Linking House Price Changes, Mortgage Debt and
Consumption Spending: New Evidence from a Micro
Panel in the Netherlands
Author: Lisanne Spiegelaar1
RUG supervisor: prof. dr. D. J. Bezemer Rabobank supervisor: dr. C. Lennartz
Master’s Thesis2
MSc Economics & MSc Finance Faculty of Economics and Business
University of Groningen
Abstract: Exploiting a unique dataset from the largest Dutch mortgage lender, Rabobank, this
paper constructs consumption measures based on bank account transactions to estimate the elasticity of consumption with respect to house prices. Potential heterogeneity in the relationship based on mortgage debt levels as proxied by LTV ratios and between different age and income groups is taken into account, as well as possible differences in the relationship between regions. The findings show an estimated elasticity of between 6.5 to 7, which results in a marginal propensity to consume of 3 to 4 cents per euro increase in housing value. The elasticity is significantly higher for high income groups and differs between provinces, with the highest elasticity observed in the province with the capital city and the largest house price volatility: Noord-Holland. Mortgage debt levels do not appear to influence the relationship, indicating that Dutch households base their consumption response on gross wealth and place their debt in a mental account to which they pay less attention.
JEL classification: D14, D15, G4, R31
Keywords: household debt, house prices, consumption, household behavior, life-cycle,
wealth effects
1 – INTRODUCTION
Projections by the Dutch central bank (DNB) show that the recovery of the housing market and its actual price developments caused the increase in private consumption since 2013 to be 60 percent higher than it would have been without the housing market recovery3. Considering the effect of this recovery on corporate investment, real disposable income, consumer confidence and unemployment leads to the conclusion that more than a quarter of the increase in GDP since 2013 can be attributed to the recovery in the housing market (DNB
Economic Developments and Outlook December 2017). This example shows that increasing
house prices can amplify the upswings of the business cycle, however, the recession of 2012 has shown that the opposite also holds true: falling house prices depressed consumption through the home equity erosion of households, amplifying the downturn of the business cycle (DNB
Economic Developments and Outlook December 2012). Volatility in the housing market is thus
an important determinant of national economic developments, mainly through its effect on private consumption. Understanding the channels through which housing market developments affect consumption offers scope for targeted interventions when necessary and allows for a clearer understanding of the effects of policy changes in the housing market, such as the gradual abolishment of mortgage interest rate deductibility agreed upon by the Dutch government in 2013.
increase in house prices positively affects private consumption through an increase in consumer confidence or an increased feeling of being wealthy. This effect is likely to be strongest during housing market upswings and in areas with excessive growth, as these are prone to more press coverage. This suggests that the relationship between house prices may be asymmetric over time, depending on the direction of the housing market, and may differ regionally based on local house price developments.
Several theories provide a formal explanation for the relationship between housing wealth and consumption. Most importantly, the life-cycle model developed by Modigliani and Brumberg (1954) explains the first channel; the wealth effect. Individuals want to smooth consumption and therefore base it not only on current income and wealth, but on the present value of expected total lifetime income and wealth. Hence, future changes in housing wealth are taken into consideration when deciding on consumption. The life-cycle model supplemented with liquidity constraints accounts for the second channel; the collateral effect. In the life-cycle model, individuals treat all parts of current and future income and wealth similarly, resulting in a positive predicted relationship between housing wealth and consumption, whereas in real life they may treat increases in housing wealth differently than, for example, increases in income. This is because households may view housing wealth as a component of wealth that cannot be easily consumed, something which is accounted for in the behavioral life-cycle theory. This theory, developed by Shefrin and Thaler (1988), is based on the behavioral finance concepts of mental accounting and frame dependence. The behavioral life-cycle theory can still predict a positive relationship between housing wealth and consumption, but the underlying argument for this relationship is fundamentally different. Other theories influencing the relationship between housing wealth and consumption are the precautionary savings theory, where households use their housing wealth as savings for times of financial hardship, and the bequest motive, where households do not spend a large fraction of their housing wealth gains, as they plan to bequeath the wealth to their children.
guilder. This paper is the first that directly studies the relationship between house prices and consumption in the Netherlands in a micro-context, utilizing a unique dataset consisting of customers from the largest Dutch mortgage lender, Rabobank, that covers 22 percent of the Dutch mortgage market. Where macro studies rely on aggregate consumption data from national accounts and micro studies often use survey-based self-reported data, this study constructs a unique measure of total consumption based on transaction data from bank accounts. The same household is tracked over several years, allowing for a precise identification of those households for which the consumption response of changing house prices is largest through a fixed effects panel. The second major contribution of this paper is the study of regional differences which may arise due to a differential sentiment caused by deviations in house price growth between regions. The only other paper studying the existence of regional effects in the relationship is Campbell and Cocco (2007). They confirm the importance of regional house prices versus national house prices when studying their relationship with consumption in the UK in a pseudo-panel. The last major contribution is the investigation of the importance of mortgage debt levels in the relationship between house prices and consumption. Research by the Dutch Central Bank (DNB Bulletin, 2018) using a macro-panel found that mortgage debt levels of households significantly affect the relationship between house prices and consumption: when there are more homeowners with a mortgage, the relationship is stronger. This is especially relevant for the Netherlands, as it has had the highest mortgage debt-to-GDP and mortgage debt-to-disposable income ratio in the European Union for many years (see section 4.2). The strengthening effect of mortgage debt levels in the relationship between house prices and consumption has positive effects in the sense that increased house prices can lead to a stronger consumption response, but it also makes the Netherlands more vulnerable to the downside of the relationship: decreasing house prices can lead to mortgages being under water, which negatively influences household consumption and bank balance sheet positions, and causes a stronger consumption decline by itself as was the case in the recession of 2012. Therefore, this paper studies the relationship between house prices and consumption in the context of household mortgage debt levels as proxied by the loan-to-value ratio (LTV), accounting for potential heterogeneity between regions, income groups and age groups.
classes, suggesting that Dutch households place their debt in a mental account to which they pay less attention. High income households are shown to respond more strongly in their consumption to house price increases than low income households.
The remainder of this paper is organized as follows: section 2 describes the different consumption theories more elaborately. Section 3 presents an overview of previously conducted empirical research on the relationship between housing wealth and consumption from a macro as well as from a micro perspective. In section 4, the Dutch institutional setting with regard to the housing market is sketched, as well as the hypotheses that are going to be tested. Section 5 describes the data and methodology. The results are stated in section 6 and section 7 concludes.
2 – CONSUMPTION THEORIES
Different theories about individual or household consumption have been developed in the past. These theories can be broadly classified as the life-cycle theory, which is strongly linked to the permanent income hypothesis, the behavioral counterpart of the life-cycle theory, the precautionary savings model and the bequest motive. This section will elaborate on these theories and shortly review some empirical evidence from research done on the extent to which these theories are able to explain reality, focusing on the life-cycle model and the behavioral life-cycle model.
