The wealthy hand-to-mouth.

Kaplan, Greg ; Violante, Giovanni L. ; Weidner, Justin 等

ABSTRACT The "wealthy hand-to-mouth" are households that hold little or no liquid wealth, whether in cash or in checking or savings accounts, despite owning sizable amounts of illiquid assets (assets that carry a transaction cost, such as housing or retirement accounts). We use survey data on household portfolios for the United States, Canada, Australia, the United Kingdom, Germany, France, Italy, and Spain to document the share of such households across countries, their demographic characteristics, the composition of their balance sheets, and the persistence of hand-to-mouth status over their life cycle. The portfolio configuration of the wealthy hand-to-mouth suggests that these households may have a high marginal propensity to consume out of transitory income changes, a prediction for which we find empirical support in PSID data. We explain the implications of this group of consumers for macroeconomic modeling and fiscal policy analysis.

**********

A valuable framework for analyzing both household survey and aggregate time-series data on the joint dynamics of income and consumption is the life-cycle permanent-income hypothesis. Nevertheless, economists have long recognized that certain aspects of these data are at odds with some of this theory's most salient predictions. This is true for both the standard version of the theory (Friedman 1957; Hall 1978) and the more recent "buffer-stock" versions (Deaton 1991; Carroll 1997). At both micro and macro levels, it is common to estimate a large sensitivity of consumption to transitory changes in income, whereas according to the theory these income dynamics should be smoothed. (1) Moreover, expected consumption growth often fails to correlate with the real interest rate, a result that implies a breakdown of the forward-looking Euler equation holding with equality, as long as the elasticity of intertemporal substitution is not zero. (2)

The most direct way to account for these facts is through the existence of a sizable share of hand-to-mouth (HtM) consumers in the population, that is, consumers who spend all of their available resources in every pay period. HtM consumers have a high marginal propensity to consume out of transitory income changes, which could account for the high correlation between consumption and the transitory component of income growth, even for anticipated income shocks. Moreover, the Euler equation does not hold with equality for HtM consumers, and thus they are a source of misalignment between movements in the interest rate and movements in aggregate consumption growth. The main challenge to this view is the claim that micro data on household balance sheets suggest that the fraction of households with near-zero net worth, and hence those who consume all their income each period, is too small for the model to quantitatively reproduce the facts discussed above.

Measuring HtM behavior using data on net worth is consistent with the vast majority of equilibrium macroeconomic models with heterogeneous agents. These models feature either a single asset or two assets with different risk profiles (but the same degree of liquidity). Notable examples are the Bewley models, which feature uninsurable idiosyncratic risk and credit constraints, in the tradition of Mark Huggett (1996), S. Rao Aiyagari (1994), Jose-Victor Rios-Rull (1995), and Per Krusell and Anthony Smith (1998), and the spender-saver models, which feature impatient and patient consumers with complete markets, in the tradition of John Campbell and N. Gregory Mankiw (1989). Spender-saver models have been revived recently to analyze macroeconomic dynamics around the Great Recession by Jordi Gali, David Lopez-Salido, and Javier Valles (2007), Gauti Eggertsson and Paul Krugman (2012), and Alejandro Justiniano, Giorgio Primiceri, and Andrea Tambalotti (2013), among others. Models by Krusell and Smith (1997) and Christopher Carroll, Jiri Slacalek, and Kiichi Tokuoka (2014a, 2014b) combine the spender-saver insight of heterogeneity in patience with a standard one-asset incomplete-markets model.

In this paper, we argue that measurements of HtM behavior inspired by the spender-saver class of models are misleading, because they miss what we call the wealthy hand-to-mouth (wealthy HtM) households. These are households that hold sizable amounts of wealth in illiquid assets, such as housing or retirement accounts but have very little or no liquid wealth, and as a result consume all of their disposable income every period. Clearly, such households would not be picked up by standard measurements since they have positive--and often substantial--net worth.

To obtain a comprehensive measurement of HtM behavior with cross-sectional survey data about household portfolios, a far better strategy is to use a model with two assets, one liquid and one illiquid, as the guiding framework. The illiquid asset yields a higher return, but it can only be accessed by paying a transaction cost. Recent analyses using this two-asset model have been carried out by George-Marios Angeletos and others (2001), David Laibson, Andrea Repetto, and Jeremy Tobacman (2003), Raj Chetty and Adam Szeidl (2007), Fernando Alvarez, Luigi Guiso, and Francesco Lippi (2012), Jonathan Huntley and Valentina Michelangeli (2014), and Greg Kaplan and Giovanni Violante (2014a, 2014b).

Viewed through the lens of this two-asset model, one discerns two types of HtM households: The poor hand-to-mouth (poor HtM), those who hold little or no liquid wealth and no illiquid wealth; and the wealthy HtM, who also hold little or no liquid wealth but have significant amounts of illiquid assets on their balance sheets. Just like the poor HtM households, wealthy HtM households have a large marginal propensity to consume out of small transitory income fluctuations. However, in this analysis we show that wealthy HtM households are more similar to non-HtM households along many other important dimensions. As a result, the wealthy HtM cannot be fully assimilated into either group. Rather, they are best represented as a third, separate class of households.

This paper investigates wealthy HtM behavior both theoretically and empirically and examines this peculiar but sizable group's implications for macroeconomic modeling and policy analysis.

First, we ask why households with significant wealth would optimally choose to consume all of their income every period, instead of using their wealth to smooth shocks. To answer this question, in section I we develop a stylized model based on Kaplan and Violante (2014a). The model reveals that, under certain parameter configurations, optimal portfolio composition has positive amounts of illiquid wealth and zero liquid wealth. Such wealthy HtM households are better off bearing the welfare loss from income fluctuations rather than smoothing their consumption. This is because the latter option requires holding large balances of cash and foregoing the high return on the illiquid asset (and, therefore, the associated higher level of long-run consumption). This explanation is consistent with calculations by Martin Browning and Thomas Crossley (2001), who show that, in a plausibly parameterized life-cycle buffer stock model, the utility loss from setting consumption equal to income, instead of fully optimizing, is second order. John Cochrane (1989) and Krusell and Smith (1996) perform similar calculations in a representative agent environment. Our model also provides useful guidance for our empirical strategy. In section II we outline this strategy in detail and explain how we approach measurement issues.

Next, we ask how large the share of wealthy HtM households is in the total population, what these households' demographic characteristics are relative to the other two groups, how their balance sheets compare with those of the non-HtM households, and how persistent their HtM status is over their life cycle. This empirical analysis is based on cross-sectional survey data on household portfolios for eight countries: the United States, Canada, Australia, the United Kingdom, Germany, France, Italy and Spain. We describe these data in section III. In the existing literature examining these data on household portfolios, the emphasis has been on the allocation between risky and safe assets (see Luigi Guiso, Michael Haliassos, and Tullio Jappelli [2002] for a thorough cross-country comparison). Instead, our focus is on the liquidity characteristics of the portfolio. In section IV, we study U.S. data, for which we have several repeated cross-sections between 1989 and 2010, as well as a two-year panel for 2007-09. In section V, we present a comparative cross-country analysis with survey data from 2010 and surrounding years.

