JEP Final Draft

A Theory of the Consumption Function,

Christopher D. Carroll

May 18, 2001

Keywords: consumption, precautionary saving, uncertainty, Permanent Income Hy-

JEL Classification Codes: A23, B22, D91, E21

Christopher D. Carroll is Professor of Economics, Johns Hopkins Univer-

Function seemed badly dated. Dynamic optimization theory not been employed much

in economics when Friedman wrote, and utility theory was still comparatively primitive,

so his statement of the “permanent income hypothesis” never actually specified a formal

mathematical model of behavior derived explicitly from utility maximization. Instead,

Friedman relied at crucial points on intuition and verbal descriptions of behavior.

Although these descriptions sounded plausible, when other economists subsequently

found multiperiod maximizing models that could be solved explicitly, the implications

of those models differed sharply from Friedman’s intuitive description of his ‘model.’

Furthermore, empirical tests in the 1970s and ’80s often rejected these rigorous versions

of the permanent income hypothesis, in favor of an alternative hypothesis that many

households simply spent all of their current income.

Today, with the benefit of a further round of mathematical (and computational)

advances, Friedman’s (1957) original analysis looks more prescient than primitive. It

turns out that when there is meaningful uncertainty in future labor income, the optimal

behavior of moderately impatient consumers is much better described by Friedman’s

original statement of the permanent income hypothesis than by the later explicit max-

imizing versions. Furthermore, in a remarkable irony, much of the empirical evidence

that rejected the permanent income hypothesis as specified in tests of the 1970s and

’80s is actually consistent both with Friedman’s original description of the model and

with the new version with serious uncertainty.

1
 There are four key differences between the explicit maximizing models developed

in the 1960s and ’70s and Friedman’s model as stated in A Theory of the Consumption

Function (and its important clarification in Friedman (1963)).

First, Friedman repeatedly acknowledged the importance of precautionary saving

motives induced by uncertainty about the future level of labor income. In contrast, the

crucial assumption that allowed subsequent theorists to solve their formal maximizing

models was that labor income uncertainty had no effect on consumption, either because

uncertainty was assumed not to exist (in the “perfect foresight” model) or because the

utility function took a special form that ruled out precautionary motives (the “certainty

equivalent’ model).1

Second, Friedman asserted that his conception of the permanent income hypothesis

implied that the marginal propensity to consume out of transitory “windfall” shocks

to income was about a third. However, the perfect foresight and certainty equivalent

models typically implied an MPC of 5 percent or less.

Third, Friedman asserted that the “permanent income” that determined current

spending was something like a mean of the expected level of income in the very near-

term: “It would be tempting to interpret the permanent component [of income] as

corresponding to the average lifetime value . . . It would, however, be a serious mis-

take to accept such an interpretation.” He goes on to say that households in practice

adopt a much shorter ‘horizon’ than the remainder of their lifetimes, as captured in

the assumption in Friedman (1963) that people discount future income at a “subjec-

2
tive discount rate” of 33-1/3 percent. In contrast, the perfect foresight and certainty

equivalent models assumed that future income was discounted to the present at market

interest rates (say, 4 percent).

Finally, as an interaction between all of the preceding points, Friedman indicated

that the reason distant future labor income had little influence on current consumption

was “capital market imperfections,” which encompassed both the fact that future labor

income was uninsurably uncertain and the difficulty of borrowing against such income

(for example, see Friedman (1963) p. 10).

It may seem remarkable that simply adding labor income uncertainty can trans-

form the perfect foresight model into something closely resembling Friedman’s original

framework; in fact, one additional element is required to make the new model generate

Friedmanesque behavior: Consumers must be at least moderately impatient. The key

insight is that the precautionary saving motive intensifies as wealth declines, because

poorer consumers are less able to buffer their consumption against bad shocks. At

some point, the intensifying precautionary motive becomes strong enough to check the

decline in wealth that would otherwise be caused by impatience. The level of wealth

where the tug-of-war between impatience and prudence reaches a stalemate defines a

‘target’ for the buffer stock of precautionary wealth, and many of the insights from the

new model can best be understood by considering the implications and properties of

this target.

A final insight from the new analysis is that precautionary saving behavior and liq-

3
uidity constraints are intimately connected.2 Indeed, for many purposes the behavior

of constrained consumers is virtually indistinguishable from the behavior of uncon-

strained consumers with a precautionary motive; average behavior depends mainly on

the consumer’s degree of impatience, not on the presence or absence of constraints.

As a result, most of the existing empirical studies that supposedly test for constraints

should probably be reinterpreted as evidence on the average degree of impatience. Fur-

thermore, future studies should probably focus more directly on attempting to measure

the average degree of impatience rather than on attempting to detect constraints.