2.1 Life-Cycle Model and Permanent Income Hypothesis
rise to unequal MPC’s out of different sorts of wealth without being refuted by introducing externally imposed liquidity constraints (Hayashi, 1985; Zeldes, 1989). In this setting, people are unable to borrow against certain assets because they are unable to liquefy them due to market imperfections, resulting in a lower MPC out of illiquid assets. When there are liquid assets still available for funding consumption, changes in the value of illiquid assets do influence the MPC out of total wealth, as they are part of this total wealth and the constraint is not yet binding. When all liquid assets have been depleted, the liquidity constraint on illiquid assets becomes binding, and their value stops to affect the MPC out of total wealth since they cannot be used for consumption purposes (Levin, 1998). The assumption that individuals are forward-looking and have one single discount rate has been questioned in the literature. The traditional LC model falls apart if individuals are myopic. In a review of the literature on time discounting and preference, Frederick et al. (2002) show that individuals do not have a single discount rate and that the different discount rates they do have are extremely high. Other research shows that individuals are most likely not forward-looking (see e.g. Poterba, 1998 for a fiscal experiment and McClure et al., 2004 for neurological evidence).
Alternatively, the bequest motive might result in higher savings and lower consumption at a later age than predicted by the traditional LC model. Homeowners with children might save housing wealth gains in order to bequeath them to their children instead of consuming the gains themselves. Skinner (1989) shows in simulations of an overlapping generations model that the bequest motive dampens the traditional LC model response to housing wealth gains. Apart from influencing the savings and consumption patterns of homeowners, transfers from bequest also affect savings and consumption of the recipients. Transfer recipients save less, buy larger homes and buy them sooner than if they had not received the transfer (Engelhardt and Mayer, 1998; Guiso and Jappelli, 2002). Overall, observed household behavior points into the direction that housing wealth is treated differently than other types of wealth, something which is accounted for in the behavioral life-cycle model described below.
2.2 Behavioral Life-Cycle Model
The main deviation of the behavioral life-cycle model (BLC, Shefrin and Thaler, 1988) from the traditional LC model is that it assumes that households view some assets as nonfungible, even when the household is not credit rationed. The BLC model draws from the behavioral concepts ‘mental accounting’ and ‘frame dependence’ (Kahneman and Tversky, 1984). According to the LC model, the marginal propensity to consume (MPC) or save should be the same across different forms of wealth and income in the absence of liquidity constraints. However, according to mental accounting theory, households divide income and wealth over several mental accounts such as current spendable income, current assets and future income, and have a different MPC for each account. As Shefrin and Thaler (1988) state; ‘[…] some
mental accounts, those which are considered “wealth”, are less tempting than those which are considered “income”’ (p. 610). Their BLC model is based on the personal characteristics
willpower effort costs are zero, or when the first-best rule for the planner, where future consumption is completely precommitted to a specific path, is available. However, neither of these conditions is likely to be satisfied, resulting in different outcomes from the BLC model. The behavioral concept ‘mental accounting’ results into different consumption responses to changes in different types of wealth. Households self-impose some sort of non-binding liquidity constraint based on the relative temptation to consume the respective wealth by putting it into a mental account labeled as illiquid. Consumption is not affected by changes in these types of wealth until more tempting, liquid assets have been depleted (Levin, 1998). This is opposite to the result from the traditional LC model as described above; there, liquidity constraints cause value changes in illiquid assets to influence spending before liquid assets have been depleted since illiquid assets are part of total wealth, but not after the liquid assets have been depleted since they cannot be liquefied and used for consumption. Additionally, consumption responses are frame dependent: they depend on the source of the increase in wealth (e.g. rise in income or rise in housing value) as well as on the good to be consumed. An example taken from Levin (1998) explains that a rise in wealth resulting from an increase in housing value will have a different effect on frequently returning small expenditures such as groceries than on infrequent and larger expenditures such as vacations. Graham and Isaac (2002) find that the BLC model is more a suitable description of consumer behavior than the LC model even when the sample consists of members of an economic university faculty.
show that there even exist asymmetric consumption responses to wealth changes which are put in the same mental account. They show that people adjust consumption differently in response to a housing wealth gain than to a housing wealth loss, with the consumption increase being larger after a gain than the decrease after a loss. In a qualitative study on three institutionally different countries (Germany, Hungary and the UK) Toussaint (2011) finds that, even though people appreciate that their dwelling can be sold or its value consumed through mortgage products, they behave as if housing wealth is illiquid, especially in times of financial hardship. According to this research, housing wealth is consumed only as a last resort and is clearly put in a mental account labeled as illiquid.
3 – RELATIONSHIP HOUSING WEALTH AND CONSUMPTION
The relationship between housing wealth and consumption has received quite some attention in the literature, although no consensus about the size and direction of the relationship has yet been reached. A summary of the research reviewed in this article can be found in Table 1.
3.1 Macro Perspective
While they also use retail sales as a proxy for consumption, they calculate the variable by dividing the state general sales tax revenue by the general sales tax rate and make 2 different categories based on the data quality. When using both consumption data categories, they find that a one dollar change in two-year lagged housing wealth changes consumption by about 5 cents, where they motivate the lag by the fact that the cost of realizing housing wealth growth is lumpy, and that homeowners may be less aware of short-run house price changes. By using a method that exploits the stickiness of consumption growth with aid of instrumental variables, Carroll et al. (2011) investigate the immediate and eventual housing and financial wealth effect in the US between 1960 and 2007. They find a significant immediate MPC out of housing wealth of 2 cents per dollar, which increases over time to a MPC of 9 cents per dollar. However, this result is only significant at the 14 percent level.
Moving away from the United States, Ludwig and Slok (2004) investigate the relationship between house prices, stock prices and consumption in a sample of 16 OECD countries between 1960 and 2000. Contrary to the previously described research, they find an insignificant housing wealth effect. However, when they split up the sample into two periods with 1985 as a cutoff point, they surprisingly find a significantly negative housing wealth effect of -0.05 for the earlier time period, and a significantly positive housing wealth effect of 0.04 for the second time period. A potential explanation for this peculiar finding is that wealth effects in general are stronger in the second time period due to financial deregulation. Additionally, they find that stock market wealth effect on consumption is larger than the housing wealth effect, something which is especially true for market-based countries. Another study that finds a larger stock market wealth effect than housing wealth effect is the one on Australia by Dvornak and Kohler (2003). They find a significant MPC out of housing wealth of 3 cents, whereas the MPC out of stock market wealth lies between 6 and 9 cents. However, they state that because Australian households’ housing assets are more than three times larger than their stock market assets, a one percent increase in housing wealth has at least as large an effect on consumption as a one percent increase in stock market wealth.
wealth effect or the substitution effect dominates in the relationship between house prices and consumption. The first effect states that consumption may increase after a rise in house prices through an increase in household wealth or an easing of credit constraints, whereas the latter effect states that, after a rise in house prices, the increased cost of housing services may reduce household consumption. As China faces large regional disparities they try to answer the question based on threshold analysis, using specific values for the health of housing market, as measured by the house price-to-income ratio, and a financial development indicator as thresholds. The authors aggregate daily per capita spending of urban residents to create a macro-measure of consumption and find that the wealth effect is dominant in healthy housing markets with higher financial development where the house price-to-income ratio is below 5.09, whereas the substitution effect is dominant in unhealthy housing markets with house price-to-income ratios between 5.09 and 6.0.