The analysis of U.S. data leads to six main findings. First, we find that between 25 and 40 percent of U.S. households are HtM, with our preferred estimate being one-third of the population. Second, we find that one-third of HtM households are poor and two-thirds are wealthy; therefore, the vast majority of this HtM group, being wealthy HtMs, would be missed by measurements of HtM behavior that are based on net worth. Third, households appear to be most frequently poor HtM at young ages, whereas the age profile of the wealthy HtM is hump-shaped and peaks around age 40. Fourth, the wealthy HtM typically hold sizable amounts of illiquid wealth: for example, the median at age 40 is around $50,000. Fifth, wealthy HtM households appear very similar to the unconstrained non-HtM in the age profiles of their income and their shares of illiquid wealth held in housing and retirement accounts. Finally, we find that wealthy HtM status is slightly more transient than poor HtM status.

Some interesting findings also emerge from a comparison of the U.S. economy with the other seven countries we study. In all the other countries, wealthy HtM households are a much greater share of the population than poor HtM households, even more so than in the United States. However, the total fraction of HtM households varies significantly across countries. As in the United States, HtM households represent more than 30 percent of the population in Canada, the United Kingdom, and Germany, but they represent 20 percent or less of the population in Australia, France, Italy, and Spain. For the euro area countries, we observe that holdings of consumer debt are minimal, suggesting that the substantial liquid wealth seen, even among the income-poor, may act as a buffer stock that substitutes for expensive and limited access to credit.

In section VI we show that a household's HtM status has strong predictive power for its consumption response to transitory shocks. We apply the identification strategy from Richard Blundell, Luigi Pistaferri, and Ian Preston (2008) to panel data on U.S. income and consumption to measure, for each type of household, the marginal propensity to consume out of transitory income shocks. We find that wealthy HtM and poor HtM households have significantly stronger responses than non-HtM households. In contrast, when we split households into HtM groups based on net worth only, we do not find a significant difference in the consumption responses of those two groups.

In section VII, we argue that the wealthy HtM deserve their own separate status in the cast of characters populating macroeconomic models. We use our empirical estimates of the share of households in each HtM group, together with simulated marginal propensities to consume from three alternative structural models of consumption behavior, to show that the wealthy HtM cannot be assimilated to either the poor HtM or the non-HtM. We highlight four areas where frameworks that do not explicitly model wealthy HtM households provide misguided intuitions about the effects of fiscal policy: the degree of nonlinearity of the marginal propensity to consume with respect to the transfer size, the asymmetry of the consumption response with respect to equal-size income windfalls and losses, the optimal phasing-out of stimulus payments with income for maximizing the impact on aggregate consumption, and the extent of cross-country dispersion in consumption responses to a fiscal transfer. Section VIII summarizes and concludes the paper.

I. Wealthy Hand-to-Mouth Behavior: A Simple Model

We start by analyzing a simple three-period model in order to illustrate the determinants of HtM behavior. In this section, we keep the presentation to a bare minimum; online appendix A contains a more thorough analysis of the problem. The model is also useful to determine how to detect a household's HtM status in the data and, as such, it provides guidance for our measurement exercise.

I.A. Household Problem

Consider a household that lives for three periods--t = 0, 1, and 2--but consumes only in the last two periods. Preferences over consumption at t= 1, 2 are given by

(1) [v.sub.0] = m([c.sub.1]) + m([c.sub.2]),

with no discounting between periods, and with u' > 0, u" < 0. The variable [c.sub.1] denotes nondurable consumption at date t.

In period 0, the household has an initial endowment to and makes a portfolio allocation decision. Two assets are available as saving instruments. An illiquid asset a pays off a gross return R before the consumption decision in period 2, but cannot be accessed at the time of the consumption decision in period 1. A liquid asset m can be accessed before the consumption decision in both periods, but pays a return 1 < R. For now, we do not allow the agent to borrow, that is, to take a negative position in the liquid asset, but we later relax this assumption.

After the initial portfolio allocation decision, households receive income y, and make their consumption and liquid saving decision at t = 1. In the last period, t = 2, they receive income y2 and consume this amount, their liquid savings from t= 1, and their savings allocated to the illiquid asset at t = 0, plus the accrued capital income. Therefore, the only two decisions to analyze are the initial portfolio allocation decision and the consumption/ saving decision at t = 1. Finally, note that since the income path ([y.sub.1], [y.sub.2]) is known at t = 0, there is no uncertainty.

Our characterization of HtM behavior concerns the asset position at the time of the f = 1 consumption decision. We define a household as non-HtM if, after consuming at t = 1, it holds a positive amount of liquid assets, that is, [m.sub.2] > 0 and a [greater than or equal to] 0. As is clear from equation 1, this household will choose [c.sub.1] = [c.sub.2]. We define a household as poor HtM if, after consuming at t = 1, it does not hold any liquid or illiquid assets: [m.sub.2] = 0 and a = 0. We define a household as wealthy HtM if, after consuming at t = 1, it holds a positive amount of illiquid assets but no liquid assets: [m.sub.2] = 0 and a > 0. Therefore, the t = 1 consumption/saving decision determines whether an agent is HtM, and the initial portfolio allocation at t = 0 determines whether an HtM agent is poor or wealthy HtM. For both types of HtM households, [c.sub.1] < [c.sub.2].

I.B. Solution

We begin with the initial portfolio allocation decision at t = 0:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where the first line is the resource constraint in the portfolio choice; the second and third lines are the budget constraints at t = 1 and t = 2; and the final line collects the inequality constraints on the choice variables. The first-order condition of this problem with respect to a gives

(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where the inequality is strict when a = 0. The derivative [partial derivative][m.sub.2]/ [partial derivative]a reflects the dependence of the liquid savings decision at t - 1 on the amount held in illiquid assets. The resulting initial portfolio allocation implicitly determines the endowment points ([y.sub.1] + [omega] - a, [y.sub.2] + Ra) immediately prior to the consumption/saving decision at t = 1.

We now turn to this consumption saving decision at t = 1, given the predetermined amount invested in liquid wealth [m.sub.1] = [omega] - a:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where the first and second lines are the budget constraints at t = 1 and t = 2, and the third line imposes the nonnegativity constraint on the choice variable. The first-order condition of this problem is

(3) u'([c.sub.1]) [greater than or equal to] u'([c.sub.2]),

where the strict inequality holds whenever the constraint binds and [m.sub.2] - 0. For example, when y, is high enough relative to [y.sub.2], the agent wants to save some of his or her income into period 2, and [m.sub.2] > 0. In contrast, when y, is low enough relative to [y.sub.2], the agent would, ideally, like to borrow and is constrained at [m.sub.2] = 0. This "short-run" Euler equation in equation 3 states that, at t = 1, the relative price of consumption between t = 1 and t= 2 is equal to one, the return on the liquid asset.

Combining equations 3 and 2 yields

(4) u'([c.sub.1]) [greater than or equal to] Ru'([c.sub.2]).

