Demand for emergency savings is higher for low-income households, but so is the cost of shocks

Precautionary behavior (Zeldes 1989; Deaton 1991) is central to many models of savings behavior and has been used to explain departures from the simplest life-cycle and permanent income hypothesis models of savings (Hubbard et al. 1995; Carroll 1997; Attanasio et al. 1999; Gourinchas and Parker 2002; Ludvigson and Michaelides 2001; Attanasio and Weber 2010). It is also frequently used as a benchmark for generating inequality in macroeconomic models (Aiyagari 1994; Krusell and Smith 1998) and to understand fluctuations in income and consumption in aggregate (Carroll et al. 1992; Parker and Preston 2005) and so precaution has been central to some of the most important questions of macroeconomics (Gourinchas and Parker 2001) and microeconomics to understanding household savings behavior. It has been used to explain why East German household saved different amounts the West German households following reunification (Fuchs-Schündeln 2008), the investment and savings decisions in Thailand (Kaboski and Townsend 2011), and the effects of banks in India (Fulford 2013).

See, for example, the campaign American Saves by the Consumer Federation of America, https://​americasaves.​org/​for-savers/​set-a-goal-what-to-save-for/​save-for-emergencies, accessed February 28, 2017. Building an emergency fund is often the first advice of financial “gurus” (see Footnote 3).

Suze Orman advocates having 6–8 months of income in liquid form (“Emergency Cash Is a Necessity, Not a Luxury” Yahoo! FinanceSuze Orman Money Matters, 2006, available at: http://​finance.​yahoo.​com/​expert/​article/​moneymatters/​7599, accessed December 21, 2010). Dave Ramsey suggests building a $1000 emergency fund as step one, paying off debt as step 2, and then building liquid savings of 3–6 months worth of expenses (“The Seven Baby Steps: Begin your journey to financial peace” Available http://​www.​daveramsey.​com/​new/​baby-steps/​, accessed December 21, 2010). David Bach recommends having savings worth at least 3 months of expenses (The Automatic Millionaire : A Powerful One-Step Plan to Live and Finish Rich, 2004, New York: Broadway Books).

Surveys that ask about the monetary rather than relative value of emergency savings include the Survey of Consumer Finances used in this paper, the Italian Survey of Household Income and Wealth used by Jappelli et al. (2008), and private surveys whose results are widely reported. For example, in reporting on a Bankrate.com survey “63% Of Americans Don’t Have Enough Savings To Cover A $500 Emergency” by Maggie McGrath (Forbes, January 6, 2016, available at:https://​www.​forbes.​com/​sites/​maggiemcgrath/​2016/​01/​06/​63-of-americans-dont-have-enough-savings-to-cover-a-500-emergency/​, accessed February 28, 2017).

Recent work that can look at expenditure volatility (measured as the percent change from month to month) within individuals suggests there are few differences across income groups (Farrell and Greig 2017). Even if the risk is the same, the utility costs of the same risks may be higher for lower income households, making them more precautionary.

See Rubin (1987) and Schafer (1997) for an explanation of multiple imputation, and Kennickell (1998) for an explanation of the multiple imputation procedure in the SCF.

The emergency savings target mean months of normal income is statistically different between the lowest two quintiles and the highest quintiles. Taking into account the multiple imputation, with 95% confidence the first quintile is significantly larger than each of the other quintiles. Similarly, the second quintile is statistically larger than the third, fourth, and fifth quintiles.

Since the semi-parametric approach does not easily admit multiple imputation or survey weights, I do not present the estimates of \(\zeta \) separately, but am instead concerned with the general shape of f. I use the same shifters in the parametric analysis later, where the coefficients are reported and the standard errors are corrected for multiple imputation.

Each of the densities are scaled to be proportional to the relative size of each group in the population, but the densities themselves are calculated from the unweighted survey to show where the nonparametric estimates are most meaningful. The densities are useful for understanding where the semi-parametric estimator has power; if the density is very low, then the slope is being estimated off of few observations and so is unreliable.

The strong negative slope for the low-income households and the zero coefficient on normal income for high-income households does not depend on the specific dividing line. The results of estimating the same regression as in the third column, but allowing the dividing line to vary are shown in Fig. A-1 in Appendix ESM. The figure shows the coefficient for the high-income, the coefficient for the low-income, and the \(R^{2}\) from the regression. The coefficient on lognormal income for high-income households is statistically indistinguishable from zero when the definition of high-income varies from around $45,000 to above $90,000, and the coefficient estimate is generally very close to zero, especially between $50,000 and $65,000, the interval of the best fit. Similarly, the negative slope with normal income for the low-income is large and significant throughout the entire range of dividing lines. The divide that provides the best fit by the average of \(R^{2}\) across the imputations is $57,526, and so I present the parametric results using this divide.

There appears to be a strong association between having zero or close to zero housing debt and high precaution, since those who pay < 5% of their income to housing also have a higher precaution. The functional form that the housing costs enter in the estimates is somewhat arbitrary, but the particulars of the form do not seem to matter. Those who owe little or nothing each month hold stronger precautionary preferences, but after zero housing cost the effect appears approximately linear since additional polynomial terms of the fraction spent on housing are neither large nor statistically significant.

The data used are the 2014 public use micro-data interview survey, available at http://​www.​bls.​gov/​cex/​pumdhome.​htm, accessed December 20, 2015.