RDP 2014-03: Household Saving in Australia 3. Cross-sectional Analysis

In this section we estimate a model of the median saving ratio that takes into account a range of household characteristics. The median saving ratio gives a better indication of how much a ‘typical’ household saves than the mean saving ratio, which can be heavily influenced by a small number of extreme values. The mean saving ratio is nonetheless important since it determines economy-wide household saving, and we return to it in Section 4.

Income is a particularly important determinant of household saving, although there is some debate as to how it effects saving. Economic orthodoxy would suggest that a household's permanent or long-run level of income should not affect their saving ratio, since households with relatively high levels of permanent income would also have relatively high levels of consumption. Aggregate time series data on national saving supports this proposition: as countries grow richer, household incomes trend higher but saving ratios do not. Conversely, the evidence from cross-sectional, household-level studies is less clear; for example, Dynan et al (2004) find that individual households' saving ratios are affected by their level of permanent income.

Our main results are estimated under the assumption that households' permanent income levels do not affect saving ratios, although our results are robust to relaxing this assumption (see Models (1) and (3) in Table B3). In particular, we assume that a household's saving ratio is a function of the deviation of their current level of income from their permanent level of income:

Here y_i is the natural logarithm of household i's current income, is the logarithm of their permanent income, and X_i represents other household characteristics pertinent to the saving decision such as age, employment status and household composition. This model implies that a household will increase their saving ratio if their current level of income rises but their permanent level of income does not, for example due to a one-off bequest. Conversely, a household will reduce their saving ratio if their current level of income falls but their permanent level of income does not, for example due to a temporary spell of unemployment.

In practice we cannot observe the permanent income of a household, and so must estimate it. We do this by regressing current income on proxies for permanent income, including households' education level, occupation, marital status and age, and taking the fitted values as measuring permanent income (see Table B1 for model results). We then use the percentage deviation of current income from the modelled estimate of permanent income as our income variable, (y_i − y_i^*).

Some authors have argued that including a measure of income in models such as ours may introduce measurement error and endogeneity issues, resulting in biased estimates. For example, Sabelhaus and Groen (2000), Brzozowski and Crossley (2011) and Meyer and Sullivan (2011) argue that large dissaving at the bottom of the income distribution in household surveys is more likely to be due to households under-reporting their income than genuine dissaving, although Browning and Lusardi (1996) argue that reporting bias in household income is unlikely to be a serious issue for most households. There is growing recognition, however, that income is too important as a driver of household saving to be excluded – see, for example, Dynan et al (2004) and Muellbauer (2007) – and so we choose to include it in the form discussed above in our main model. As a robustness check we also estimate a model excluding any measure of income – Model (3) in Table B3.

In addition to income, we explore other possible drivers of saving. These drivers and the main variables used in our modelling are outlined below.^[6]

• Life-cycle motives. Although an ageing population cannot explain the rise in the aggregate saving ratio, age is an important determinant of household saving in the cross-section, and so we include it in our median regression model. Age dummy variables are used to capture the saving behaviour of different age groups: young (less than 30 years old), pre-retirement (50–64 years) and old (65 years and over). The reference household is middle-aged (30–49 years).

• Credit constraints. An increase in the incidence of credit constraints would be expected to lift household saving, since some households that may wish to borrow to fund consumption would be precluded from doing so. Credit constrained households are identified from households' answers to questions regarding financial stress; households are assumed to be credit constrained if they answer in the affirmative to at least two out of seven financial stress questions. The reference household is not financially constrained.

  Note that our credit constraint variable will only capture households that are currently credit constrained. In an overlapping generations framework, Kent, Ossolinski and Willard (2007) show that the adjustment to a new equilibrium following a change in credit constraints can take many years to complete. As such, the lowering of credit constraints that occurred in the late 1980s and early 1990s may still have been impacting household behaviour during our sample period. In particular, with a decline in the number of households who purchased housing during the earlier period of elevated credit constraints and relatively low house prices, the share of households likely to experience very large capital gains on selling their homes (and therefore needing to save less than otherwise similar households would) falls.

• Precautionary motives. We seek to capture precautionary motives in a number of ways. Similar to Chamon and Prasad (2010), we construct a variable that seeks to measure a household's risk of unemployment, a risk that is likely to influence a household's saving behaviour. (Chamon and Prasad, in their study of Chinese households, estimate the risk of incurring a large health expense). One might expect that employed households that face a relatively high risk of becoming unemployed in the future save more than other households (see, for example, the models outlined in Zeldes (1989), Deaton (1991), Carroll (1992) and Carroll and Samwick (1997)). Each household's risk of unemployment is estimated using a logit model of the probability of a household containing one or more unemployed people. If a household's fitted probability of unemployment is greater than 10 per cent, the risk of unemployment variable is set equal to 1. (The logit model is based on a number of independent variables including geographical location, wealth, age, migrant status and personal debt status; see Appendix B for more detail).