This is because 'u([c.sub.1]) = u\[c.sub.2]) when [m.sub.2] is interior, and because [m.sub.2] is unaffected by a marginal change in a when the household is at a constraint. This long-run Euler equation in equation 4 states that, from the agent's viewpoint at t = 0, the relative price of consuming at t = 1 versus t = 2 is R. Comparing equations 4 and 3, the intertemporal trade-off appears to change between t = 0 and t = 1 because the illiquid asset is available as a saving instrument only at t = 0.

The "short-run" Euler equation (3) implies

(5) [m.sub.2] = max {[y.sub.1] + [omega] - [y.sub.2]- (1 + R)a/2, 0}.

Since we are interested in characterizing HtM behavior, we focus on the case where [m.sub.1] - 0. Equation 5 reveals that a sufficient condition for this case is [y.sub.2] [greater than or equal to] [y.sub.1] + [omega]: for a given initial endowment, income in period 2 is so large, relative to period 1, that even when the total endowment [omega] is saved into the liquid asset, the household still desires to consume more at t = 1.

To make further progress on the solution, we assume that u is in the constant elasticity of substitution class, with elasticity of intertemporal substitution [sigma]. Then, the long-run Euler equation (4) gives

(6) a = max{[R.sup.[sigma]] ([y.sub.1] + [omega])-[y.sub.2]/R + [R.sup.[sigma]],0}

From equation 6, we conclude that the household is wealthy HtM when

(7) R > [([y.sub.2]/ [y.sub.1] + [omega]).sup.1/[sigma]]

and is poor HtM when the opposite (weak) inequality holds.

It is useful to explain the role of the model's parameters in determining wealthy HtM behavior. A high relative return R makes the illiquid asset more attractive by raising its effective return, thereby inducing the agent to tolerate wider consumption differences across periods in order to achieve a higher overall consumption level. Steep income growth [y.sub.2]/[y.sub.1] reduces the appeal of the illiquid asset as a saving instrument, since the income path already guarantees high consumption later in life. The higher the elasticity of intertemporal substitution a, the more the household is willing to absorb a jump in consumption across periods, and so the more likely it is to save into the illiquid asset even if [y.sub.1] is low relative to [y.sub.2]. (3)

Since the model is deterministic, wealthy HtM households choose to invest in the illiquid asset at t - 0, even though they know with certainty that they will be constrained in the next period. By acting this way, they consume even less at t = 1 and make themselves even more constrained. Put differently, the shadow value of an additional unit of income at t = 1 is higher for the wealthy HtM than for the poor HtM. If we let this multiplier be [lambda], for a poor HtM [lambda] = u'([y.sub.1] + [omega]) - u'([y.sub.2]), and for a wealthy HtM agent [lambda] = u'([y.sub.1] + [omega] - a) - u'([y.sub.2] + Ra), which is larger. Nevertheless, this choice is optimal because the welfare gain from the rise in the overall level of lifetime consumption more than compensates for the welfare loss from the consumption gap between t = 1 and t = 2.

MARGINAL PROPENSITY TO CONSUME OUT OF A TRANSITORY SHOCK Suppose that after the initial portfolio allocation decision, but before the consumption decision at t = 1, the household receives an unexpected income shock, such as a transfer [tau] from the government. What is the household's marginal propensity to consume out of this transfer? A non-HtM household has a marginal propensity to consume of exactly one-half, since there is no discounting and it smooths the payment equally across the two periods. If the transfer is small enough not to throw the agent off its kink ([m.sub.2] = 0), then the HtM household's marginal propensity to consume out of the transfer will be one. This occurs as long as [tau] [less than or equal to] [y.sub.2] - ([y.sub.1] + [omega]) + (1 + R) a. This condition is weaker for a wealthy HtM than for a poor HtM because, as explained above, the former household is more constrained. (4) Finally, note that all these results carry over to the case of an anticipated transfer, as long as the transfer is small enough that it does not change HtM status at t = 1.

I.C. Taking Stock

Our two-period model is an extremely stylized environment. It is useful to describe how wealthy HtM behavior can arise as a result of giving up gains from additional consumption smoothing in exchange for the opportunity of investing in a high-return asset that yields higher levels of average lifetime consumption. This insight also survives in more general environments. We now briefly discuss five extensions.

First, for some illiquid assets like housing or large durables such as vehicles, the most significant component of their return is the flow of services they provide to the owner. At the same time, they have a consumption commitment component, meaning they require periodic expenditures that cannot be avoided, such as maintenance and repair. Consider a version of our model with the following in period t - 1. The illiquid asset yields a utility flow [phi])a proportional to the stock, and these services are perfect substitutes with c, (housing can be rented out and thus transformed into [c.sub.1]); and the illiquid asset's owner must incur expenditures [kappa]a. Then, the counterpart of condition (7) is one where R is simply replaced by R/(1 - [kappa] + [phi]), the effective return on the illiquid asset.

Second, when the agent can access unsecured credit, there is a second kink in the budget constraint at the credit limit; this is in addition to the kink at [m.sub.2] = 0. The model in online appendix A shows that in this case, households can be wealthy HtM or poor HtM either at the zero kink or at the credit limit.

Third, as we showed in earlier work (Kaplan and Violante 2014a), in the presence of income uncertainty a wealthy HtM prefers bearing the welfare loss from income fluctuations to holding the large balances of cash required for consumption smoothing. Saving in the liquid asset means forgoing the high return on the illiquid asset and the associated higher level of long-run consumption. This explanation is reminiscent of calculations made by Cochrane (1989), Krusell and Smith (1996), and Browning and Crossley (2001), who demonstrate that in several different contexts the utility loss from setting consumption equal to income, instead of fully optimizing, can be second order.

Fourth, in the model the illiquid asset is inaccessible in the intermediate period. In a more general environment where the illiquid asset can be accessed by paying a fixed transaction cost, the household may decide to deposit an unexpected positive windfall into the illiquid account, or to smooth a negative shock by withdrawing from the illiquid account. This behavior could potentially alter the model's implications for the marginal propensity to consume of wealthy HtM agents. In Kaplan and Violante (2014a), we show that this is the case only if the shock is large relative to the transaction cost. We return to this point in section VII.

Finally, in our two-period model, we have abstracted from discounting, but it is easy to see that with geometric discounting between periods, all the qualitative conclusions remain intact. Hyperbolic discounting introduces an additional reason to save in illiquid assets, since illiquidity protects quasi-hyperbolic households from future consumption splurges (see Angeletos and others 2001; Laibson, Repetto, and Tobacman 2003), and therefore makes it even easier to generate wealthy HtM behavior.

II. Identifying Hand-to-Mouth Households in the Data

For both types of HtM household discussed in section I--wealthy and poor--there are two kinks in the intertemporal budget constraint where marginal propensity to consume out of small income changes can be large; at zero liquid assets and at the unsecured credit limit. (5) According to the theory, a household is HtM at the zero kink in period t if it consumes all its cash-on-hand for the period, and carries zero liquid wealth between t and t + 1. Similarly, a household is HtM at the credit limit if, at the end of period t, it has borrowed up to the limit.

