of males in the dataset, as in general, men are more likely to be homeowners than women (Vignoli, Tanturri & Acciai, 2016). Furthermore, table 1 reports a high amount of main wage earners among
homeowners. This is normal for a sample of homeowners, since most of these individuals are also the breadwinners for the household. The last notable number from table 1, is the net income. For the net income, 111 observations are negative. This follows from the fact that most values for net income are not filled in by individuals, but calculated by CentERdata staff (see:
CentERdata, 2017, p.5). Different questions are used to compute a net income, since for many studies the net income value is of importance, but is not filled in by respondents. Negative net incomes can be a result of negative values for profit (i.e. a loss for business owners, or a
relatively high amount of paid alimony. If this is the case for an individual observation depends on the answers the individual has filled in for the other financial questions. The reason for negative consumption values is explained in the following paragraph.
Table 1: Descriptive statistics of the sample used for the analysis (N individuals= 2,518, N observations= 9,828)
Minimum Maximum Mean SD
Individual demographics
Year of birth 1915 1994 1958.17 15.75
Sex (1 = Male, 2 = Female) 1 2 1.30 0.05
Highest level of
Education (1=Kindergarten, 6=University 1 6 4.23 0.6
Household size 1 9 2.43 1.22
Children present 0 7 0.62 1.04
Composition of the household
Couple without children 0.48
Couple with children 0.29
Single 0.19
Other 0.04
Degree of Urbanization (1 = very high, 5 = very low) 1 5 3.16 .3
Occupation
Employed 0.60
Freelance 0.04
Unemployed 0.03
(Partly) disabled 0.03
Looking for work 0.02
Own business 0.01
Table 1 (continued)
Minimum Maximum Mean SD
Economic status
Main wage earner (yes = 1, no = 0) 0 1 0.88 0.32
Net income (€) -4800 689704.1 31148.1 13041.7
Consumption (€) -288146 365828.4 30442.3 36015
WOZ value (€) 10000 249292.1 1100000 115485
3.2 Consumption Construction
In the DHS panel, no direct questions are asked about consumption that can provide an estimation of the consumption of individuals. However, questions about assets, debts and income are asked which can be used to construct a consumption variable. The basic strategy is to take to net income or net profit (if an individual is self-employed) (𝐶𝑡 = 𝑁𝑡) and subtract the total property in- or decrease (𝐶𝑡 = 𝑁𝑡− 𝑃1) that has taken place between year 0 and year 1 (∆P1 =
Pt− Pt−1), and add the in-or decrease in total debt (𝐶𝑡 = ∆𝑁1− ∆𝑃1+ ∆D1) between year 0 and year 1 (∆D1 = Dt− Dt−1). The appendix clarifies which indicators are used to compute the
fluctuation in property (Appendix A) and debt (Appendix B). The starting point is the net income, as this is the main source for individuals to consume. If individuals buy financial products or store it in their bank account, they have less money to spend on consumption
products, therefore the change in financial wealth is subtracted from their net income. Contrary if more debt is taken on, this increases the room for consumption expenditures and is therefore added to net income. Stocks and funds are corrected for the average value change per year on the Amsterdam Exchange Index (AEX). This is done to take into account the big increase in value on stock markets all over the world, in the years towards the financial crisis of 2008 and the large
decrease in value in the years after the crisis. The AEX is the most important stock exchange in the Netherlands and that is why the fluctuations for this exchange are used, as most Dutchmen have stocks listed on this exchange, if they have stocks at all. The construction of the
consumption variable does offer an useful estimation of the consumption expenditures of
individuals however, 730 negative values are computed. This is partly due to negative values for net income, as reported in table 1. However, this still leaves 619 negative consumption values for 507 different individuals unaccounted for. These negative values can be the result of a mistake in the construction design of the consumption variable or due to inaccurate answers by respondents. The latter is a result of individuals making mistakes in reporting their finances in surveys, which can lead to negative values. To rule out the possibility of mistakes in the design of the
consumption variable, the thesis supervisor and the literature is consulted about possible problems when taking on a similar approach, both did not lead to the finding of mistakes in the design of the variable. Therefore it is assumed that in general, the value of the consumption variable is correct, but in some individual cases does not reflect the consumption expenditures precisely. The latter does not seem to be a very big problem, because even if not all the values are correct, there are enough observations in the sample to balance out this deviation.