The small and sometimes even negative housing wealth effects observed in the macro literature, especially when found in combination with significantly positive and larger stock market wealth effects, could be an indication of the dominance of the behavioral life-cycle theory. Stock market wealth is generally easier to liquify, and value changes are easier to observe and comprehend by households. The findings could also be a reflection of the dominance of stock markets over other wealth forms in Anglo-Saxon countries. However, macro research may be subject to endogeneity or cover up potentially large inter-household heterogeneity. Therefore, the next section considers literature that analyzed the relationship in a micro perspective.
3.2 Micro Perspective
wealth and consumption. Using aggregate US data from 1950 to 1992, he finds an elasticity coefficient of 0.1. The micro relationship is investigated through the effect of housing wealth on savings and uses PSID data covering the time period 1983-1989. The results show that savings significantly decline after an increase in housing wealth, with the effect being larger for young households. Engelhardt (1996) uses the same PSID data as Skinner (1993) and confirms his results.
Lehnert (2004) explicitly tests the traditional LC model assumption that the MPC is constant over the age of the household using a sample of PSID data covering 1968 until 2003. He finds the strongest consumption effect for the age group 52-62 with an elasticity between 3.36 and 4.67 percent followed by the youngest age group with an elasticity between 3.05 and 3.75. According to the author, potential explanations are that young homeowners are more likely to realize their housing wealth gain as they are more likely to move, or that they act as liquidity constrained consumers who are more likely to spend shocks to their wealth in the face of rising future income. A drawback of the study by Lehnert (2004) is that the main variable of interest, consumption, is measured solely as food expenditure.
Khalifa et al. (2009) use PSID data in a threshold analysis to examine the effect of different income levels on the relationship between housing wealth and non-durable consumption. They find two relevant threshold income levels; $71,000 and $501,000 per year. When income is below $71,000, changes in housing wealth significantly affect consumption with a coefficient of 0.01. In between the two threshold levels, a significant coefficient of 0.02 is found, whereas for income levels higher than the largest threshold, the wealth effect is insignificant. Another study using data from the US is the one by Bostic et al. (2009). However, instead of using PSID data, they use triennial survey data from the Consumer Expenditure Survey and the Survey of Consumer Finance. They use both total consumption and durable consumption as dependent variable. The authors find a housing wealth effect on total consumption of around 0.04 – 0.06 when using market values, and 0.02 – 0.04 when using the market value net of any outstanding housing debt. The effect of housing wealth on durable consumption is largely insignificant.
consumption. The effect of negative housing wealth shocks on consumption is directly measured in Mian et al. (2013). The authors investigate the effect specifically during the Great Recession and incorporate heterogeneity in household debt levels in their results. Measuring total retail consumption based on MasterCard transactions and county-level retail expenditure, they find a MPC out of housing wealth of 0.05 to 0.07, which is much higher for poorer households, highly leveraged households, and those most likely to end up underwater. These findings support the liquidity constraints argument of the traditional life-cycle theory where those who are liquidity constrained show a larger consumption response.
Campbell and Cocco (2007) study the relationship between house prices and non-durable consumption in a pseudo panel setting in the UK, using data from the Family Expenditure Survey (FES) which covers the time period 1988 to 2000. They compare elasticities for different age groups as well as for homeowners versus renters. They find the largest elasticity for older homeowners (1.7 percent) and an insignificant elasticity for younger renters. Additionally, their evidence suggests that UK house prices are correlated with aggregate financial market conditions, since they find that predictable national house price changes significantly affect the consumption of renters as well as homeowners.
In an earlier paper, Attanasio and Weber (1994) also use UK FES data in a cohort analysis to investigate to what extend rising house prices can explain the consumption boom of the late 1980s. They find that rising house prices largely explain the rising consumption of older cohorts, but not that of younger cohorts. Since the younger cohorts were mainly responsible for the consumption boom, they conclude that rising house prices on its own are not a sufficient explanation.
Burrows (2018) examines the relationship between house prices and consumption indirectly in an innovative manner, namely through a bivariate probit model which measures whether UK households are more likely to increase mortgage equity withdrawal or savings. She finds that realized house price gains and expected as well as unexpected house price changes have a significantly positive effect on the probability of withdrawing equity through increasing mortgage borrowing, where the effect is mainly driven by younger households. This, in turn, is found to have a negative impact on household savings, which suggests that increases in house prices do lead to increased consumption.
renters and a negative effect for old renters, they claim that there must be some unobserved common factor which simultaneously affects consumption and house prices, such as income expectations.
On the contrary, when measuring consumption as total consumption expenditure minus housing expenditure, Browning et al. (2013) do not find evidence of an overall housing wealth effect as predicted by the life-cycle theory when studying Danish households between 1987 and 1996. However, when focusing on a credit reform in 1992 which allowed homeowners to use their housing equity to increase their mortgage for non-housing consumption purposes, they do find a significantly positive housing wealth effect, but only for young homeowners. Another study finding conflicting results is the one by Geiger et al. (2016) on German households. They study a broad range of links between German household portfolios, income and total consumption, including the housing wealth effect. Surprisingly, they find a significantly negative housing wealth effect. The authors explain this through the absence of financial products designed for home equity withdrawal, the stringent loan-to-value lending standards, and the low level of home ownership in Germany. Because of these reasons, homeowners cannot easily cash any housing gains, future homeowners need to save more for a deposit, and renters can expect higher rents, resulting in a negative housing wealth effect for the entire population. A negative housing wealth effect on total consumption is also found by Cho (2011) in a study on Korean households, however, only for those households who are in the lowest income quintile. For those households in the highest income quintile, the housing wealth effect is positive.