Given the theoretical definition of HtM status, ideally we would observe balances of liquid wealth at the end of the pay period--the period that starts at income receipt and ends just before the next income receipt. Unfortunately, surveys either report average balances over the period or report balances at a random point in time (the interview date). As a result, HtM status will be measured with error.

To illustrate this issue, consider a continuous-time generalization of the model in section I where income is paid discretely at the beginning of the period as liquid wealth, but consumption occurs continuously--and is constant--over the period. Given the timing mismatch between the discrete income payment and the continuous consumption expenditures, one would expect to observe positive (or above-credit-limit) balances of liquid wealth, even for the HtM households: this makes their identification especially challenging. In online appendix B, we lay out this enriched version of the model.

We now describe our identification strategy--which builds upon one we used in a separate paper (Kaplan and Violante 2014a)--starting with the case where liquid balances observed from the survey are averages over the period.

II.A. Average Balances

Let [y.sub.it] denote the income of household i in pay period t, let [a.sub.it] denote holdings of illiquid wealth, and let [m.sub.it] denote average balances of liquid wealth over the pay period.

The left-hand panel of figure 1 depicts the dynamics of income and average cash-on-hand [m.sub.it] over a pay period for an HtM household that starts and ends the period at the zero kink. Its liquid balances peak at [y.sub.it] when income is paid into the liquid account at the beginning of the pay period, and are depleted constantly until they reach zero at t + 1. Average balances over the period are equal to one-half income.

A conservative criterion to identify HtM agents on the zero kink in the data is therefore to count those survey households whose average liquid wealth balances are positive (to capture the fact they are not borrowing), but are equal to or less than half their earnings per pay period, where "half" is due to the assumption that resources are consumed at a constant rate. Specifically, a household is poor HtM at the zero kink if

(8) [a.sub.it] [less than or equal to] 0, and 0 [less than or equal to] [m.sub.it] [less than or equal to] [y.sub.it]/2

and a household is wealthy HtM if

(9) [a.sub.it] > 0, and 0 [less than or equal to] [m.sub.it] [less than or equal to] [y.sub.it]/2

[FIGURE 1 OMITTED]

The case [a.sub.it] < 0 is very rare in survey data. It occurs when housing equity is negative because a decline in house prices has pushed the market value of the house below the residual value of the mortgage. We include these households among the poor HtM because, even though they own illiquid assets, they effectively have no means of using them to smooth consumption and, as such, these households are more similar to the poor HtM.

This estimator of the number of HtM households provides a lower bound because, although all non-HtM households would always hold average liquid balances above half their earnings, some HtM households may also hold, on average, liquid balances above half their earnings. For example, a household that starts the period with positive liquid savings, in addition to its earnings, and ends the period with zero liquid savings is HtM, but its average liquid balance is above half its earnings, and so it would not be counted as HtM by this criterion. (Online appendix B makes this point formally.)

Next, consider an HtM household at the credit limit -[[m.bar].sub.it] < 0. This is a household that consumes all its cash-on-hand for the period, as well as all its available credit. For consistency with the strategy above, we propose to count a household as poor HtM at the credit limit if

(10) [a.sub.it] [less than or equal to] 0, [m.sub.it] [less than or equal to] 0, and [m.sub.it] [less than or equal to] [y.sub.it]/2 - [[m.bar].sub.it],

and to count it as wealthy HtM at the credit limit if

(11) [a.sub.it] > 0, [m.sub.it] [less than or equal to] 0, and [m.sub.it] [less than or equal to] [y.sub.it]/2 - [[m.bar].sub.it],

The right-hand panel of figure 1 depicts the dynamics of income and average cash-on-hand [m.sub.it] over a pay period for an HtM household that starts and ends the period at the credit limit. It is easy to see that this criterion is also conservative: a household that starts the period at t with liquid wealth above its credit limit and ends the period at t + 1 having exhausted all its borrowing capacity would carry an average balance above the limit, and would therefore escape our criterion based on equations 10 and 11.

II.B. Balances at a Point in Time

Some surveys report balances of liquid wealth at the interview date, which can be thought of as a random point during the pay period. Is it still true in this case that our estimator, based on the criteria in equations 8 through 11, provides a lower bound on the fraction of HtM households? In online appendix B we show that we would always miss some truly HtM households. However, we might mistake a non-HtM household for an HtM household if its end-of-period liquid balances are less than one-half of its income away from zero or from the credit limit if it is borrowing. For a biweekly pay period, this means that the only problematic households are those with one week or less of income in excess of their kink--households which, for practical purposes, one may want to identify as HtM anyway.

CONSUMPTION COMMITMENTS Recent literature has emphasized the existence of precommitted consumption expenditures--expenditures that a household is committed to incur every pay period, unless it pays a transaction cost (either monetary or in terms of time) to modify its previous commitments (see, for example, Chetty and Szeidl 2007; Stephen Shore and Todd Sinai [2010]). These expenditures include rent, mortgage or other loan payments, utility bills, fees for school, gym, or clubs, and alimony. The key feature of committed expenditures is that they are bulk expenditures incurred at a point in time that discretely deplete a household's balance of liquid wealth.

How does the presence of such expenditures affect our identification strategy? Let [[bar.c].sub.it] be the amount of committed expenditures for household i at date t. If [[bar.c].sub.it] is incurred at the beginning of a pay period, the criterion to identify an HtM household (say, at the zero kink) should be amended as [m.sub.it] [less than or equal to] ([y.sub.it], - [[bar.c].sub.it])/2, while if it is incurred at the end of the period the criterion should be [m.sub.it] - [[bar.c].sub.it] [less than or equal to] [y.sub.it]/2. In the first case, our baseline measurement overestimates HtM status, and in the second case it underestimates it. Instead, if committed expenditures are incurred smoothly over the period or are paid in the middle of the pay period, then the criterion should be, [m.sub.it] - [[bar.c].sub.it]/2 [less than or equal to] ([y.sub.it], - [[bar.c].sub.it])/2 which is the same as our baseline measurement. We verify the robustness of our estimates with respect to those consumption commitments that we can measure in our survey data by using these alternative assumptions about the timing of expenditures.

DEFINITION OF HTM IN TERMS OF NET WORTH For comparison with theories of HtM behavior based on net worth, we also compute the fraction of HtM agents in terms of net worth. Let [n.sub.it] = [a.sub.it] + [m.sub.it] be the net worth of agent i in period t. Then, a household is HtM in net worth (net-worth HtM) if

(12) 0 [less than or equal to] [n.sub.it] [less than or equal to] [y.sub.it]/2 or, [n.sub.it] [less than or equal to] 0 and [n.sub.it] [less than or equal to] [y.sub.it]/2 - [[m.bar].sub.it].

II.C. Direct Survey Questions

Finally, whenever the data allow, we also use direct survey questions as alternate estimates of the fraction of HtM households. These questions typically ask whether expenditures over the last month have exceeded income, abstracting from purchases of large durable goods such as housing or cars, and whether the household usually spends more than its income. Counts of HtM households derived from these questions provide a useful check on the reliability of our identification strategy based on reported liquid wealth and income.