3.3 Empirical Model
The basic model that is used in this thesis to analyse the panel data is a fixed effects model. This model estimates the relationship between the independent and dependent variables using the aggregated changes over time for each person. In other words: the model creates a unique constant for each individual, which takes into account the unique characteristics of the individual (Torres-Reyna, 2007). This model is chosen to maximally benefit from the advantages
of the panel data, namely the different waves the same respondents answer the questions. The dependent variable (consumption) and the independent variable (WOZ value) are transformed to their inverse hyperbolic sine value (IHS) of the natural log. This transformation provides the same results and benefits as the natural log value would, but also allows for negative values. The big advantages of transforming the dependent and independent variable into their IHS value, is that it reduces the impact of outliers in the analysis, as it converges the values and it gives us a result that we can read in terms of percentage change.
The formula that is used to transform the Consumption variable is 𝑌𝑖𝑡 = log (𝑦𝑖𝑡+ (𝑦𝑖𝑡2 + 1)12)
For the WOZ variable, the same transformation formula is used to maintain a comparable scale. 𝑊𝑖𝑡 = log (𝑊𝑖𝑡+ (𝑊𝑖𝑡2+ 1)12)
The basic equation (1) that is used for the estimation is 1) 𝑌𝑖𝑡 = 𝛼𝑖+ 𝛽𝑊𝑖𝑡 + 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in years and 𝑖 is an individual; 𝛼𝑖 captures the unobserved time-invariant
individual effect and controls in this way for the different starting position for each individual. Variable 𝜋𝑡 represents year dummies for each year (except for the first) absorbing in that way the time trends (Lewis, 2015), as the aim of this study is to only capture the influence of housing wealth on consumption without economic prospects which influence both. 𝑊𝑖𝑡 denotes the IHS
transformation of the natural log of the WOZ value, for each individual 𝑖 in year 𝑡; 𝑌𝑖𝑡 reflects the natural log of the IHS transformation of consumption expenditures at a point 𝑡 in time for an individual 𝑖; 𝛽 is the coefficient of main interest as it estimates the influence of WOZ value
fluctuations (housing wealth) on consumption. Lastly, the error term is displayed as 𝜀𝑖𝑡. Equation 1 is the basic formula for the estimation of the influence of housing wealth on private
consumption. This equation is applied to the total sample to answer the first sub-question. I. To what extent does a relationship between housing value fluctuations and
consumption expenditures exist in the Netherlands?
For the other estimations also equation (1) is used, but applied to different subgroups of the total sample, to allow for a more in depth analysis. These subgroups are a result of a division for the total sample, into two or three groups at a certain boundary (e.g. younger and older than 50 years old). To ensure the estimations between the different subgroups for the same division are significantly different from each other, an Ordinary Least Square2 (OLS) regression is
applied. Using an OLS regression instead of a FE model to verify if the different subgroups respond differently to housing wealth fluctuations, may perhaps seem strange at first glance. However, if a FE model would be applied, the variation through difference in age, for instance, would be absorbed by the FE model itself. This creates a situation in which it is impossible to detect the real effect difference between the two age groups. Therefore a OLS regression is used, which is explained in detail for each set of subgroups based on the same division.
In the observed period, it is possible that some people have moved to another house. To ensure the correlation between housing wealth and consumption is not, to a large extent, caused by people moving into another house, equation (1) is applied again. However, this time the people that have moved to another house in the observed period, are left out of the analysis. The
2 For increasing and decreasing WOZ values the FE model is used in a different manner, this will
two variables that are used to derive which people have moved are: WO53 (has the individual moved to another province) (CentERdata, 2017) and WOD52L (is the individual certain to move away in the coming two years) (CentERdata, 2017). Furthermore, OLS regression (2) is used to assure the estimated coefficient for the people that have stayed in the same house is significantly different from the estimated effect for the people that have moved between 2004 and 2017.