This overview of the micro literature clearly shows the lack of consensus on the relationship between house prices and consumption, especially when looking at differential effects between age groups. Some studies confirm the prediction of the traditional life-cycle model that younger, liquidity constrained households and/or older, dissaving households show a larger consumption response (Lehnert, 2004; Skinner, 1993; Windsor et al., 2015), whereas others show the weakest or even an insignificant effect for younger households (Attanasio and Weber, 1994; Campbell and Cocco, 2007). The same discrepancy is observed for income groups; some find a significantly positive consumption response for poorer households, in line with the liquidity constraints arguments (Khalifa et al., 2013; Mian et al., 2013), whereas others find a significantly negative consumption response for poorer households (Cho, 2011). This paper will investigate the relationship between house prices and consumption based on different age and income groups as well in an attempt to create more consensus.
4 – THE DUTCH HOUSING MARKET
4.1 Housing Market
When ranking EU28 countries based on homeownership data from 2015, the Netherlands has the 7th lowest homeownership rate. Homeownership rates have risen gradually from 63.9 percent in 2005 to 67.8 percent in 2015, which is slightly below the relatively constant EU28 average of around 70 percent. However, ranking EU28 countries based on homeownership rates using homeowners with an outstanding mortgage or housing loan, the Netherlands differs dramatically from the average situation. Whereas the EU28 average of homeowners with a mortgage was just 26.9 percent in 2015, the Netherlands had the highest ratio that was observed: 60.1 percent (source: Eurostat)4. This could lead to a deviation in behavior of Dutch homeowners in consumption responses to house price changes from previously observed empirical evidence, as they are more indebted than the average European household.
Estimations of Statistics Netherlands (CBS) on the distribution of homeownership rates across the different provinces in the Netherlands show significant differences (see Table 2). For the past twenty years, the lowest homeownership rates were observed in Noord- and Zuid-Holland, whereas the highest ownership rates were found in Drenthe and Zeeland, with the difference amounting up to 20.15 percent in 1998. Incidentally, the province with the lowest
4 Homeownership rate is defined as the fraction of all households types owning a dwelling over the total of all
Table 1 - Literature
authors research question technique area time period coefficient housing wealth -
consumption* Macro perspective
Case et al. (2001) relationship housing wealth, financial wealth and consumer spending
panel model 14 developed countries 1975 - 1996, annually 0.11 - 0.17 US states 1982 - 1999, quarterly 0.05 - 0.09 Benjamin et al. (2004) relationship housing wealth, financial wealth
and consumer spending
panel model with ARMA terms
US aggregate 1952 - 2001, quarterly
0.079 - 0.157 Zhou and Carroll
(2012)
relationship lagged housing wealth growth, lagged financial wealth growth and consumer spending growth
panel model US states 2001 - 2005, semiannually
0.047 - 0.058
Carroll et al. (2011) immediate and eventual MPC out of housing wealth and financial wealth
IV model based on the stickiness of
consumption
US aggregate 1960 - 2007, quarterly
0.018 - 0.087
Ludwig and Slok (2004)
relationship housing wealth, financial wealth and consumer spending
cointegrated panel model 16 OECD countries 1960 - 2000, quarterly insignificant Boone et al. (2001) effect of financial liberalisation on the
consumption-to-disposable income ratio through financial and housing wealth
OLS 6 OECD countries
1975 - 2000, semiannually
-0.06 - 0.34
Dvornak and Kohler (2003)
relationship housing wealth, financial wealth and consumer spending
OLS, IV and OLS panel model
Australia 1984 - 2001, quarterly
0.01 Barrell et al. (2015) relationship housing wealth, financial wealth
and consumer spending
IV model based on the stickiness of
consumption and DOLS
Italy and UK 1972 - 2012, quarterly
Italy: 0.01 UK: 0.03
Dong et al. (2017) relative importance of wealth effect vs. substitution effect in relationship between house prices and consumption
threshold analysis China 2003 - 2014, annually
healthy housing market: 0.09 unhealthy housing market: -0.15a Micro perspective
Skinner (1989) effect of housing wealth changes on consumption
cross-section time series and fixed effect panel model
US 1976 - 1981, annually
Skinner (1993) effect of housing wealth changes on savings quantile regressions US 1984 - 1989, annually
elasticity is positive, larger for younger households
Engelhardt (1996) effect of housing wealth changes on savings mean and median regressions
US 1984 - 1989, annually
MPC mean saver: 0.14 MPC median saver: 0.03
Lehnert (2004) effect of housing wealth changes on consumption in the presence of credit constraints (defined by age)
panel model US 1968 - 2003, annually
average elasticity: 3.94 age quintiles: 3.05, 0, 2.91, 3.36, 2.90b
Khalifa et al. (2013) effect of housing wealth changes on consumption for different income groups
threshold analysis US 2001 - 2005, annually MPC if income >74k: 0.01 74k - 501k: 0.03 >501k: insignificant
Bostic et al. (2009) relationship housing wealth, financial wealth and consumer spending
year-specific and pooled OLS US 1989 - 2001, triennial elasticity based on market value: 0.04 -0.06 net value: 0.02 - 0.04 pooled OLS: 0.05
Mian and Sufi (2011) effect of housing wealth gains on household borrowings and the use of these borrowings
IV panel model US 1997 - 2008, annually
households borrowed 25 cents per dollar increase in housing value, which was used for real outlays such as consumption
Mian et al. (2013) effect of negative housing wealth shocks on consumption in the presence of household debt during the Great Recession
IV panel model US 2006 - 2009, annually
MPC: 0.05 - 0.07, much higher for poorer households, highly leveraged households, and those most likely to end up underwater
Campbell and Cocco (2007)
relationship house prices and consumption pseudo panel UK 1988 - 2000, quarterly
elasticity: 1.22 Attanasio and Weber
(1994)
can the UK consumption boom of the 1980s be explained by an increase in house prices
cohort analysis UK 1974 - 1988, annually
rising house prices largely explain rising consumption of older cohorts, but not that of younger cohorts Burrows (2018) the impact of house prices on mortgage equity
withdrawal and household savings
recursive bivariate probit model
UK 1995 - 2007, annually
Windsor et al. (2015) effect of housing wealth changes on consumption
panel and pseudo panel model
Australia 2003 - 2010, annually
MPC panel: 0.01 - 0.03 MPC pseudo panel: 0.03 - 0.04, highest for youngest homeowners Browning et al. (2013) effect of unanticipated house price changes on
consumption
panel model Denmark 1987 - 1996, annually
insignificant Geiger et al. (2016) the link between the German housing market,
household portfolios, income and consumption maximum likelihood in a six-equation system Germany 1981 - 2012, quarterly -0.07
Cho (2011) effect of housing wealth changes on consumption for different income groups
panel model Korea 1987 - 2008, quarterly
elasticity
low income: -0.07 high income: 0.09c
Gan (2010) effect of lagged housing wealth changes on consumption
panel model Hong Kong 2000 - 2002, quarterly
elasticity: 0.17 * coefficients taken from the main research specification.
a Dong et al. (2017) specify the health of the housing market based on the house price-to-income ratio. In this table, healthy housing market is defined as having a ratio below
5.08, whereas an unhealthy housing market has a ratio between 5.08 and 5.96.
b Lehnert (2004) specifies 5 age quintiles, with the ranges being respectively 25-34, 35-42, 43-51, 52-62, 63-95.
homeownership rate in 2015, Noord-Holland, which is the province where the Dutch capital city is located, also had the highest price index level of all the Dutch provinces in that year.