III. Survey Data on Household Portfolios

The eight countries included in our study are the United States, Canada, Australia, the United Kingdom, and the four largest economies in the euro area: Germany, France, Italy, and Spain. Data for the first four countries come from their own separate surveys, the U.S. Survey of Consumer Finances (SCF), the Canadian Survey of Financial Security (SFS), the Household, Income and Labour Dynamics in Australia (HILDA) survey, and the United Kingdom Wealth and Assets Survey (WAS). Data for the euro area countries come from the Household Finance and Consumption Survey (HFCS), a joint project administered by all of the central banks of the Eurosystem. Online appendix C contains a detailed description of all these cross-sectional surveys.

In order to categorize a household as wealthy HtM, poor HtM, or non-HtM, we need information on its labor income and on the amounts of assets and liabilities held in various categories of its balance sheet. In the rest of this section, we discuss sample selection and comparability across surveys. Next, we present some descriptive statistics on the asset and liability distribution across countries.

III.A. Sample Selection and Data Comparability

Each individual survey is tailored to its own country and, as such, the questions asked and the definitions of particular asset classes vary across surveys. Our main goal is to be as consistent as possible in selecting the sample, and in defining income, liquid, and illiquid wealth across surveys.

SAMPLE SELECTION In all surveys, we restrict our analysis to households in which the head is between 22 and 79 years of age, and we drop households only if their income is negative or if all of their income originates from self-employment. (6) Table 1 summarizes the survey years we use for each country, the sample selection, and the final sample sizes. Since all these surveys oversample the rich, we always use weights to construct sample statistics.

INCOME In choosing our definition of income, we try to include all labor income plus any government transfers that are regular inflows of liquid wealth. We exclude interest, dividend, and other capital income because these forms of income are realized more infrequently. For the United States, we define income (from the U.S. SCF) as gross wages and salaries, self-employment income, regular private transfers such as child support and alimony, public transfers such as unemployment benefits, food stamps, and Social Security Income (SSI), and regular income from other sources excluding investment income. For Canada, we define income (from the SFS) as after-tax total income, and there is no distinction between labor, capital, and self-employment income. For Australia, income (from the HIFDA survey) is wages and salaries, self-employment income, regular private transfers such as child support and alimony, and public benefits such as the Australian Government Parenting Payment. For the United Kingdom, we define income (from the WAS survey) as net employee earnings, net self-employment income, and any public benefits such as the Jobseeker's Allowance and Maternity Allowance. For Germany, France, Italy, and Spain, we define income (from the HFCS) as gross income from wages, salaries, and self-employment, unemployment benefits, regular private transfers such as child support and alimony, and regular public transfers. (7)

The main discrepancy in income measurement across surveys is that income in Canada is reported after taxes, whereas all other countries survey gross income before taxes. For most households, except the self-employed, taxes are withheld at the source, hence the amount paid into the liquid account--and available for spending--is net of taxes. Thus, using income before taxes somewhat overstates the fraction of HtM households by inflating the liquid wealth threshold. Whenever possible, we verify the robustness of our results to an adjustment for the individual tax liability.

LIQUID WEALTH Our definition of liquid wealth differs slightly across the surveys, depending on the specific categories of wealth that are available. In the U.S. SCF, our definition of liquid assets consists of checking, saving, money market, and call accounts as well as directly held mutual funds, stocks, corporate bonds, and government bonds. Liquid assets in the Canadian SFS are deposits in financial institutions as well as holdings in mutual funds, other investment funds, and stocks and bonds. In the Australian HILDA, liquid assets include balances in bank accounts, equity investments, and cash investments (bonds). In the U.K. WAS, liquid assets include bank accounts, individual savings accounts (ISAs), and holdings of shares, corporate bonds, and government bonds. (8) For the euro area HFCS, liquid assets are cash, sight (also called current, draft, or checking) accounts, mutual fund holdings, shares in publicly traded companies, and corporate or government bond holdings.

The main shortcoming in the definition of liquid wealth is the absence of information on cash holdings. To address this problem, we resort to an imputation procedure based on data from the 2010 Survey of Consumer Payment Choice, administered by the Federal Reserve Bank of Boston (see Kevin Foster, Scott Schuh, and Hanbing Zhang 2013). We compute the ratio of average cash holdings measured in that survey to the median value of checking, saving, money market, and call accounts from the 2010 U.S. SCF. We then inflate the value of each household's checking, saving, money market, and call accounts by this ratio in all surveys. (9)

We define liquid debt in the U.S. SCF as the sum of all credit card balances that accrue interest, after the most recent payment. Liquid debt in the SFS is credit card and installment debt. Liquid debt in the Australian HILDA is credit card debt. In the U.K. WAS, liquid debt is credit card debt, plus any balances on store cards, hire purchases, and mail orders. In the euro area HFCS, liquid debts are considered to be the balance on credit cards after the most recent payment that accrue interest, together with any balances on credit lines or bank overdrafts that also accrue interest.

The measure of liquid wealth that we use to compute HtM status is net liquid wealth, or liquid assets, minus liquid debt. We also examine liquid wealth by comparing our baseline results both with results from a narrower definition that excludes directly held mutual funds, stocks, and bonds from liquid assets and with results from a broader definition, which includes outstanding debt in home-equity lines of credit. Considering alternative distinctions between liquid and illiquid wealth affects the split between poor and wealthy HtM, but does not affect the total number of HtM households.

ILLIQUID WEALTH Net illiquid wealth in the U.S. SCF includes the value of housing, residential and nonresidential real estate net of mortgages and home equity loans, private retirement accounts (such as 401(k)s, IRAs, thrift accounts, and future pensions), cash value of life insurance policies, certificates of deposit, and saving bonds. Net illiquid wealth in the Canadian SFS is the value of the principal residence and other real estate investment less mortgages on the properties and lines of credit that use property as collateral. It also includes retirement savings such as Registered Retirement Savings Plans, Registered Retirement Income Funds, employer pension plans, and other retirement funds. In the Australian HILDA, net illiquid wealth is net equity in home and other real estate properties plus life insurance policies and superannuation (government-supported, compulsory private retirement funds). (10) In the U.K. WAS, net illiquid wealth includes the value of the main residence, other houses, and land net of mortgages and land debt, plus occupational and personal pensions, insurance products, and National Savings products. The definition of net illiquid wealth in the euro area HFCS is the value of the household's main residence and other properties net of mortgages and unsecured loans specifically taken out to purchase the home, plus occupational and voluntary pension plans, cash value of life insurance policies, certificates of deposit, and saving bonds.

We also explore broader definitions of illiquid wealth that include the value of businesses for the self-employed, the resale value of vehicles net of the loans taken out to purchase them, and other nonfinancial wealth not included in our baseline, such as antiques, artwork, jewels, and gold. (11)

REFERENCE PERIOD The reference period for the liquid and illiquid wealth questions varies across surveys. In the U.S. SCF, for most assets it is the interview date; for some assets, such as checking and saving accounts, when a respondent is unsure about balances the interview can prompt for an average balance over the month. The Canadian SFS asks for information on assets and debts for "a time as close as possible to the date of the interview." Both the U.K. WAS and Australian HILDA ask for current balances or values of assets and liabilities. In the HCFS, France, Germany, and Spain use the interview date, and Italy uses December 31, 2010.