2) 𝑌𝑖𝑡 = 𝛼 + 𝛽1𝐷𝑖𝑡 + 𝛽2𝑆𝑖+ 𝛽3𝐷𝑖𝑡𝑆𝑖 + 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in year and 𝑖 denotes an individual; 𝑌𝑖𝑡 reflects the change between year 1 and 0, for the natural log of the IHS transformation of the constructed consumption variable; 𝛽1𝐷𝑖𝑡 is the effect of the difference in WOZ value between year 1 and 0, on consumption; 𝑆𝑖 is a dummy variable which is one if people have moved and zero if people did not move to another house; 𝛽2 denotes the effect of people that have moved on consumption, if the difference in WOZ value between year 1 and year 0, is zero. Time patterns are controlled for trough time dummies which 𝜋𝑡 reflects. The error term is indicated by 𝜀𝑖𝑡; 𝛽3𝐷𝑖𝑡𝑆𝑖 displays the effect of the difference in WOZ value for the people that have moved, compared to the people that did not move to another house. This is factor of main interest, as it shows to what extent the coefficient for the people that have moved significantly differs from people that did not move. The latter result, will be briefly discussed in the Results chapter, all the outcomes from equation (2) will be reported in the appendix.
To answer sub-question II, the credit-constrained individuals have to be identified, which is difficult. In real life, borrowing money depends on many different factors, such as: income, type of employment, interest rates, assets and age.
II. To what extent is the relationship between housing wealth and consumption caused by credit constraint individuals (i.e. the collateral effect)?
For this research, it would be one step too far to make a model which incorporates all of the things that influence borrowing capacity for individuals. Even if all of the relevant factors are identified, it would be highly doubtful that all of these factors are known for the individuals in the sample. There is no need for a complex model, since for this study it is enough to divide the sample in a group that contains individuals that are certainly not credit constrained and a group, that certainly contains the people who are constrained. For this study this is complex enough, since we only want to compare those two groups and it is not vital to identify precisely the individuals that are constrained. The division between the two groups is made using the net income to debt ratio. This ratio is certainly not a perfect measure, however for money lenders this is one of the main factors when assessing if a person can borrow additional money. The individuals on the lower end on the scale of this ratio as a group, definitely contain the
individuals with problems borrowing money. The boundary between the two groups is set on the 25% lowest to highest ratios. This boundary is somewhat arbitrary, but since no boundary is known in the literature, this percentage is chosen based on the assessment of the researcher and his supervisor. The choice is made to include individuals in the sample of credit-constrained individuals, even if they have a ratio that meets the criteria for only one year. This is done to ensure that the most credit constrained individuals are included in this group, who at least once were very likely to have problems borrowing money. Equation (1) is first applied to the
individuals belonging to the 25% lower of the lower end of the net income debt ratio and secondly for the income to debt ratio that is part of the 75% higher end of the sample. In this
paper, the aim is to distinguish the influence of the two effects. The credit constrained individuals are hypothesized, to be influenced by both the wealth and collateral effect, as increased housing value offers them a possibility to borrow more money against. For the non-credit constrained group, only the wealth effect is assumed to be of influence, since these
individuals already can borrow additional money, regardless if their collateral increases in value. To distinguish the influence of the wealth and collateral effect on consumption, the estimation for the non-credit constrained group is subtracted from the estimation for the credit constrained group, while correcting for the difference in group size. The outcomes of the combined
individual effects are compared to the outcome for the full sample, to check if the two effects indeed account for the co-movement between housing values and consumption expenditures. Dong et al. (2017) use the housing price-to-income ratio as a similar indicator for the collateral effect. This seems an unexpected measure to identify the credit-constrained
individuals. However, Dong et al. (2017) find that regimes where the ratio is above 5.0882, the MPC out of housing wealth is significantly higher than below this boundary because of credit constraints. The outcome from Dong et al. (2017) has not been applied in micro-level research and is therefore only used to check the prior result for the collateral effect in this thesis. The last paragraph of this section explains this in more depth.