The development of house prices can be observed by means of the house price index of existing owner-occupied dwellings (Prijsindex Bestaande Koopwoningen, HPI), which is published quarterly by Statistics Netherlands. As opposed to the average purchase price, this index considers the quality of the sold dwelling by controlling for its cadastral value (WOZ waarde). As can be seen in Figure 1, the Dutch housing market has suffered from a four-year bust starting in 2009 and lasting until the first quarter of 2014. Regionally, the bust was most pronounced in Noord-Brabant where the index decreased by 18.6 percent during this four-year period, and was least pronounced in Zeeland, where the decrease in the index was 10.6 percent. Over the period 2001-2017, house prices grew least in Limburg (15.4 percent) and most in Zeeland (58.0 percent). Up to this date, only three out of the twelve provinces have regained their pre-bust index levels. The number of sold dwellings, which already started to decrease in 2007, has returned to its pre-bust levels everywhere. A notable observation from Figure 1 is that in the most recent years, growth in the number of sold dwellings is slowing down, while the house price index and average purchase price continue to show increasing growth rates. According to market expects, this is an indication of an overheated housing market. The development of the house price index and the number of sold dwellings per province can be found in the Appendix (Table A.1).
Table 2 - Homeownership and Price Index per County
4.2 Mortgage Lending
In 2016, the EU28 average mortgage debt-to-GDP ratio was 47.1 percent. With 95.3 percent, the Netherlands had the highest mortgage debt-to-GDP ratio of this region and has had so for many years previously. The same holds for the mortgage debt-to-disposable income ratio which amounted up to a staggering 198.6 percent in the Netherlands, compared to an EU28 average of only 69.9 percent (EMF Hypostat 2017). This fact, in combination with the finding that mortgage debt levels affect the relationship between house prices and consumption (DNB
Bulletin, 2018), makes the study of this relationship the Netherlands especially relevant. Figure
2 shows developments in the Dutch mortgage lending market in relation to the development of house prices. While the bust in the housing market started in 2009, mortgage lending growth rates still remained positive for several years. It was only in 2013 that the amount of outstanding mortgages and the provision of new mortgage loans started to shrink. However, whereas house prices started their recovery in 2014, recovery of mortgage lending, especially mortgage lending by banks, has lagged behind and only begun in the last quarter of 2016. The Dutch Central Bank (DNB) gives several explanations for this lag, the key explanation being the large sum of voluntary repayments which took place since 2013 due to low interest rates and the tax exemption of gifts for home purchases. Other reasons are the increasing share of elderly, asset-rich people in the housing market and the fact that housing equity of many households remains limited and sometimes still is negative (DNB Financial Stability Report Autumn 2017). By 2016, the number of homeowners with negative housing equity amounted up to 20 percent, having decreased from 40 percent in 2013 (DNB Financial Stability Report Spring 2017). Even though recovery of mortgage lending by banks showed a significant lag, the increasing
-40 -30 -20 -10 0 10 20 30 40
Figure 1 - Development Dutch housing market
year-on-year growth rate
House price index Number of sold dwellings Average purchase price
Source: Statistics Netherlands (CBS)
importance of non-bank mortgage providers did ensure that total mortgage lending already started expanding again in 2015.
The high mortgage debt levels are a likely to be a significant factor of most Dutch household portfolios. The ratio of mortgage debt-to-disposable income of households was 198.6 percent in 2016, after having reached an all-time-high of 214.9 percent in 2010 (EMF Hypostat
2017). Apart from the risks posed by these high household debt levels, the most predominant
form of mortgage owned by households also poses threats to household finances and the Dutch financial sector. As Figure 3 shows, the main type of mortgage owned by Dutch households is an interest-only mortgage. During the loan term, only interest payments are made, whereas the principal amount needs to be repaid in full upon maturity. Popularity of this type of mortgage and the generally observed substantial mortgage debt levels are likely to be the result of high maximum loan-to-value (LTV) ratios allowed by institutions and the mortgage interest rate deductibility (MID) rule of 1893. Initially introduced in response to a new tax system to compensate homeowners for the fact that they now had to pay tax on their houses, the rule has evolved into being a stimulant for homeownership and an encouragement for citizens to take on mortgage debt because it allows them to deduct their mortgage interest payments from their gross income before calculating income tax. The rate at which mortgage interest can be deducted is comparatively high in the Netherlands. Groot and Lejour (2017) show that, before
-10 -5 0 5 10 15
Figure 2 - Year-on-year growth house prices and mortgage loans
House price index
Residential mortgage loans
Residential mortgage loans excluding voluntary repayments (available from 2014) New mortgage loans provided by Dutch banks
%
Sources: Statistics Netherlands (CBS), DNB.
the decline in interest rates as a result of the Great Recession, it was not beneficial in the short-run for Dutch households to repay mortgages due to the loss of maximal MID, resulting in increased tax payments that were not outweighed by decreased interest payments.
The prevalence of interest-only mortgages and high mortgage debt levels are likely to have real effects on the Dutch economy. Firstly, research shows that high debt ratios lead to larger consumption increases during economic upswings and larger consumption cuts during or shortly after times of economic downturn. For example, Mian et al. (2013) show that in areas in the US where pre-crisis LTV ratios were higher, consumption declines post-crisis were larger. Similarly, Bunn and Rostom (2015) investigate UK households and find that spending during the financial crisis decreased more for households with higher mortgage-to-income ratios. The largest spending reduction was observed for households with a mortgage-to-income ratio of above 4; these households reduced spending 10 time more than households with mortgage-to-income ratios of below 1. Secondly, interest rate sensitivity of households with mortgages, especially those with variable rate mortgages, is higher as well and has been shown to negatively affect consumption (Auclert, 2017). Thirdly, the negative correlation between interest rates and house prices influences financial stability through financial institutions with mortgage exposure, who experience a higher loss-given-default in the sense that when interest rates are higher, debt service burdens and thus default risks are higher, but recovery values of collateral are lower. Lastly, consumption responses to macro prudential measures may also be
after an interest rate hike than the spending expansion observed after an interest rate cut (Sufi, 2015), because they save a large part of the gain from decreased interest rates in order to minimize harmful consumption effects from increased debt service burdens after potential future interest rate hikes (Di Maggio et al., 2017). The actual effects of high debt levels on the Dutch economy may be less pronounced than suggested by literature due to the high levels of pension wealth accumulated in the Netherlands. Dutch pension funds’ assets amounted up to 180 percent of GDP in 2016, a level which is unprecedented by any other nation in the world (source: OECD Pension Statistics). These high wealth levels may allow for higher debt levels without experiencing any adverse effects. Unfortunately, due to data limitations I am not able to distinguish between different types of mortgages households own. However, the impact of overall mortgage debt levels on the relationship between house prices and consumption is taken into account through LTV ratios.