III.B. Descriptive Statistics

Table 2 reports some basic descriptive statistics on household income, liquid and illiquid wealth holdings, and portfolio composition, for each country in the sample.

In all countries, the typical household portfolio structure is rather simple. It comprises a small amount of liquid wealth in the form of bank accounts, some housing equity, and a private retirement account. In particular, the median holdings of other financial assets such as directly held stocks, bonds, mutual funds, and life insurance are zero everywhere. This is a well established fact borne out by empirical studies of household portfolios (see Guiso, Halassios, and Jappelli, 2002).

However, there are some interesting cross-country differences in household portfolios. First, the ratio of median net worth to median income varies widely across countries: from just above 1:1 in Germany and the United States to over 6:1 in the United Kingdom, Italy, and Spain. With respect to net liquid wealth, consumer credit appears much less frequently in the euro area: less than 10 percent of households have credit card debt in France, Italy, and Spain, compared to 30 to 40 percent in the Anglo-Saxon countries. Figure 2, which plots the distribution of net liquid wealth to monthly income for the eight countries, reinforces this observation.

Housing equity forms the majority of illiquid wealth for households in every country with the exception of Germany, where median housing wealth is zero, since only 48 percent of the population are homeowners. This homeownership rate is at least 10 percentage points less than in the other seven countries (see also Eymann and Borsch-Supan 2002). The median value of housing equity relative to median annual income is especially remarkable in Italy and Spain, where it exceeds 6:1.

There are also large differences in the fraction of households with positive private retirement wealth: in the Anglo-Saxon countries, at least half of all households hold a personal retirement account, whereas in France, Italy, and Spain less than one-tenth do. Surely, a big part of the explanation is in the generosity of the public pension system in these countries: according to the OECD, replacement rates for the median earner are between 60 and 70 percent in these countries, compared to 40 percent in the United Kingdom and the United States (see OECD 2013). The size of private retirement wealth in Australia and the United Kingdom is astonishing. In Australia, this is partly due to the "superannuation" regulations that require all employers to generously contribute to tax-deferred retirement accounts on behalf of their employees. (12) In the United Kingdom, the Pension Schemes Act of 1993 created tax-free employer-sponsored (defined benefits) occupational pensions and (defined contributions) personal pensions, while the Pension Act of 2008 established that workers must choose to opt out of an employer's occupational pension plan, rather than opt in (see Banks and Tanner 2002 for more details of the options available for retirement savings in the United Kingdom).

Finally, the proportion of households with life insurance in their portfolio is much higher in the euro area than in the Anglo-Saxon countries.

[FIGURE 2 OMITTED]

We conjecture that solid intergenerational family ties and a stronger precautionary savings motive linked to the lower rate of female participation in the workforce may account for these differences.

IV. United States

Next we report the main findings for the United States, using data from the 1989-2010 waves of the U.S. SCF. We begin by estimating the fraction of HtM households and assessing the robustness of our estimates to a variety of aspects of the definition adopted in section II. We then analyze the key demographic characteristics of non-HtM, poor HtM, and wealthy HtM households, and we examine their portfolio composition in more detail.

IV.A. The Share of HtM Households

Our definition of HtM status is based on equations 8 through 12. Since the U.S. SCF does not report individual data on the frequency of pay, we need to make an assumption that applies to all households. Consumer Expenditure Survey data from 1990 to 2010 reveal that 32 percent of respondents are paid weekly, 52 percent of respondents are paid biweekly, and the rest are paid monthly or at lower frequencies. (13) Based on these findings, in the benchmark analysis we set the pay frequency to two weeks. In the benchmark, we also set the household credit limit to one month of income. The U.S. SCF asks respondents to report their credit limit, but most of the other surveys do not, so for comparability we choose a common limit. (14)

The lower panel of figure 3 plots the fraction of HtM households in the U.S. population over the period 1989-2010 and shows the split between wealthy and poor HtM. Our estimates indicate that, on average, 31 percent of U.S. households are HtM over this period. Of these, roughly one-third are poor HtM and two-thirds are wealthy HtM. This is our paper's first main result: the vast majority of hand-to-mouth households own illiquid assets. Looking at changes over time across the two decades covered by our data, the fraction of HtM households remains fairly stable and the split between poor and wealthy does not change significantly. The first line of table 3 reports that the share of U.S. households that are HtM in terms of net worth is less than 14 percent. Thus, looking at wealth distribution through the eyes of net worth alone misses more than half of the HtM households in the United States. (15)

[FIGURE 3 OMITTED]

[FIGURE 4 OMITTED]

The lower panel of figure 3 explores the illiquid asset portfolio of the wealthy HtM households by plotting the share of wealthy HtM households that own housing, nonhousing illiquid wealth, or both. About half of wealthy HtM households have both, about a third have positive housing but no nonhousing illiquid wealth, and a sixth have nonhousing illiquid wealth but no housing wealth. A deeper look into the portfolio of HtM households reveals that, if we condition on homeownership, the leverage ratio is a strong predictor of HtM status. Figure 4 shows that the fraction of HtM households doubles from 20 to 40 percent as the leverage ratio rises toward one, as regular mortgage payments absorb a significant fraction of disposable income and leave households with little or no liquid savings.

[FIGURE 5 OMITTED]

ROBUSTNESS Figure 5 and Table 3 summarize our sensitivity analyses. In figure 5, which covers the United States, the upper-left panel plots the shares of poor and wealthy HtM households weighted by income. Not surprisingly, the weighted fraction of HtM households is smaller than its unweighted counterpart: HtM households represent roughly 20 percent of total U.S. income, since their income is below the U.S. average. When we weight by income, however, wealthy HtM households represent three-quarters of all HtM households. The upper-right panel of figure 5 plots HtM shares when the pay period is set to one month instead of two weeks: the fraction of HtM households increases by 9 percentage points and wealthy HtM households account for most of the difference with the baseline.

Symmetrically, the fourth line of table 3 shows that, when the pay period is set to one week, the share of wealthy HtM households drops by 5 percentage points. In the lower-left panel of figure 5, we verify the robustness of our estimates with respect to the tightness of the credit limit. When we use the self-reported credit limit in the U.S. SCF, the fraction of HtM households drops by 5 percentage points, with a lower number of wealthy HtM households accounting for all of the drop. Finally, the lower-right panel shows that by including vehicles as illiquid wealth, we move roughly half of the poor HtM into the wealthy HtM group but, by construction, the total share of HtM households in the population is unchanged.

Table 3 contains a number of other sensitivity analyses. We begin with direct questions on HtM status. The U.S. SCF contains a combination of sequential questions aimed at assessing whether "over the past year, [household] spending exceeded, or was about the same as, income, and such expenditures included purchases of a home or automobile or spending for any investments." (16) Based on this definition, the share of HtM households is around 44 percent. Wealthy HtM households account for two-thirds of the total, and fluctuations in this measure over time very closely follow those in the baseline definition of figure 5 (upper-left panel). The third row of table 3 also reports results for another sequence of direct questions in the U.S. SCF. The first question asks households, "Which of the following statements comes closest to describing your saving habits?" We label a household as HtM if it responds "Don't save--usually spend more than (or as much as) income." Roughly 24 percent of households are HtM according to this definition.