Moreover, the effect for the group that is not credit constrained (the other 75%), is actually the group for which the collateral effect is not relevant, since they already can borrow more money if they want to. In other words, this is the group that is only affected by the wealth effect. To
answer sub-question (III), this group of people is used, as it provides an estimation of the wealth effect solely.
III. To what extent is the relationship between housing wealth and consumption caused by the fluctuation in an individual’s permanent income (i.e. the wealth effect)? To verify if the coefficient for the group that is credit constrained is significantly different from the group that is not credit constrained, equation (3) is used.
3) 𝑌𝑖𝑡 = 𝛼 + 𝛽1𝐷𝑖𝑡 + 𝛽2𝐿𝑖+ 𝛽3𝐷𝑖𝑡𝐿𝑖+ 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in year and 𝑖 denotes an individual; 𝑌𝑖𝑡 reflects the change between year 1 and year 0 in the natural log of the IHS transformation of consumption expenditures, 𝛽1𝐷𝑖𝑡 is the effect of the difference in WOZ value between year 1 and year 0 on consumption; 𝐿𝑖 is a dummy variable which is one if the individual is identified to be credit constrained and zero if not; 𝛽2 is the effect of people that are credit constrained on consumption, if the difference in WOZ value between year 1 and year 0, is zero. Time patterns per year are absorbed by time dummies which is reflected by 𝜋𝑡. The error term is represented as 𝜀𝑖𝑡; 𝛽3𝐷𝑖𝑡𝐿𝑖 denotes the effect
of the difference in WOZ value for the people that are credit constrained compared to the non-credit constrained people. This is the factor of main interest, as it shows to what extent the effect for the people that are credit constrained is significantly different from the people that are non-credit constrained. In the Results chapter, the latter outcome will be highlighted, as it is the result of interest. The appendix will also include all the other outcomes for equation (3).
Following the paper of Engelhardt (1996) a distinction between increasing and decreasing housing values is made to answer sub-question IV.
IV. Do individuals respond differently to house value increases than to decreases? This is done at the level of the individual, instead of the macro level, to precisely isolate the expenditures in consumption at the moment individuals know their house has gone up or down in
value. Every WOZ value for an individual higher than the one in the previous year, is used in combination with the value prior to the increasing value. For decreasing housing values, the same method is used as for increasing values. In contrast to the estimations for the other subgroups, besides FE equation (1) not an OLS regression but another (FE) regression (4) is applied. This is done, because the division between the two groups, is unlike for the other
subgroups, not a division between individuals, but between observations of the individual. In this situation, a fixed effects model is preferable to measure, if the estimations for both groups
significantly differ from each other. It is closest to the (FE) model that estimates the coefficients in the first place (equation 1), which ensures consistent results. Contrary to the other groups, it does not create the problem that the (FE) absorbs the variance between the groups, if it is used to check if the groups are significantly different affected by housing wealth fluctuations. The reason for this is, that the variation which is used to make a distinction, is based on only two
observations not on a division between individuals. This means that two observations of one individual can be assigned to the ‘increasing group’ and two other observations to the
‘decreasing group’. The factors that determine if a WOZ value increases or decreases, are not related to an individual, as they are a result of market conditions which are randomly distributed per individual. Regression (4) is mathematically displayed as:
4) 𝑌𝑖𝑡 = 𝛼𝑖 + 𝛽1𝑊𝑖𝑡 + 𝛽2𝑀𝑖 + 𝛽3𝑊𝑖𝑡𝑀𝑖 + 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in year and 𝑖 denotes an individual; 𝑌𝑖𝑡 reflects the natural log of the IHS transformation of the consumption variable; 𝛽1𝑊𝑖𝑡 is the effect of the difference in WOZ value between year 1 and year 0 on consumption; 𝛼𝑖 captures the unobserved time-invariant individual effect; 𝑀𝑖 is a dummy variable which is one if the WOZ value is decreasing and zero if it is
increasing; 𝛽2 is the effect of decreasing WOZ values on consumption, if the difference in WOZ value between year 1 and year 0, is zero. Time patterns per year are absorbed by time dummies which is reflected by 𝜋𝑡.The error term is indicated by 𝜀𝑖𝑡. 𝛽3𝐷𝑖𝑡𝑀𝑖 displays the effect of the
difference in WOZ value increasing WOZ values compared to decreasing WOZ values. This is the factor of main interest, as it shows to what extent the coefficient for decreasing WOZ values is significantly different from increasing values. The latter result will be briefly discussed in the results chapter and all the outcomes from equation (4) will be reported in the appendix.