Research shows that there exists a relationship between mortgage debt and house prices (see amongst others De Haas and De Greef, 2000; Brissimis and Vlassopoulos, 2009; Favara and Imbs, 2015). Damen et al. (2016) find that the ability to pay (ATP) of mortgage owners is a long-run fundamental of house prices and that this relationship is strongly influenced by MID regulations and the characteristics of the country’s mortgage portfolio, indicating that the relationship between mortgage debt and house prices observed in the literature works through borrowers’ ATP. They explain that, ceteris paribus, the ATP of borrowers is higher in a country where mortgage interest is deductible at a higher rate or where interest-only mortgages are dominant than in a country where the MID rate is lower or where annuity-based mortgages are more common. The authors show that a 1 percentage point increase in interest rates would,
ceteris paribus, decrease the ATP of Dutch mortgage owners by 17.4 percent and house prices
by 19.6 percent, which is the strongest effect observed for all countries in their sample5. Moreover, they show that the relatively high MID rate and the large share of interest-only mortgages in the Netherlands imply that a complete abolishment of MID decreases the ATP of Dutch mortgage owners by 35.0 percent and through this, decreases house prices by 39.6 percent. These are very significant effects that should be considered by policymakers when deciding on changing the MID regulations.
5 Damen et al. (2016) use a sample of 8 OECD countries including Belgium, the Netherlands, United Kingdom,
The substantial levels of mortgage debt and the high toll the MID rule took on the Dutch government’s treasury6 led to the decision in 2013 to start phasing out MID. Since then, only borrowers with annuity-based mortgages with maturities less than 30 years are allowed to use MID. Moreover, the maximum rate at which mortgage interest is deductible is being reduced by 0.5 percent every year until a maximum deductibility of 37 percent is reached in 2042. Because the reform is spread over such a long horizon, the negative effects predicted by Damen
et al. (2016) are likely to be counteracted by income and anticipation effects. As can be seen in
Figure 3, the policy reform has not yet resulted in a large decrease in the outstanding amount of interest-only mortgages, however, there has been an increase in the prevalence of annuity-based mortgages due to the reform (DNB FSR Autumn 2017). Another measure taken by the Dutch government to curb growth in the outstanding amount of mortgage debt to households is to reduce maximum loan-to-value (LTV) ratios which could amount up to 125 percent in some cases. A reduction of 1 percent per year was initiated in 2013, until a maximum LTV ratio of 100 percent was reached in 2018. Debates are currently ongoing on whether to reduce the maximum ratio even further to 90 percent. Since literature found that mortgage debt levels affect the relationship between house prices and consumption, these policy adjustments may also affect this relationship.
4.3 Hypotheses
Recent research by the Dutch Central Bank (DNB Bulletin, 2018) found that the mortgage debt of households significantly affects the relationship between house prices and consumption: when there are more homeowners with a mortgage, this relationship is stronger. The presence of favorable tax rules such as the MID described in section 4.2 and mortgage equity release products7 which both promote higher leverage are potential explanations for the role of mortgage debt in the relationship under question, as they free up resources that can be used for consumption. When households understand that an increase in house prices results in a lower LTV ceteris paribus, they might feel richer and therefore adjust their spending upwards. In combination with a relaxation of liquidity constraints and less need for precautionary savings after an increase in wealth, this suggests that the relationship between house prices and consumption is positive. Therefore, the main hypothesis in this paper is as follows:
6 In 2013, the costs of the mortgage interest rate deduction to the state were 13.8 billion euros. Source: Miljoenennota 2018, Rijksoverheid.
Hypothesis 1: Increasing house prices positively affect consumption in the Netherlands.
The relationship is likely to be affected by household LTV ratios, which may be especially relevant in the Netherlands given the high levels of mortgage debt owned by Dutch households (see section 4.2). The effect of the level of LTV ratios in the relationship between house prices and consumption is not clear cut, but depends on the underlying reason for the consumption response. If consumption responses to increased house prices are due to a relaxation of liquidity constraints, high LTV households should show a stronger consumption response as they face a more binding constraint. However, if the consumption response is due to a reduction in precautionary savings, households with low LTVs should show a stronger consumption response. These households are more likely to have desirable levels of precautionary savings and are therefore able to use increases in housing wealth for consumption purposes, whereas high LTV households show a weaker consumption response because they use the housing wealth gains to increase precautionary savings.8 Another reason for the moderating effect of high LTV households in the relationship between house prices and consumption apart from an increase in precautionary savings is that these households might use increases in housing wealth for accelerated deleveraging to reduce upcoming debt service costs instead of for consumption purposes. However, this may be dependent on the phase of the financial cycle and its related sentiment. During periods of negative sentiment, households may be less willing to increase consumption and prefer to reduce debt, whereas during periods of positive sentiments the willingness to reduce debt may be smaller, resulting in larger consumption responses for high LTV households than observed during the periods of negative sentiments. The impact of consumer sentiment is left as a scope for future research. In order to test the impact of LTV ratios and whether the liquidity constraints or precautionary savings argument appears to be the underlying reason for consumption responses to house price changes, the following hypothesis is tested:
Hypothesis 2: The relationship between house prices and consumption is more moderated for high LTV households.
Testing whether households behave according to the behavioral life-cycle model would require in-depth interviews and psychological tests, which is beyond the scope of this research. However, testing whether the relationship between house prices and consumption differs per age group as claimed by the traditional life-cycle model can give an indication as to whether
behavior follows this version of the model. According to the consumption-smoothing assumption underlying this model, younger and older homeowners should show the strongest consumption response to a house price increase. Young homeowners have income levels below their lifetime average and house price gains may reduce their liquidity constraints, allowing them to increase consumption, whereas older homeowners are in a stage of dissaving. They are more likely to have more financial resources and lower debt levels. This, in combination with the generous pension scheme at place in the Netherlands which takes away the need for housing wealth as a precautionary saving, might allow them to spend a larger fraction of their housing wealth gains as well.
Hypothesis 3: The relationship between house prices and consumption is stronger for younger and older homeowners.