It is reassuring that our baseline estimate of HtM households sits in between the counts based on these two direct questions, since we interpret the first question as providing an upper bound and the second as providing a lower bound. Our baseline calculations refer to the current HtM status for a household. In the first set of direct questions, although households that spent more than their income over the past year because they dis-saved or borrowed are not truly HtM, they would still be classified as such based on the questions. Conversely, the second set of direct questions asks about the usual HtM status, and therefore those households that are, at the time of the survey, temporarily in an HtM status would answer the question negatively. The cross-sectional correlation between our indicator of HtM status and the one provided by these two questions is about 0.3 for each.

Our estimates of HtM households are related to calculations of "financially fragile" households by Lusardi, Schneider, and Tufano (2011). Based on an ad-hoc survey, they document that a quarter of U.S. households report that they would certainly be unable to come up with $2,000 in 30 days, and a similar fraction reports that they could probably not come up with the funds to deal with an ordinary financial shock of this size. These authors also emphasize that there are many solidly middle-class households in this last group. In line three of table 3, we compute the fraction of households that are less than $2,000 away from the liquid wealth thresholds for being defined as HtM. We find that 50 percent of households are "financially fragile" according to this definition. Of these, 17 percent have no illiquid assets, but 33 percent own housing or retirement wealth (or both). The poor HtM could be mapped onto the ad-hoc survey respondents who would certainly not come up with this amount, and the wealthy HtM could be mapped onto those who would probably be unable to cope.

Overall, our estimates are in line with those of Lusardi, Schneider, and Tufano, but they also suggest a more nuanced interpretation. Households in the second group (who could "probably not come up the funds") should have the means to deal with a shock of this size, for example by using their illiquid wealth as collateral for a loan. They may choose not to do so because the transaction costs involved would dominate the welfare gain from smoothing such a small shock, but for larger shocks they would choose to adjust and smooth consumption. We return to this shock-size asymmetry of behavior in section VII. (17)

The other robustness checks in table 3 are conducted with respect to the definition of illiquid wealth, debt, income, and the timing of consumption expenditures. Using a higher illiquid wealth threshold in the definition of wealthy HtM ($1,000 instead of $1) moves about 1 percentage point of households from the wealthy HtM category into that of poor HtM. Broadening the definition of illiquid wealth to include business equity, directly held stocks and bonds, or other valuables (such as artwork, antiques, and jewels) has small effects relative to the baseline. (18) Including all private retirement wealth as liquid wealth for households headed by persons age 60 or above reduces the share of wealthy HtM households by less than 1 percentage point.

Around one-quarter of U.S. households simultaneously have positive liquid assets above y 12 and some revolving credit card debt. (19) One may worry that many of these households have net liquid wealth close to zero and they would therefore be counted as HtM, even though they have slack in both liquid wealth and credit. In table 3 we show that excluding this group does not affect our calculations much, because the distribution of HtM status within this group is not very different from the population distribution. Home equity lines of credit (HELOCs) were virtually nonexistent before the year 2000, but in the last decade they became a more common instrument to extract liquidity from housing. (20) Changing the definition of liquid debt by including used-up HELOCs--while simultaneously increasing the credit limit by the total available line of credit--decreases the fraction of HtM households, as expected, but by only 1 percentage point.

The U.S. SCF collects data on a household's normal, or usual, income as well as its actual income. This alternate definition of income has no effect on our calculations. Recall that our definition of income is gross income before taxes and tax credits. Through the National Bureau of Economic Research's TAXSIM data files, we have constructed, household by household, a measure of after-tax income. (21) Under this income measure, the total fraction of HtM households declines, but quantitatively this effect is very small. The reason is that, in the United States, the effective average tax rate is very small at the low end of the income distribution (around zero), mainly because of the Earned Income Tax Credit; even in the middle quintile it is only 10 percent.

Finally, as explained in section II, accounting for committed expenditures has an ambiguous effect on the share of HtM agents, depending on whether the expenditures occur primarily at the beginning or at the end of the pay period. Table 3 shows that these two opposite timing assumptions bound the share of total HtM households between 27 and 42 percent.

[FIGURE 6 OMITTED]

IV.B. Demographics, Portfolio Composition, and Status Persistence

DEMOGRAPHICS We now turn to the demographic characteristics of the three groups of HtM households. Figure 6 plots the share of the population that is wealthy HtM and poor HtM by age. (22) The bulk of poor HtM household behavior is observed in the early stages of the life cycle. The fraction of poor HtM households drops sharply until age 30, and keeps falling steadily over the life cycle until reaching roughly 5 percent in retirement. By contrast, the age profile of the fraction of wealthy HtM households is markedly hump-shaped: it peaks at around age 40, when over 20 percent of U.S. households are wealthy HtM, and it remains above 10 percent throughout the life cycle. Accordingly, the share of non-HtM individuals increases steadily from 50 percent at age 22 to 80 percent in retirement.

The first three panels of figure 7 report some demographic characteristics of the three HtM groups by age. (23) Non-HtM households have on average one more year of education than wealthy HtM households, which, in turn, have one more year of education than poor HtM households. In terms of marital status, non-HtM and wealthy HtM households are indistinguishable, whereas poor HtM households are 30 percent less likely to be married. In contrast, poor HtM and wealthy HtM are both more likely to have children than are non-HtM households.

The middle-right panel of figure 7 shows that poor HtM households are income-poor, with median annual income around $20,000 (in 2010 dollars) during the working years, while the non-HtM are high-income households whose median earnings are $70,000 at their life-cycle peak. The most surprising finding is that the wealthy HtM look a lot like the non-HtM in their income path. The same conclusion holds for the incidence of unemployment and for the likelihood of receiving welfare benefits, which are both much lower for non-HtM and wealthy HtM households than for poor HtM households.

PORTFOLIO COMPOSITION Figure 8 digs deeper into the balance-sheet composition of the three groups of HtM households. The upper-left panel shows that median net liquid wealth holdings are zero at virtually every age for both the poor HtM and the wealthy HtM. Median net liquid wealth for non-HtM households grows steadily from about $2,500 at age 25 until retirement, where it levels off at roughly $15,000. (24) The upper-right panel reveals that the wealthy HtM households hold significant amounts of illiquid wealth: for example, median holdings at age 40 exceed $50,000. Hence, wealthy HtM households are not just poor HtM households with small amounts of savings in less liquid assets. The two lower panels of figure 8 articulate this observation further, plotting the age profiles of the average fraction of illiquid wealth held in housing and retirement accounts for wealthy HtM and non-HtM households. The conclusion is striking: the lines are on top of each other, indicating that the portfolio allocation of these two groups is nearly identical.