To answer sub-question (V), the sample is split into three age groups, following the paper of Attanasio and Weber (1993) and Atalay, et al. (2016) to analyse the effect per age group.
V. Do older homeowners adjust their consumption expenditures more as a result of housing value fluctuations, than young homeowners?
The group young people is between 20 and 39 years old, the middle-aged between 40 and 59 and the last group is 60 years or older in 2004. For each group equation (1) is used to analyse the effect for individuals born after 1961, between 1943 and 1961 and before 1943. To verify if the coefficients for the three different age groups are significantly different from each other, equation (5) is used.
5) 𝑌𝑖𝑡 = 𝛼 + 𝛽1𝐷𝑖𝑡+ 𝛽2𝐾𝑖+ 𝛽3𝐷𝑖𝐾𝑖 + 𝛽4𝑂𝑖+ 𝛽5𝐷𝑐𝑂𝑖 + 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in year and 𝑖 denotes an individual; 𝑌𝑖𝑡 reflects the change between year 1 and year 0 in the natural log of the IHS transformation of consumption expenditures; 𝛽1𝐷𝑖𝑡 is the effect of the difference in WOZ value between year 1 and year 0 on consumption; 𝐾𝑖 is a dummy
variable which is one if the individual is born between 1943 and 1961 and zero if not; 𝛽2 is the
in WOZ value between year 1 and year 0 is zero. 𝛽3𝐷𝑖𝑡𝐾𝑖 displays the effect of the difference in WOZ value for people being born between 1943 and 1961 compared to individuals born after 1961. This one of the two factors which are especially, as it shows to what extent the coefficient for people being born between 1943 and 1961 is significantly different from individuals born after 1961. Time patterns per year are absorbed by time dummies which is reflected by 𝜋𝑡.The
error term is represented as 𝜀𝑖𝑡. 𝑂𝑖 again is a dummy variable, which is one if the individual is
born before 1943 and zero if not; 𝛽4 is the effect of people individuals being born before 1961 on consumption, if the difference in WOZ value between year 1 and year 0, is zero. 𝛽5𝐷𝑖𝑡𝑂𝑖
displays the effect of the difference in WOZ value for people being born before 1943 compared to individuals who were born after 1961. This the second main factor of interest, as it shows to what extent the coefficient for people being born before 1943 is significantly different from individuals born after 1961.
Furthermore, since it is hard to know beforehand if the latter division in age groups is useable for our sample, the group is split between young and old individuals at the age of 50 in 2004. This means that the effect of housing wealth on consumption for people born in 1953 and later is analysed individually and the effect on people born before 1953 is individually analysed using equation (1). For this division, it is also important to analyse if the coefficient for the ‘younger than 50’ group is significantly different from the ‘older than 50’ group. Equation (6) is applied to the sample for this purpose.
6) 𝑌𝑖𝑡 = 𝛼 + 𝛽1𝐷𝑖𝑡+ 𝛽2𝐹𝑖+ 𝛽3𝐷𝑖𝑡𝐹𝑖 + 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in year and 𝑖 denotes an individual; 𝑌𝑖𝑡 reflects the change between year 2 and year 1 in the natural log of the IHS transformation of consumption expenditures, 𝛽1𝐷𝑖𝑡 is the
effect of the difference in WOZ value between year 1 and year 0 on consumption; 𝐹𝑖 is a dummy variable which is one if the individual is born before 1953 and zero if the birth year is after 1953; 𝛽2 is the effect of people born before 1953 on consumption, if the difference in WOZ value between year 1 and year 0, is zero. Time patterns per year are absorbed by time dummies which is reflected by 𝜋𝑡. The error term is indicated by 𝜀𝑖𝑡. 𝛽3𝐷𝑖𝑡𝐹𝑖 displays the effect of the difference
in WOZ value for the born before 1953 compared to the people born after 1953. This is the factor of main interest, as it shows to what extent the coefficient for the people that are born before 1953 is significantly different from people born after 1953. In the results chapter, the latter outcome will be highlighted, as it is the result of interest. The appendix will also include all the other outcomes for equation (6).