  Precautionary motives may also be captured in other variables that describe households with less secure incomes or those who are more vulnerable to income shocks, such as migrant households, single-parent households, those with a worse standard of living compared with a year ago and those who rely on government payments for a large share their income. We also control for households likely to be vulnerable to an asset price shock: self-funded retirees and households that draw more than 20 per cent of their income from investments. The reference household is born in an English-speaking country (possibly Australia) and has the same or a better standard of living compared with a year ago.

• Wealth effects. Higher wealth has been found to have a significantly positive effect on household consumption in Australia, and therefore a negative effect on saving, all else equal (Dvornak and Kohler 2003; Yates and Whelan 2009; Windsor, Jääskelä and Finlay 2013). We include the ratio of household wealth relative to income and the gearing ratio (debt relative to assets), as well as home ownership dummies, to capture wealth effects in our model. We interact all of these variables with age because there are theoretical reasons to believe that the saving response to shocks in these variables may differ by age. The reference household is a renter.

Other controls include household size; the number of children in the household (relative to household size); state or territory of usual residence; region of state (rural/urban); education status; skill level of occupation; marital status; gender of the household head; and dummy variables that identify if a household obtains more than 20 per cent of their income from wages and salaries, business income, government payments, and other income.^[7] The reference household is a single male with high school as their highest level of education who lives in urban NSW and works in a high-skilled occupation.

3.1 Regression Output

Table 1 shows results from the median regressions for 2003/04 and 2009/10, where the dependent variable is the saving ratio and the independent variables are as described above. The differences in coefficients across the two time periods are also presented. Full regression outputs are presented in Table B3 where, for robustness, we also show results from a regression where the logarithm of current income is used instead of the deviation of current income from permanent income (Model (1)), and where no measure of income is included (Model (3)).

3.2 Results

Income

As expected we find that the coefficients on the deviation of current income from permanent income are significant and positive, meaning that households whose current level of income is above their permanent level of income save more, all else equal. The value of the coefficient on income suggests that in the cross-section, a 1 percentage point increase in current income relative to permanent income is associated with a 0.6 percentage point increase in the saving ratio, all else equal; this is within, but at the upper end, of the range of estimates presented in Dynan et al (2004) using US data.^[8]

Education, which is often used as a proxy for permanent income, is found to have a significant impact on saving.^[9] This suggests that our estimate of permanent income used to derive the income variable may not be perfect, that the permanent income hypothesis does not hold, and/or that education is capturing other factors. For example, precautionary motives may be lower for highly educated households because they may face less employment risk.

Financial constraints

Households that are financially constrained according to our criterion tend to have higher saving ratios, holding all else equal, although this effect is only statistically significant in 2003/04. As discussed earlier, this accords with intuition.

Variables related to precautionary motives

Single-parent households and those who rely on government payments for a large share of their income tend to save more than other households, all else equal. Being at risk of unemployment is also associated with higher saving, although the effect is not statistically significant.

Households where the household head was not born in an English-speaking country tend to save more than households where the household head was born in an English-speaking country. This is consistent with the results of Islam et al (2013), who find that migrants have a higher propensity to save compared with Australian-born households with similar characteristics. While this effect could reflect the differing priorities of newly arrived migrants compared with existing residents, it could also be evidence of precautionary saving if being born in a non-English-speaking country is associated with less certainty regarding employment.

Variables related to wealth

Turning to the effect of household wealth on saving behaviour, we find that, overall, higher wealth-to-income ratios are associated with lower saving ratios (and therefore more consumption). In general, the wealth effect is smaller for the oldest households, which is consistent with Windsor et al (2013), who interpret this as evidence against a traditional wealth effect on consumption. Rather, they suggest that rising household wealth increases consumption by reducing liquidity constraints, which are more likely to be binding on the young.

Owning a dwelling outright tends to be associated with higher saving for younger households and lower saving for older households. For the young, this effect may be capturing personality traits rather than wealth per se, with those who own their home outright by the age of 30 being inherently diligent savers. For the older age groups, owning a home is likely to be associated with a higher degree of financial security, obviating the need to save in case of emergency.

Turning to the effect of debt on saving behaviour, our results suggest that the more debt a household has relative to their assets, the less the household saves; for households aged under 30 years this effect is statistically significant in 2003/04.

Life cycle

Perhaps unsurprisingly, we find that pre-retirement households save more than middle-aged households (the control group), who in turn save the same or more than the young. Older households, all else equal, tend to save more than middle-aged or younger households would, were they to face similar living circumstances, suggesting that the low level of saving by older households is predominantly due to their circumstances rather than their age per se.

Footnotes

See Table C1 for a full list of definitions of variables used in the modelling. [6]

Other income includes private pensions, superannuation, child support, scholarships, other regular sources and income from family members not living in the household. [7]

Note that if we drop all other controls from our model, the coefficient on income falls to around 0.2, which is more typical of that found in other studies. This highlights the importance of controlling for a range of household characteristics. [8]

Education is widely used as a proxy for permanent income; Attanasio and Weber (2010), for instance, document that more highly educated households tend to have steeper income profiles than those headed by less-educated individuals. [9]