RDP 2019-06: The Effect of Mortgage Debt on Consumer Spending: Evidence from Household-level Data 3. Data

3.1 The Household, Income and Labour Dynamics in Australia (HILDA) Survey

We use household-level panel data from Wave 17 of the HILDA Survey to pin down the causal relationship between debt and spending (DSS and Melbourne Institute 2018). These data allow us to exploit the heterogeneity in debt holdings and spending patterns across and within households over time. The unique scope and depth of the information in the survey also allows us to examine the underlying mechanisms.

The HILDA Survey is an annual Australian survey that has tracked the same representative group of individuals (roughly 17,000 persons from 9,000 households) since 2001. We use data up to 2017. Through personal interviews and self-completed questionnaires, the survey collects detailed information on household economic behaviour, including their spending, income, debt and assets.^[3] The wide range of information collected as well as the panel nature of this dataset makes it well-suited to answering our research questions.

There are a few features of the dataset that are worth highlighting. First, the survey collects information on both non-durables spending (between 2006 and 2017) and durables spending (between 2006 and 2010). Non-durables spending, as defined in this paper, covers frequently purchased items, such as groceries, fuel and utilities, while durables spending covers infrequently purchased items, such as whitegoods, motor vehicles and computers.^[4] While the time series for durables spending is more limited than that for non-durables, it does cover the GFC period. Our model estimates mainly focus on total spending, though in some cases we provide separate model estimates for durables and non-durables spending.

Second, information on owner-occupier housing debt and assets are available each year, while the other components of the household balance sheet (e.g. financial assets, investor housing debt, and non-housing debt) are only available every four years. The model is estimated on an annual basis to capture relatively high-frequency changes in debt and spending. This means that we limit ourselves to studying the effect of owner-occupier housing debt rather than total household debt.

In Australia, owner-occupier housing debt is by far the largest component of total household debt (at more than 60 per cent of the aggregate), so our results for its effect on spending at the household level should also matter for consumption at the aggregate level.^[5]

Third, we use the level of owner-occupier housing debt as our measure of debt whilst controlling for income and housing wealth, as this is most closely tied with the model outlined in Appendix A and allows us to test whether the depth of household balance sheets (the level of both debt and assets) matters. This is in contrast to most of the existing literature, which uses debt relative to income or assets, or mortgage repayments relative to income. In Appendix E we provide results for these alternative debt indicators.

To limit the influence of large outliers on the results and ensure that households with non-positive debt, assets, housing equity or income are not excluded from the sample by default, we follow Dynan (2012) and apply the inverse hyperbolic sine (IHS) transformation to our main variables of interest. This is likely to be a serious problem in household-level studies that use the natural logarithm transformation, as a reasonable share of households have no debt or assets, or negative incomes and housing equity in some years. Also, some households report no durables spending in a given year. The IHS transformation allows us to keep these observations.

We need to place a couple of restrictions on the sample used in our estimations. Importantly, we use the sample of households that held owner-occupier debt in the previous year.^[6] We drop all household-year observations with zero mortgage debt in the previous year for two reasons: first, we want to abstract from any short-term increase in spending due to taking on new debt (e.g. first home buyers furnishing their new home); and second, we observe that some previously indebted households are likely to have misreported having zero debt given that they return to a similar debt level the following year. We also remove some outlier observations since reporting error is an issue, particularly with the spending measures in the HILDA Survey.^[7] Specifically, we remove household-year observations in the top or bottom 1 per cent of income, house prices, and spending growth from the sample.^[8] Table C1 provides some descriptive statistics for the remaining sample as well as for the non-indebted households excluded from the sample.

The data suggest that there are some differences in spending across households with different levels of debt. Figure 2 shows the median level of durables and non-durables spending between 2006 and 2010 for highly indebted households (i.e. those in the top quartile of the debt distribution) compared with less-indebted and non-indebted households. First, highly indebted households tend to have higher levels of spending than other households since they are also more likely to be asset-rich and in the peak spending years of the life cycle (Carroll and Summers 1991; Deaton 1992; Ellis, Lawson and Roberts-Thomson 2003). Second, during the GFC, which had its peak unemployment effect in Australia in 2009, durables spending fell by more for highly indebted households than for other households. This is consistent with durables spending being more discretionary in nature and more easily postponed than non-durables spending.