The paper by Dong et al. (2017) states that for individuals with a housing price-to-income ratio above 5.0882 the influence of housing price increases is much higher than for ratios below this number. Dong et al. (2017) use this ratio as a comparable indicator for credit constraint individuals. As mentioned earlier in this paper, it is not very clear were the boundary lies between credit constraint individuals and non-credit constraint individuals. Dong et al. (2017) found a boundary at 5.0882 for the housing price-to-income ratio, but this has not yet been confirmed in other research. Therefore, this boundary is only used to verify our previous results for the credit constraint group. The same method for the division between the credit constrained and the non-credit constrained group is used, as before. That is, even if an individual has only for one year a ratio that meets the criteria, he falls in the credit constrained group. Equation (1) is first applied to the group individuals that at least once had an observed housing price-to-income ratio that was above 5.0882. This is also done for the group of individuals who never had a
housing price-to-income ratio above 5.0882 to compare the effect of housing wealth fluctuations on consumption for both groups. Besides applying FE equation (1) for both ratio groups, OLS regression (7) is used, to verify if the estimated coefficient for the group that has a housing price-to-income ratio above 5.0882 is significantly different from the group with only ratios below this boundary.
7) 𝑌𝑖𝑡 = 𝛼 + 𝛽1𝐷𝑖𝑡+ 𝛽2𝑅𝑖+ 𝛽3𝐷𝑖𝑡𝑅𝑖 + 𝜋𝑡+ 𝜀𝑖𝑡
Where 𝑡 is time in year and 𝑖 denotes an individual; 𝑌𝑖𝑡 reflects the change between year 1 and year 0 in the natural log of the IHS transformation of consumption expenditures; 𝛽1𝐷
𝑖𝑡 is
the effect of the difference in WOZ value between year 1 and year 0 on consumption; 𝑅𝑖 is a dummy variable which is zero if the individual has had a housing price-to-income ratio above 5.0882 and zero if the individual never had a ratio in above 5.0882. Time patterns per year are absorbed by time dummies which is reflected by 𝜋𝑡. The error term is represented by 𝜀𝑖𝑡. 𝛽2 is
the effect of people with a ratio above 5.0882 on consumption, if the difference in WOZ value between year 1 and year 0 is zero. 𝛽3𝐷
𝑖𝑡𝑅𝑖 displays the effect of the difference in WOZ value for
people with a ratio below 5.0882 compared to the people with a ratio above this number. This is the factor of main interest, as it shows to what extent the coefficient for the two groups are different from each other. In the results chapter the latter outcome will be highlighted, as it is the result of interest. The latter result will be briefly discussed in the results chapter and all the outcomes from equation (7) will be reported in the appendix.
In this chapter has been made clear what the research design and approach is to analyse the relationship between housing wealth and consumption expenditures.
4
Results
In this chapter, the results of the analysis are reported in tables and discussed in text. The order in which the different results are explained, is the same as that of the sub-questions of this paper. Before the result for the sub-question is discussed, the question itself is repeated to remind the reader what it was.
I. To what extent does a relationship between housing value fluctuations and consumption expenditures exist in the Netherlands?