Apart from a non-linear effect based on age, research has also shown that income levels significantly affect the relationship (see amongst others Khalifa et al., 2013; Mian et al., 2013; Cho, 2011 as described in section 3.2). There does not exist consensus on whether poorer households exhibit a stronger or a weaker consumption response. The liquidity constrains argument, where an increase in housing value and thus homeowner equity relaxes liquidity constraints, is likely to be the most relevant for those for whom the constraints are binding, i.e., low income households. Therefore, the following hypothesis has been formed:
Hypothesis 4: The relationship between house prices and consumption is stronger for low income households.
interest-only loans are much less attractive for new homeowners. Therefore, even though this is a limitation of this research, it is not likely to be detrimental to the results.
As shown in section 4.1, house price developments in the Netherlands differ regionally. Claeys et al. (2017) show for 6 EU countries including the Netherlands, that house price fluctuations in capital cities are stronger and more volatile than in the rest of the country. Consumption has been shown to differ regionally as well. The differences can be explained by urbanization characteristics (MacMillan et al., 1972), inequality and income levels (Charles and Lundy, 2013) or other demographic, cultural or lifestyle characteristics (Larson, 1998). Research shows that prices and store supply of consumption goods are affected by neighborhood characteristics such as ethnic and socioeconomic conditions (see Moore and Diez Roux, 2006, for store supply and Myers et al., 2011, for prices). To control for the regional consumption differences and the regional house price growth differences described previously, it is tested whether the relationship differs regionally by investigating the effect separately for different province-clusters (details on construction of the clusters are stated in section 6.2 and Appendix Table A.5). To follow up on the research by Claeys et al. (2017) it is tested whether the housing market in large cities creates an upward bias in the estimated relationship. The 4 largest cities in the Netherlands are taken into consideration; Amsterdam (province Noord-Holland), Rotterdam (province Zuid-Noord-Holland), Den Haag (province Zuid-Holland) and Utrecht (province Utrecht). These cities are taken together to reflect the Dutch urban area and to ensure a sufficiently large sample size.
Hypothesis 5: The relationship between house prices and consumption is stronger in urban areas with stronger house price fluctuations.
Finally, the presence of an asymmetric relationship is investigated. Housing market developments in areas showing excessive growth are prone to more press coverage, which may strengthen relationship between house prices and consumption through the confidence channel, where consumption responses are stronger due to an increased feeling of being wealthy. This effect is likely to be stronger during market upswings, which may give rise to an asymmetric relationship over time depending on the direction of the housing market. The sample does not cover periods of nation-wide increasing and decreasing house prices, however, using province-level incidences of falling and strongly rising house prices, an attempt is made to investigate the presence of an asymmetric relationship.
5 – DATA AND EMPIRICAL MODEL
This section described the data collection process, sources, descriptive statistics and data limitations, as well as the construction of the main variable of interest, consumption, and the empirical specification designed to test the hypotheses stated in section 4.3.
5.1 Household data
Household data is obtained from Rabobank Group, an internationally operating Dutch bank which is the largest bank in the Netherlands based on the number of customers. In 2017, Rabobank had 7.3 million Dutch customers and covered 22 percent of the Dutch mortgage market (Rabobank Annual Report 2017). The raw sample covers 41,991 households selected based on having both a mortgage as well as a bank account at Rabobank. Details from the mortgage application are used to determine LTV’s, LTI’s, location of residence and other household-level controls. This is linked to transaction data from the households’ bank accounts, which is used to form the consumption variable. Specified variables are the account’s end of month balance, spending registered as interest payments, monthly returning transactions, which are likely to be expenses such as insurance, non-period savings account transactions, which are sums transferred from the bank account to the savings account in excess of the monthly returning transactions to the savings account, salary received and mortgage interest deduction received. The accounts of all individuals registered on the mortgage application are taken together to represent one household.
Descriptive statistics on the household data can be found in Table 3. For a full list of variable definitions and sources, see Appendix Table A.2. Households are observed for an average period of 21.4 months. The average household in the sample had a disposable 54,146 euros at the time of the mortgage application and started off with an LTI of 3.57. The oldest member of the household is, on average, 54 years old. The average household has an LTV of 0.61 and savings of 28,013 euros during the sample period. According to data from Statistics Netherlands on the financial risks of mortgage debt from homeowners, average disposable income of homeowners between 2013 and 2015 was 42,500 euros, the average LTV was 0.64 and the average LTI 3.7. Based on LTV and LTI ratios, the sample appears to be a near perfect
Table 3 - Summary statistics household data
Variable Mean Median Std. Dev. Min Max Observations
income 54,415.83 48,234.00 38,058.99 17,000.00 2,000,000.00 n = 31,429 N = 672,633 salary 3249.67 2755.00 4401.52 1.00 441,855.00 n = 19,955 N = 376,363 LTV 0.61 0,57 0.34 0.00 8.80 n = 31,139 N = 668,365 LTI 3.57 3,49 1.93 0.01 77.24 n = 31,429 N = 672,632 savings 28,012.88 13,228.76 48,868.39 0.00 1,375,532.00 n = 31,429 N = 672,633 age 54.32 54.00 13.70 20.00 101.00 n = 31,429 N = 672,632
end of month balance 4,973.11 3,262.00 9,558.53 -308,034.60 525,491.40 n = 31,429
N = 672,633 interest payments 593.60 416.00 662.18 0.00 17,138.62 n = 31,429 N = 672,632 monthly returning transactions 1,405.79 965.50 2,345.88 0.00 233,556.00 n = 31,429 N = 672,632 non-periodic savings account transactions 931.81 0.00 3,465.97 0.00 235,000.00 n = 31,429 N = 672,632 mortgage interest deduction 264.91 0.00 416.44 0.00 16,341.86 n = 31,429 N = 672,632 debit card transactions 908.36 717.00 749.81 0.00 16,065.81 n = 31,429 N = 672,632 cash transactions (ATM) 325.69 150.00 366.07 0.00 9,975.18 n = 31,429 N = 672,632 credit card transactions 23.42 0.00 133.31 0.00 4,022.92 n = 31,429 N = 672,632
Note: n is the number of households, N is the total number of observations. The bank account transaction variables
reflection of the Dutch population. However, the sample households do appear to be richer than the average Dutch homeowner. Overall, the sample characteristics are such that the results can be generalized relatively safely for the entire Dutch population during the sample period.
When looking at the sample distribution of income quartiles over LTV quartiles in Table 4, it is observed that households in the lowest income quartile also fall in the lowest LTV quartile, and the richest households, those that fall in the highest income quartile, have the highest LTV ratios. This is perhaps not surprising, as the maximum mortgage households can obtain not only depends on the value of the property (maximum LTV ratios, see section 4.2) but also on the households’ income, but one could argue that high income households may have higher savings which they can use to reduce their required loan. However, the fiscal benefits of the MID regulation are larger for high income-households, stimulating them to take on larger loans. The combination of this and the fact that high income-households are allowed to lend more is an explanation for the observation that the highest income households have the highest LTV ratios.