1 The Modern Model(s) of Consumption

Current graduate students rarely appreciate how difficult it was to forge today’s canon-

ical model of consumption based on multiperiod utility maximization. The difficulty

of the enterprise is attested by the volume of literature devoted to the problem from

the 1950s through the ’70s, beginning with the seminal contribution of Modigliani and

Brumberg (1954). The model that eventually emerged has several key characteristics.

Utility is time separable; that is, the utility that consumption yields today does not

depend on the levels of consumption in other periods, past or future. Future utility

is discounted geometrically, so that utility one period away is worth β units of this

period’s utility, utility two periods away is worth β 2 , and so on, for some β between

0 and 1. Furthermore, the utility function must satisfy various criteria of plausibility

4
like decreasing marginal utility, decreasing absolute risk aversion, and so on. Finally,

the model must incorporate a mathematically rigorous description of how noncapital

income, capital income, and wealth evolve over time.

One of the unpleasant discoveries in the 1960s and ’70s was that when there is

uncertainty about the future level of labor income, it appears to be impossible (under

plausible assumptions about the utility function to derive an explicit solution for con-

sumption as a direct (analytical) function of the model’s parameters. This is not to

say that nothing at all is known about the structure of optimal behavior under uncer-

tainty; for example, it can be proven that consumption always rises in response to a

pure increment to wealth. But an explicit solution for consumption is not available.

1.1 The Perfect Foresight/Certainty Equivalent Model

Economists’ main response to this problem was to focus on two special cases where the

model can be solved analytically: The “perfect foresight” version in which uncertainty

is simply assumed away, or the “certainty equivalent” version in which consumers

are assumed to have quadratic utility functions (despite unattractive implications of

quadratic utility like risk aversion that increases as wealth rises, and the existence of

a ‘bliss point’ beyond which extra consumption reduces utility).

The perfect foresight and certainty equivalent solutions are very similar; for brevity,

I will summarize only the perfect foresight solution, in which the optimal level of

consumption is directly proportional to total “wealth,” which is the sum of market

5
wealth Wt and ‘human wealth’ Ht ,

Ct = kt (Wt + Ht ), (1)

where market wealth Wt is real and financial capital, while human wealth is mainly

current and discounted future labor income (though in principle Ht also includes the

discounted value of transfers and any other income not contingent on saving decisions;

henceforth I refer to these collectively as ‘noncapital’ income). The constant of pro-

portionality, kt , depends the time preference rate, the interest rate, and other factors.

A simple example occurs when consumers care exactly as much about future utility

as about current utility (β = 1); the interest rate is zero; and there is no current or

future noncapital income (Ht = 0). In this case, the optimal plan is to divide existing

wealth evenly among the remaining periods of life. If we assume an average age of

death of 85, this model implies that the marginal propensity to consume out of shocks

to wealth for consumers younger than 65 should be less than (1/20), or 5 percent –

since the change in wealth will be spread evenly over at least 20 years. Furthermore,

the theory implies that the MPC out of unexpected transitory shocks to noncapital

income (“windfalls”; e.g. finding a $100 bill in the street) is the same as the MPC out of

wealth, because once the windfall has been received, it is theoretically indistinguishable

from the wealth the consumer already owned. When the model is made more realistic

by allowing for positive interest rates, consumers younger than 65, etcetera, it still

implies that the average MPC should be quite low, generally less than 0.05.

6
 In contrast, Friedman (1963) asserted that his conception of the permanent income

hypothesis implied an MPC out of transitory shocks of about 0.33 for the typical

consumer.3 Friedman (1963) provided an extensive summary of the existing empirical

evidence tending to support the proposition of an MPC of roughly a third.4 From

today’s perspective, however, the most surprising aspect of Friedman’s (1957, 1963)

arguments is that their main thrust is to prove an MPC much less than one (to discredit

the ‘Keynesian’ model that said consumption was roughly equal to current income),

rather than to prove an MPC significantly greater than 0.05.

The 15 years after the publication of A Theory of the Consumption Function pro-

duced many studies of the MPC. Particularly interesting were some natural exper-

iments. In 1950, unanticipated payments were made to a subset of U.S. veterans

holding National Service Life Insurance policies; the marginal propensity to consume

out of these dividends seems to have been between about 0.3 and 0.5. Another natural

experiment was the reparations payments certain Israelis received from Germany in

1957-58.5 The marginal propensity to consume out of these payments appears to have

been around 20 percent, with the lower figure perhaps accounted for by the fact that

the reparations payments were very large (typically about a year’s worth of income).

On the whole, these studies were viewed at the time as supporting Friedman’s model

because the estimated MPCs were much less than one.

The change in the profession’s conception of the permanent income hypothesis

in the 1970s from Friedman’s (1957, 1963) version to the perfect foresight/certainty

7
equivalent versions (with their predictions of an MPC of 0.05 or less) is nicely illustrated

by a well-known paper by Hall and Mishkin (1982) that found evidence of an MPC

of about 0.2 using data from the Panel Study of Income Dynamics (PSID). Rather

than treating than this as evidence in favor of a Friedmanesque interpretation of the

permanent income hypothesis, the authors concluded that at least 15-20 percent of

consumers failed to obey the PIH because their MPCs were much greater than 0.05.

