Keep calm and consume? Subjective uncertainty and precautionary savings

Benhabib and Bisin (2005) develop a neuroeconomic life-cycle model based on the assumption that a consumer can decide to override his/her automatic consumption choices with a cognitively controlled choice. This is related to the concept of mastery or locus of control (that is the belief that one can oneself affect one’s future economic outcomes) which has been shown empirically to affect wealth accumulation (Cobb-Clark et al. 2013). Exercising such self-restraint, however, is costly.

This should be particularly important for low-income households that struggle to build up sufficient savings; those households often rely on increased unsecured debt in the event of unemployment (see Sullivan 2008). Low-income household’s reliance on (typically expensive) unsecured debt could have long-term, undesirable distributional effects.

The level of wealth at any given point in time is the result not only of a series of behavioural choices, but also of a history of income shocks, starting conditions, and luck in terms of investment returns.

For detailed information on the HILDA Study, its survey design and collected information see Summerfeld et al. (2015).

The data was extracted using PanelWhiz, software that provides a graphical interface for assessing many panel data sets from around the world in Stata. Haisken-DeNew and Hahn (2010) provide a general introduction to the use of PanelWhiz, and Hahn and Haisken-DeNew (2013) discuss the software’s usage specifically for Australian datasets including HILDA.

This leaves a sample of one-family households, where this family is a couple-family with or without children, a lone-parent-family or a lone person without children.

We also estimated a version of equation (1) with added household fixed effects to control unobserved heterogeneity. The results do not change. However, panel attrition reduces the sample available for the fixed effects estimation. We thus prefer the cross-sectional estimates and present them in this paper; the panel estimates are available on request.

This imputation procedure is, although drawing on two different data sets, conceptually similar to a two-stage instrument variable estimation, and the prediction of expectations can be interpreted to represent the “first-stage” estimation results, while the main estimation of equation (1) represents the “second stage”. In that sense, our analysis follows Lusardi (1997) and Mastrogiacomo and Alessie (2014) who use instruments to deal with attenuation bias in measures of subjective uncertainty. The ‘first stage’-estimation includes controls for state of residence, age, gender, occupation, education, household size, income and year; these characteristics are also controlled in the “second stage” - the main estimation of equation (1) (X_itand ly_itT), with the exception of gender. In addition, the probability of losing one’s job, leaving it voluntarily and finding a new one at least as good, as well as one’s self-rated financial situation compared to last year’s, are included in the first stage, but excluded from the second stage. For this strategy to be valid, these “instruments” must, first, be strongly correlated with economic expectations, and second, the household’s subjective assessment of the future state of the economy must be the only link between the “instruments” and the household’s savings rate. The F-statistic for a test on joint significance of these variables in the first stage is 68.10 (see Table 1), indicating that the first criterion is indeed fulfilled. In order to fulfil the second condition, a number of other crucial variables have to be controlled in the second stage: in particular variables that describe the households’ past and future real financial situation. As already discussed in Section 2, the estimation equation (1) will include current assets held in accounts and ac_it, other wealthln(wealth_it), future income ly_itT, and the variance of future income lvarly_itTto ensure that these conditions hold.

Income is transformed into 2014 Australian dollars using the Consumer Price Indices provided by the Australian Bureau of Statistics.

Where income information is not available in one or two of the four waves, it is assumed to be equal to the average household income as calculated from the waves for which the information was available. If income was missing in more than two of the four waves, observations have been removed from the sample. The within-household variance of income is calculated using true income observations only.

We have run robustness checks that additionally include the number of dependent children. Results do not change and are available from the authors on request.

This includes individually as well as jointly owned cheque accounts, savings accounts, keycard/EFTPOS accounts, other transaction accounts, fixed term deposits and cash management trusts.

All asset components have been transformed into 2014 Australian dollars using the Consumer Price Indices provided by the Australian Bureau of Statistics.

A number of items are missing in less than 1% of all cases; these are rare items for which most individuals indicate to not have this asset type or debt category at all: cash investments, and the value of and debt on investment properties. For somewhat more commonly held assets and debt categories, exact values are missing in more than 1% and less than 5% of all cases; these are the value of and debt on an owner-occupied home, the value of businesses, collectibles and vehicles, as well as the financial assets “equity investments”, “life insurance” and “trust funds”. The most commonly held types of assets and debt, which are missing in more than 5% of all cases, are positive and negative balances on bank accounts and credit card accounts, other forms of personal debt (including unpaid household bills), and pension savings. Among these categories, values are missing for 7.9% (other debt) to 18.3% (pension savings) of all observations.

The imputation is based on nearest-neighbour regression (separately for zero-values and non-zero-values), improved by imputation according to the Little and Su (1989) longitudinal imputation method where this was possible. Summerfeld et al. (2015) and Hayes and Watson (2009) provide details.

Evaluated at all other variables set to zero, this would imply a cash savings rate of 5.7%, which is substantially higher than the empirical observed value of 1.5%. However, a household with all controlled characteristics at zero does not exist in the data.

To explore which of the two explanations is likely at play, we repeated the estimation adding a dummy-variable that indicates whether a household is currently paying off a mortgage. In that specification, the coefficient on non-financial assets is reduced by about 50% and becomes insignificant at the 10%-level; at the same time, households that currently pay off a mortgage are observed to accumulate on average 1.26 percentage points more cash relative to their income than other households. However, the underlying estimated coefficient is insignificant at the 10%-level.

We also estimate a fully interacted model, in which not only subjective and objective economic uncertainty, but also all control variables are interacted with the set of year dummies. The results on key coefficients (subj_it and lvarly_itT) are robust to this change and available from the authors.

While this is not a measure of intelligence per se, the test’s predictive power for general intelligence tests is very high. The main advantage of this test compared to more direct measures of intelligence is that it is quick to administer in a survey setting.