Disentangling the effect of household debt on consumption

Abstract

We estimate the relationship between household mortgage debt and consumption for the period 2006 to 2015. Using Dutch administrative data, we find that the average consumption of households with high mortgage debt prior to the financial crisis has decreased much more during the crisis than that of other households. We also find that the willingness or ability among households to use new mortgage debt to finance one-off high consumption decreased during the crisis. On the macro-level, the drop in Dutch consumption during the financial crisis is predominantly driven by households who were already in high debt. They are responsible for a 7 percentage points drop in macro-consumption at the worst point of the financial crisis. Within this group, the drop in consumption of households with negative home equity explains 6 percentage point of the total drop in macro-consumption.

Appendices

Data cleaning procedures

In this appendix, we describe the handling of the data more extensively. First, we describe how we clean income and wealth data. Second, we describe how we correct for the price effect, when using the change of wealth data as a proxy for household savings. Finally, we discuss the cleaning procedure of the administrative consumption proxy.

1.1 Cleaning income and wealth data

To start we have to make the timing of wealth and income data consistent with each other. Income is measured by the tax authority on the last day of the year, while the stock of wealth is measured at the first day of the year. To make the timing consistent, we set the wealth data back by one day, effectively lagging wealth by one year.

Next, we clean the data. We drop all observations for which we do not observe disposable income. Furthermore, we delete all observations for which net household income is below a threshold value: $75\%$ of the yearly social welfare level for a single household in 2009 (5760 euro).

Wealth data consists of ten different categories.¹³ In all wealth categories, we set missing values to zero. Furthermore, depending on the wealth category, we set negative values to zero if it is clear that this category cannot have negative values. In four wealth categories (i.e., value of the house, mortgage, value of other properties and assets of privately owned firms), we use linear interpolation if observations are missing, since the values of these categories are very stable over time and we can thus infer missing observations from surrounding years. For more information on cleaning these wealth categories, we refer to Bijlsma and Mocking (2017). Lastly, we drop any observation for which the value of the house is below 50 thousand and above 2.5 million euros.

Finally, we want to remark that the distribution of the savings account data changes over time due to improved measurement of the small amounts (below 500 euro).¹⁴ We cannot clean for this. Instead we choose to ignore this problem, because it manly concerns small values and the resulting measurement error is most likely captured by the time FE.¹⁵

1.2 Accounting for the price effect in wealth data

Because we use the change in wealth as a proxy for household savings, we need to correct for the price effect in the change of wealth, as explained in Sect. 3.3. This appendix describes how we correct for the price effect in each wealth category.¹⁶

We assume that the price effect on savings accounts, assets of privately owned firms and other assets can be neglected. For the first asset type, this assumption is not too problematic, since the interest rate in our sample period is close to zero. For the second there is most likely a strong price effect. However, it is not easy to correct for this and the number of households that privately own firms are small.¹⁷ The third type consists of a wide range of assets, both liquid (e.g., cash) and illiquid (e.g., art), and therefore, correction for the price effect is not possible.¹⁸

For stocks and bonds, we correct for the price effect by using the national mutation in stocks and bonds due to financial transactions (savings) and due to changing prices (the price effect) from the national account data of the CBS. These series allow us to calculate the share of the mutation due to changing prices with respect to the total mutation in each year. Under the assumption that for each household the change in prices is equal to that on the national level, we can calculate which share of the change in their stocks and bonds portfolio is due to investment. This approach is more accurate than calculating the excess return of a household’s portfolio compared to a benchmark portfolio based on the Amsterdam Exchange (AEX) index. The latter method leads to substantial deviations compared to macro-level mutations in stocks, especially during the financial crisis.¹⁹

We only consider a change in housing wealth when a household moves to a new address, because only in this case the change in housing wealth is due to investment in housing, i.e., savings, instead of a price effect. In all other cases we assume it is a price effect. However, we do not know the exact selling price of the old house and the purchasing price of the new house for each household that moves. Furthermore, we only observe the value of the house at the end of each year. However, we do observe the month in which a household changes its address, so we impute the value of the old and new house to this specific month using the housing price index of the CBS and assume that this is the actual selling and purchasing price of the old and new house, respectively.

We do not observe what other real estate is comprised of. This category consists of second houses, but also holiday homes. We correct for the price effect by assuming that a household only (dis)invests in real estate if the year-on-year growth rate is (lower) higher than 15%. All changes below this threshold we classify as a price effect.²⁰ All smaller changes are assumed to be caused by the price effect. The choice of this benchmark is based on the Dutch national housing price index, year-on-year it only varies between $5\%$ and $-7\%$. Because the growth in real estate prices varies widely between regions, we let the benchmark be substantially higher than the maximum absolute change in the price index.

1.2.1 Cleaning the administrative consumption proxy

After constructing the administrative consumption proxy (see Sect. 3.3), we clean the new proxy. First, we drop all observations for which we do not observe any spending. On a household level, we drop all households if they have an aggregate consumption that is negative, if their average consumption over time is lower than $75\%$ of the social welfare level in 2009 (5760 euro) or if their average consumption is 120 thousand euro larger than their disposable income. Lastly, we winsorize consumption at the bottom, at 5760 euro, and at the top, at one million. The latter threshold is chosen because disposable income is also winsorized at one million by the CBS.

Fig. 11

Histogram of DTI and real financial wealth. In the DTI histogram, the dashed red and blue line indicate the top quantile and decile, respectively. For the purpose of display we exclude all observations for which DTI equals zero. Remark that our data are constructed such that the DTI ratios are capped at 10 (see Sect. 3.2). We scale real financial wealth by household’s real disposable income. The dashed blue and red line indicate the bottom decile and quantile, respectively. For the purpose of display, we exclude all observations for which financial wealth is equal to zero

Next, we drop all observations in which a major change in the household type occurs: from single to a couple or vice versa. We do this because in the year of this event household income and wealth change substantially since wealth is split between or aggregated over the two individuals in the household.

Finally, we construct our dependent variable $c_{i,t}/y_{i,t-1}$. To prevent outliers, most likely caused by measurement errors, from influencing our results, we winsorize our dependent variable at the top at 2.5 and at the bottom at 0.2. This is a conservative approach winsorizing on both sides a little less than 5% of the observations.

Additional descriptive statistics

We display two additional histograms of DTI and real financial wealth scaled by household’s real disposable income, where we define financial wealth as in Sect. 4: the sum of the savings accounts, stocks, bonds and other assets. Figure 11a displays the position of the top quantile and decile, which we use in our regression analysis to define $\hbox {DTI}_{i,t}^\textrm{high}$ (see Sect. 4). Figure 11b displays the bottom decile and quantile, which we use in Sect. 6.1 for the sample split between households with low and high liquidity, respectively.