Table 4 – Distribution income - LTV 1st quartile income 2nd quartile income 3rd quartile income 4th quartile income 1st quartile LTV 9.32 6.65 4.76 4.28 2nd quartile LTV 7.13 6.58 5.91 5.39 3rd quartile LTV 5.54 6.07 6.47 6.92 4th quartile LTV 3.02 5.71 7.86 8.40
Note: Percentage of all observations that fall in the respective class. 1st quartile
LTV/income are households with an LTV ratio/income level in the lowest 25 percent, respectively. 4th quartile LTV/income are households with an LTV ratio/income level in the highest 25 percent, respectively.
5.2 House price data and control variables
House price data and unemployment rates are obtained from Statistics Netherlands, whereas mortgage lending rates are obtained from DNB. House price and unemployment statistics are available only on a quarterly base. To obtain monthly data, it is assumed that the reported statistic applies to the last month of the respective quarter. The first two months of the quarter are linearly imputed using this assumption and the statistic for the preceding quarter. For a description of house price data, see section 4.1. Unfortunately, the time period under consideration does not include periods of nationally decreasing housing prices. When arbitrarily defining months with national growth rates above 0.4 percent (resulting in a yearly growth rate of slightly below 5 percent) as periods of stronger growth, the sample covers 11 months of stronger growth and 21 months of relatively flat prices. 8 provinces showed decreasing housing prices at some point during the sample period. The average unemployment rate in the Netherlands, which is used as a control for business cycle movements, decreased gradually from 7.7 percent in May 2014 to 5.4 percent in December 2016. The average unemployment rate over the sample period is 6.7 percent. The 5 to 10-year collateralized mortgage lending rate, used to control for the potentially changing costs in debt service after a renegotiation of loan terms during the time period, showed a gradual decrease from 3.77 percent in May 2014 to 2.25 percent in December 2016. The average mortgage lending rate over the sample period is 2.89 percent.
5.3 Data limitations
Due to data availability I am not able to control for life events which may impact consumption, such as a divorce or entering unemployment. Data on the composition of the household (e.g. number of children) is also not available, as well as data on the education level of the head of the household. Absence of these control variables may cause a bias in the results, however, the magnitude of the bias is not expected to be such that it undermines the quality of the main results to a large extend, as some of the missing variables are covered by the household fixed effects.
44 percent of this variable has a value equal to zero. It is impossible to establish whether household income was truly zero in these months, or whether it was positive but not picked up by this category because it is labeled differently. For example, profit distribution from self-employed or entrepreneurs may not be labeled as ‘salary’, whereas they did in fact have a positive income. Therefore, when salary equals zero, the variable is treated as missing. Because this greatly reduces the number of households covered by the regressions, total income as registered at the mortgage application is used as an alternative proxy for income as a robustness check.
The main variable of interest, consumption, is constructed based on data on transactions from bank accounts. Certain transaction categories are distinguished, however, it remains impossible to differentiate between durable and non-durable consumption. Due to the nature of durable goods, which are often more expensive and provide consumption services for a longer period, durable consumption may respond differently to wealth changes. Following the logic by Bostic et al. (2009), if households use durable goods for diversification purposes or buy them after unanticipated wealth increases, durable consumption should respond more strongly to house price changes than non-durable consumption, whereas if households view durable goods as a long-term consumption good, the consumption response of durables may be smaller.
Lastly, the time period under consideration doesn’t include a nation-wide downturn in house prices. Therefore, for the estimation of a potential asymmetric effect, I rely on incidences of monthly decreases. As these are less likely to receive press coverage due to their short nature, the confidence argument explained in section 4.3 is less likely to hold. Therefore, the results regarding an asymmetric effect should be interpreted with caution.
5.4 Construction of consumption variable
To measure consumption, two variables are constructed that both rely on transaction data from bank accounts and the accounts’ end of month balance. The bank accounts of all individuals registered on the mortgage application are summed together to represent one household.
The main variable used to measure consumption growth in this paper is based on all transactions done with debit card, cash (measured as ATM transactions) and credit cards:
The subscripts i and t indicate the household and time period, respectively. The difference of the log of the sum is taken to transform the variable into a growth rate and to ensure that the coefficients resulting from the regressions can be interpreted as elasticities. It is assumed that the cash retrieved from ATM transactions is used for consumption purposes rather than for saving. Any interest payments on credit card balances are not taken into account. This measure is likely to be the most reliable measure of total consumption as it directly measures expenses. However, this measure is unable to take into account online transactions which are becoming an increasingly important form of consumption, and it may give rise to large coefficients due to its volatile nature. Because the total value of monthly transactions is relatively small, 1,257 euros on average with a standard deviation of 975 euros, a small increase of for example 125 euros will already give rise to a growth rate of 10 percent. Therefore, a second measure of consumption is developed based on the combined bank accounts’ end of month balance as a robustness check which measures the value of consumption in euros and looks as follows:
𝐶𝑜𝑛𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛𝑖,𝑡= 𝑒𝑜𝑚𝑖,𝑡 − (𝑒𝑜𝑚𝑖,𝑡−1− 𝑖𝑛𝑡𝑒𝑟𝑒𝑠𝑡𝑖,𝑡− 𝑀𝑅𝑇𝑖,𝑡 − 𝑁𝑆𝐴𝑇𝑖,𝑡+ 𝑠𝑎𝑙𝑎𝑟𝑦𝑖,𝑡+ 𝑀𝐼𝐷𝑖,𝑡) (2) where consumptioni,t is consumption measure 2, eomi,t is the end of month balance, interesti,t are loan interest or amortization payments, MRTi,t are monthly returning outflowing transactions that aren’t labeled as any of the specified categories such as insurance costs, NSATi,t are non-periodic savings account transactions from the bank account to the savings account,
salaryi,t is salary received and MIDi,t is mortgage interest rate deduction received. The subscripts i and t indicate the household and time period, respectively. By adjusting the end of month balance as described in Equation 2, changes due to income and expenses that are not consumption are taken into account, although in an imperfect manner, as there are likely to be income and non-consumption expenses that not covered by any of the specified categories. Examples of inflows that are not covered are unemployment benefits, rent and care allowance or child support, and an example of non-consumption expenses is money withdrawn for investment purposes. The construction of the variable assumes that any decreases in the end of month balance after subtracting the end of month balance of the previous month corrected for the specified categories are due to consumption. Therefore, even though this measure can take into account online transactions, it is likely to be less reliable than the first measure of consumption growth. The results of this second consumption measure are reported in the Appendix and shortly described in section 7.4.1.