PERSISTENCE How persistent is a household's HtM status? We answer this question by exploiting the 2007-09 panel component of the U.S. SCF. Table 4 reports the 2-year transition matrix across the three HtM statuses for U.S. households. The diagonal elements of the matrix reveal that non-HtM status is by far the most persistent, and wealthy HtM status the most transient of the three. These transition probabilities imply that the expected length of HtM status is around 3.5 years for wealthy HtM households, 4.5 years for poor HtM households, and 11 years for the non-HtM.

[FIGURE 7 OMITTED]

[FIGURE 8 OMITTED]

V. Cross-Country Evidence

The previous section showed that around 30 percent of households in the United States are HtM, and that of these households one-third are poor HtM and two-thirds are wealthy HtM. In this section we use household portfolio data from seven other developed economies to assess whether the prevalence of wealthy HtM households is a common feature of the wealth distribution across countries and, if so, whether the demographic, income, and balance-sheet characteristics of wealthy HtM households in these countries are similar to those in the United States.

As discussed in section III, we focus our attention on three other Anglo-Saxon countries--Canada, Australia, and the United Kingdom--and the four largest euro area economies, Germany, France, Italy, and Spain. While data are available for more than one point in time for most of these countries, in order to keep the discussion manageable we focus on the most recent single cross-section in each country. For Australia and the European countries this is 2010, for the United Kingdom it is 2009, and for Canada it is 2005. For the sake of comparability, we use only the 2010 wave of the SCF for the United States.

Figure 9, upper panel, shows the fraction of poor and wealthy HtM households in each country. There is a striking similarity among the United States, Canada, and the United Kingdom in their overall fraction of HtM households as well as the breakdown between poor and wealthy HtM. These three countries have a large share of HtM households, exceeding 30 percent. Australia is an outlier among the Anglo-Saxon countries in two ways: first, its total fraction of HtM households is roughly half the fraction in the other three countries (the United States, the United Kingdom, and Canada); and second, 90 percent of its HtM households are wealthy. Among the euro area countries, France, Italy, and Spain have smaller shares of HtM households than the United States, the United Kingdom, and Canada--at around 20 percent--whereas in Germany this share is closer to 30 percent. For all eight countries, there are more wealthy than poor HtM households; even for the euro area countries, the fraction of wealthy among the HtM households exceeds two-thirds. Thus, a widespread feature of household portfolios across countries is that a complete characterization of the fraction of the population that is likely to exhibit HtM behavior requires going beyond an examination based simply on low net worth.

[FIGURE 9 OMITTED]

The lower panel of figure 9 reveals that there are significant differences in the portfolio composition of wealthy HtM households across countries. In Italy and Spain, virtually all the wealthy HtM own some housing wealth. Homeowners are also dominant among the wealthy HtM in the United States and Canada. In contrast, around half of the wealthy HtM in Australia, Germany, and Canada have no housing wealth; rather, the majority of their illiquid assets are held in private retirement accounts. Table D1 in the online appendix provides more information on the cross-country portfolio composition.

What explains the fact that the euro area countries have a smaller fraction of HtM households than the United States? In the euro area countries, households hold more liquid wealth relative to their income than is the case in the United States. As is clear from figure 2, this fact can be partly attributed to differences in liquid debt. The fraction of poor HtM households in the euro area countries with negative liquid wealth is two to four times smaller than in the Anglo-Saxon countries (see online appendix table Dl). Presumably, lower access to unsecured credit in Europe implies that households have more incentives to hold large balances of liquid wealth for transaction and precautionary reasons. For example, Daniela Vandone (2009) documents that, in 2006, the total value of consumer credit amounted to 25 percent of disposable income in the United Kingdom, as compared with 15 percent in Germany and Spain, 12 percent in France, and only 10 percent in Italy.

Australia is the country with the largest share of wealthy HtM among its HtM households. Online appendix table Dl shows that this can be traced to the very high share of the country's population that owns private retirement wealth. As explained in section III, the high ownership rate of retirement accounts in Australia is largely due to the country's superannuation regulations. When we exclude superannuation accounts as a component of wealth, the fraction of poor HtM in Australia rises from 3 to 9 percent and the fraction of wealthy HtM drops accordingly.

AGE PROFILES Age profiles of the fraction of poor and wealthy HtM households in each country are shown in figure 10. For most countries, the fraction of poor HtM households declines monotonically with age. The exceptions are Australia and France, where the age profiles of the poor HtM are flat. There are some marked differences in the age profiles of the wealthy HtM that can be explained by differences in portfolio holdings across countries. In countries where housing wealth is a substantial part of household portfolios, such as the United States, Canada, and the United Kingdom, the age profile is hump-shaped, peaking in the early 40s. In contrast, in Australia and Germany, where a high fraction of wealthy HtM households hold retirement accounts, the share of wealthy HtM decreases with age.

[FIGURE 10 OMITTED]

An important caveat to these results is that because we infer age profiles from a single cross-section, we necessarily confound age, cohort, and time effects. This could explain, for example, why in Spain the share of wealthy HtM falls steadily with age. This pattern may reflect time effects, since recent 25- to 35-year-olds have faced much harsher economic conditions upon entry into the labor market than earlier cohorts. (25)

V.A. Robustness

Table 5 contains an extensive sensitivity analysis of our definitions of poor HtM and wealthy HtM households that parallels Table 3.

Questions on whether household spending exceeded income in the past year are present in all surveys. As we found in the United States, in the other seven countries we find larger shares of both poor HtM and wealthy HtM households when we use these direct spending vs. income questions to measure the incidence of HtM behavior. The difference is especially marked for Italy and Spain where, according to this criterion, more than 60 percent of households--and hence three times the baseline estimate--are HtM. Extending the credit limit from one month of income to one year of income has a substantial effect for the Anglo-Saxon countries, but virtually no impact for the euro area countries. This finding is in line with the empirical distribution of liquid assets documented in figure 2, which showed that households with negative net liquid wealth are extremely rare in the euro area countries. (26)

The fraction of "financially fragile" households (those with liquid balances lower than the threshold plus 2,000 local currency units (27)) is only 10 to 15 percentage points larger than the share of HtM households in the Anglo-Saxon countries, but in most of the euro area countries it is 30 percentage points larger. This result is consistent with the distributions of liquid wealth reported in figure 2, which show that in euro area countries there is a large mass of households just to the right of the threshold.

Shortening the pay period from the biweekly baseline to one week (or extending it to a month) has a small impact on the fraction of poor HtM households, but it decreases (or increases, respectively) the fraction of wealthy HtM households by 5 percentage points on average. Including vehicles as illiquid wealth shifts HtM households from poor to wealthy in every country, although to a lesser extent than it does in the United States. In two countries, Canada and Italy, including other nonfinancial assets (such as valuables and collectibles) in the definition of illiquid wealth shifts 12 percent (Canada) and 5 percent (Italy) of households from poor to wealthy HtM. (28) Including HELOCs among liquid debt has no effect, except in Canada, where the share of HtM increases by 8 percentage points.

Our baseline measure of income is income after transfers but before taxes, except for Canada, where it is disposable income. For three countries--the United States, the United Kingdom, and Italy--we can analyze the effect of netting taxes at the source for every household, and find that the effect of this correction is minor. (29)