3.2 The Australian Mortgage Market and Liquidity-constrained Households

In examining the drivers of any debt overhang effect, we use several measures for liquidity constraints. One such measure proposed in the literature is whether households are ‘hand-to-mouth’, that is if they hold little liquid wealth and consume almost all their current income each period. Kaplan, Violante and Weidner (2014) (hereafter, KVW (2014)) provide a framework for identifying such ‘hand-to-mouth’ households – defined as households whose liquid wealth is less than half their income each pay period.

In the KVW (2014) framework, all housing wealth is illiquid. However, this is not the case in Australia due to several important features of the mortgage market discussed in Appendix B. These features increase the ability of Australian households to prepay their mortgages (at near zero cost) and thereby build up prepayment buffers which are essentially liquid wealth. This implies that prepayment buffers should be taken into account when measuring liquid wealth.

A unique feature of the HILDA Survey is the availability of data on the home purchase history of each home owner. Every four years, households are asked how much their home cost when they originally bought it and how much they borrowed at the time. By combining this information with a standard bank (credit foncier) formula that links loan repayments to the interest rate, loan term and loan amount, we can estimate the scheduled mortgage balance of each household:

where the scheduled mortgage debt balance (D_S) is a function of the stock of debt at origination (D₀), the nominal mortgage interest rate (i), the age of the loan in years (k), and the term of the mortgage in years (T).^[9] The scheduled balance is the total amount that the borrower is contracted to repay at any given time based on this formula. But given the capacity to prepay, most borrowers have an actual mortgage balance that is lower than the scheduled balance. The difference between the actual and scheduled balance is an estimate of each household's prepayment buffer.

Based on this, we estimate that around 14 per cent of Australian households are liquidity constrained (Figure 3). This is slightly lower than the estimate based on the KVW (2014) definition that does not adjust for prepayments. However, the adjustment only affects home owners with mortgage debt. So the differences are much more pronounced when focusing solely on indebted households. On average, around 5 per cent of indebted homeowners are liquidity constrained using the new measure compared with 13 per cent using the KVW measure.

Footnotes

Expenditure items are collected through the self-completed questionnaire, which has a lower response rate than the interview (85 to 95 per cent response rate, depending on the wave). [3]

Non-durable items collected are: groceries; meals out; leisure activities; child care; alcohol; cigarettes and tobacco; public transport and taxis; clothing and footwear; motor vehicle fuel, maintenance and repairs; health care (e.g. medical fees, private health insurance); telephone and internet charges; home maintenance and repairs; and education. Durable items collected are: holidays; motor vehicles; computers and related services; audio visual equipment; household appliances; and furniture. [4]

We have run the analysis on the limited sample of years for which total household debt is available, and find that the results hold for owner-occupier housing debt but not investor housing debt. This may reflect the small sample of households with investor housing debt. Because of the importance of owner-occupier debt in total housing debt, its effect on aggregate consumption is likely to dominate the effects of other types of debt. [5]

Similar results are found if we extend the sample to include previously indebted households. [6]

The value of some expenditure items reported in the HILDA Survey, such as consumer durables, are estimated to differ by as much as 10 per cent from the more accurate estimates produced by the cross-sectional Household Expenditure Survey from the Australian Bureau of Statistics. This is mostly due to differences in data collection procedures (Wilkins and Sun 2010). [7]

The top and bottom 1 per cent of total spending growth are removed from the sample for the models of total and durables spending, while the top and bottom 1 per cent of non-durables spending growth are removed from the sample for non-durables spending models since it covers a longer time period where total spending is unavailable. [8]

The HILDA Survey provides an estimate of debt at origination and whether the mortgage has a variable or fixed rate, or a combination of both. However, we do not have information on the contract term or the interest rate on the mortgage. We assume a standard 25-year mortgage for the loan term and we assume the mortgage interest rate is equal to a standard mortgage indicator rate based on the year in which the loan was taken out. This standard interest rate is adjusted for average discounts and any reported refinancing by the household. We apply the formula separately for variable-rate and fixed-rate mortgages. To the extent that some households have mortgages with longer maturity, their scheduled mortgage balance at any point in time might be higher than our estimate suggests. As a consequence, prepayment buffers for these households may be larger than estimated. [9]