In the first column of table 2, a positive and highly significant coefficient of 0.21 is reported for the relationship between housing wealth and consumption expenditure. This
coefficient means that an increase in housing wealth of 1% is associated with a 0.21% increase in consumption spending, under the ceteris paribus assumption. Since this number is highly
significant, it is assumed that this number is not a result of coincidence but, that there is indeed a relationship between the two. Column 2 of table 2 shows, that even if the individuals that have moved to another house in the observed period are removed from the sample, the result holds. The significance in the second column is a little lower than in the first column, which is most likely a result of the decreased sample size. However, the estimation still is almost the same as in the first column, 0.23% instead of 0.21%. The people that moved, have created a small
downward bias in the result for the full sample, but this effect is only small 0.02% (0.23-0.21). Furthermore, from the OLS regression (Appendix C) we can conclude that the coefficients reported in column (1) and (2) of table 2 are not even significantly different from each other. In other words: it is not even certain that the estimations are actually the same and that the
TABLE 2: Main results, influence of housing value on private consumption (Fixed effects model) (1) Full sample (2) Sample Stay (3) Sample Credit Constrained (Collateral and Wealth effect) (4) Sample Not Credit Constrained (Wealth effect) (5) Sample Growth (6) Sample Decrease
Dependent variable is the IHS transformation of Consumption IHS transformation of the WOZ value .21***
(.06) .23** (.08) .35*** (.13) .18** (.08) .22*** (.07) .13 (.09)
Year fixed effects Yes Yes Yes Yes Yes Yes
Number of observations 9,828 7,640 2,907 6,921 7,019 4,9273
Number of individuals 2,532 2,039 920 1,610 2,406 1,393
Note: ***, **, * denote significance at the 1%, 5%, 10% level, respectively. The numbers between parentheses denote standard errors.
Sub-question II and III are discussed together, as it is hypothesised in the methodology chapter, that the two combined cause the co-movement between housing wealth and
consumption, if macroeconomic prospects are controlled for.
II. To what extent is the relationship between housing wealth and consumption caused by credit constraint individuals (i.e. the collateral effect)?
III. To what extent is the relationship between housing wealth and consumption caused by the fluctuation in an individual’s permanent income (i.e. the wealth effect)? From column 3 of table 2 it becomes clear that, as was hypothesized, credit constrained individuals increase their consumption much more (0.35), than non-credit constrained people (0.18) to housing wealth fluctuations. The credit-constrained individuals are exposed to both the wealth and collateral effect. Non-credit constrained people are only influenced by the wealth
3The combined observations and individuals combined are higher than the number of observation and individuals for the full sample, since for the decreasing and increasing also the value prior to the decrease or increase is included to maximize the FE model effectiveness.
effect, because this group already could borrow more money, prior to their house value increase. This means, that the estimation in column 4 only measures the wealth effect (0.18) and that the collateral effect can be deducted as the difference between the two groups, if the difference in sample size is accounted for. After this correction, the effect for the credit-constrained
individuals is 0.204. The difference in effect between the two groups is 0.025. The strength of the
relationship between housing wealth and consumption for the collateral effect, is hence estimated as 0.02%. The collateral and wealth effect combined account for 0.20 of the total 0.21 MPC from column 1. This leaves 0.01 unaccounted for. The reason for this is not very clear, but can be due to the model not being 100% accurate. However, it is possible that the actual measure of the wealth effect is 0.19%, since the wealth effect (0.18) is only measured at a 95% confidence level. If this would be the case, it would leave us with a more conclusive outcome, unfortunately we cannot be sure if this is true, based on this study. However, the unaccounted part is relatively small, hence, it is assumed that the model in general is right, although it is not 100% accurate. Furthermore, this statement is supported by the result in Appendix (D), which shows that the effect for the credit constrained group is significantly different from the non-credit constrained group.
IV. Do individuals respond differently to house value increases than to decreases? Column 5 of table 2, shows a similar estimation for the influence of housing wealth on consumption expenditures for increasing housing values (0.22), as the main result in column 1. For decreasing values, a highly insignificant (lower) estimation (0.13) is found, even at a 90%
4 Calculation: (920:1610*0.35) 5 Calculation: (0.20-0.18)
confidence level (column 6 of table 2). Moreover, the results from Appendix (E) show, that there is no significant difference detectable between the effect for increasing, or decreasing values. Based on these estimations, we can only conclude that we have found no evidence for a different effect for increasing and decreasing housing values on consumption. The most straightforward explanation for this is, that the hypothesised difference between increasing and decreasing housing values in effect, does not exist. On the other hand, it could be possible that the applied model was not sensitive enough, or needed more data, to detect a different effect. Based on this research, it is not certain what the suitable explanation is. All things considered, it is only certain that this study did not lead to the finding of evidence for the hypothesis, that increasing and decreasing housing values have a different effect on consumption.
V. Do older homeowners adjust their consumption expenditures more as a result of housing value fluctuations, than young homeowners?
In table 3, results are presented for different age groups. The first column of this table repeats the main result, to be able to quickly compare it with the results in table 3. In column 2, 3 and 4, the results of the same age cohorts, as used by Attanasio and Weber (1993) and Atalay et al. (2016), are reported for the DNB survey. For young (column 4) and old homeowners (column 6), the relationship is a little higher (0.26 and 0.25) than the main result in column 1 (0.21). The estimation for the middle-aged individuals (0.10) in column 5 is remarkable, as it is highly insignificant and also relatively low. These results suggest, that young and old individuals
increase their consumption more if housing wealth increases than the average effect of 0.21. This can be caused by the middle-aged group, as it is not even that the individuals respond at all to housing wealth fluctuations. The latter would be surprising, since this is the largest subgroup for
this age division in the analysis. However Appendix (F) shows that, it is not certain the
estimations for the three age groups are significantly different from each other. This means that it is possible that the three different age groups are affected to the same extend by housing values. Statements about how housing wealth fluctuations affect consumption behaviour of the different age groups are not possible. However, the results of this thesis lean into the direction, that young and old individuals change their consumption to a large extent to housing wealth fluctuations, whereas the middle-aged individuals barely respond. This study finds very weak evidence which point into this direction. Future research could incorporate this aspect, in order to verify or reject this idea.
TABLE 3 – Influence of housing wealth on consumption spending for different age groups (Fixed effects model) (1) Full sample (2) Sample Under 40 (3) Sample Between 40 and 60 (4) Sample Over 59 (5) Sample Under 50 (6) Sample Over 50
Dependent variable is the IHS transformation of Consumption IHS transformation of the WOZ value .21***
(.06) .26** (.11) .10 (.09) .25* (.14) .17** (.09) .17* (.09)
Year fixed effects Yes Yes Yes Yes Yes Yes
Number of observations 9,828 2,982 4,518 2,328 4,714 5,114
Number of individuals 2,532 1,092 1,006 434 1,534 998
Note: ***, **, * denote significance at the 1%, 5%, 10% level, respectively. The numbers between parentheses denote standard errors.
In columns 5 and 6, the same coefficient for people older than 50 is reported, as for individuals who are 50 years and under (0.17). Following the LCH model, the expectation was to find a higher coefficient for older than the younger group. Following the PIH model, no difference between age cohorts was assumed. The expectations of the PIH model best fit the results in
column 5 and 6, as they estimate the same effect for people above and under 50 years old. This statement is supported by Appendix G, which proves that there is no difference between the effect for the two different age groups. However, if we look back at the results in column 2,3 and 4, it is possible that the division in two groups at the age of 50 is too broad. From the division into three age groups, it seems as if, between the age of 40 and 60, something interesting is going on. This age group does not seem to respond to housing wealth fluctuations. Dividing the sample into groups of under and over 50 years old, could have led to a model, which is not sensitive enough to this possible phenomenon. In the future, scholars could split the middle-aged group up in smaller groups, to get a better understanding of the behavior of this group to wealth changes. Table 4 follows the logic of the paper of Dong et al. (2017), in which a relationship is found between the housing price-to-income ratio and the level of MPC from housing wealth. The first column is used to repeat the main result of table 2 once more, as it can be helpful to compare it to the estimations in table 4. Column 2 of table 4, indicates that individuals with a housing price-to-income ratio above 5.0882 have a higher MPC from housing wealth (0.35) than the individuals with ratios below this number in column 3 (0.18). Appendix H supports this result, as it shows that the effect for the people with a ratio above 5.0882, is significantly different from the
estimation for the people below this ratio. As explained in the methodology chapter, the housing price-to-income ratio is used as another indicator to verify our results for the collateral and wealth effect. People with a housing price-to-income ratio above 5.0882 are assumed to have higher debts, which leads to a smaller chance they can loan additional money. Hence, both the collateral and the wealth effect is hypothesized to affect this group. For the wealth effect, the alternative strategy leads to the same estimation (0.18) as the baseline approach. However, the