1.2 The New Model

The principal development in consumption theory in the last 15 years or so, start-

ing with Zeldes (1984), is that spectacular advances in computer speed have allowed

economists to relax the perfect foresight/certainty equivalence assumption and deter-

mine optimal behavior under realistic assumptions about uncertainty.

A preliminary step was to determine the characteristics of the income uncertainty

that typical households face. Using annual income data for working-age households

participating in the PSID, Carroll (1992) found that the household noncapital income

process is well approximated as follows. In period t a household has a certain level of

‘permanent noncapital income’ Pt , which is defined as the level of noncapital income the

household would have gotten in the absence of any transitory shocks to income. Actual

income is equal to permanent income multiplied by a transitory shock, Yt = Pt t where

permanent income Pt grows by a factor G over time, Pt = GPt−1 . Each year there

is a small chance (probability 0.005) that actual household income will be essentially

8
zero (t = 0), typically corresponding in the empirical data to a spell of unemployment

or temporary illness or disability. If the transitory shock does not reduce income all

the way to zero, that shock is distributed lognormally with a mean value of one and a

standard deviation of σ = 0.1. Carroll (1992) and subsequent papers also find strong

evidence for permanent as well as transitory shocks to income, also with an annual

standard deviation of perhaps 0.1. However, because permanent shocks complicate

the exposition without yielding much conceptual payoff, I will suppress them for the

purposes of this paper and compensate by boosting the variance of the transitory

component to σ = 0.2; for the version with both transitory and permanent shocks,

see Carroll (1992). The PSID also shows the annual household income growth factor

to be about G = 1.03 or 3 percent growth per year for households whose head is in the

prime earning years of 25-50.

The next step in solving the model computationally is to choose values for the

parameters that characterize consumers’ tastes. For the simulation results presented

in this paper, I will assume a rather modest precautionary saving motive by choosing

a coefficient of relative risk aversion of ρ = 2, toward the low end of the range from

1 to 5 generally considered plausible.6 I follow a traditional calibration in the macro

literature and choose a time preference factor of β = 0.96 implying that consumers

discount future utility at a rate of about 4 percent annually, and I make a symmetric

assumption that the interest rate is also 4 percent per year.

We are now in position to describe how the model can be solved computationally.

9
As is usual in this literature, it is necessary to solve backwards from the last period

of life. For simplicity, we will assume that the income process described above, with

constant income growth G, holds for every year of life up to the last. (For a version

with a more realistic treatment of the lifetime income profile, including the drop in

income at retirement, see Carroll (1997)).

In the last time period, the solution is easy: The benchmark model assumes there

is no bequest motive, so the consumer spends everything. Following Deaton (1991),

define cash-on-hand X as the sum of noncapital income and beginning-of-period wealth

(including any interest income earned on last period’s savings). In the second-to-last

period of life, the consumer’s goal is to maximize the sum of utility from consumption

in period T −1 and the mathematical expectation of utility from consumption in period

T , taking into account the uncertainty that results from the possible shocks to future

income YT . For any specific numerical levels of cash-on-hand and permanent income

in period T − 1 (say, XT −1 = 5 and PT −1 = 1.4), a computer can calculate the sum

of current and expected future utility generated by any particular consumption choice.

The optimal level of consumption for {XT −1 , PT −1 } = {5, 1.4} can thus be found by a

computational algorithm that essentially tries out different guesses for CT −1 and homes

in on the choice that yields the highest current and discounted expected future utility.

Note that for each different combination of {XT −1 , PT −1 }, the utility consequences

of many possible choices of CT −1 must be compared to find the optimum, and for

each CT −1 that is considered, the numerical expectation of next period’s utility must

10
be computed. The solution procedure is basically to calculate optimal CT −1 for a

grid of many possible {XT −1 , PT −1 } choices, and then to construct an approximate

consumption function by interpolation (‘connect-the-dots’).

Once the approximate consumption rule has been constructed for period T − 1, the

same steps can be repeated to construct a consumption rule for T − 2 and so on.

This begins to give the flavor for why numerical solutions are so computation in-

tensive. Indeed, the problem as just described would be something of a challenge even

for current technology. Fortunately, there is a trick that makes the problem an order of

magnitude easier: Everything can be divided by the level of permanent income. That

is, defining the cash-on-hand ratio as xt = Xt /Pt and ct = Ct /Pt , it is possible to find

the optimal value of the consumption-to-permanent-income ratio as a function of the

cash-on-hand ratio, so that rather than solving the problem for a two-dimensional grid

of {XT −1 , PT −1 } points one can solve for a one-dimensional vector of values of {xT −1 }.

This and a few other tricks turn the problem into one that can be solved with the

amounts of computer power that began to be widely available in the 1980s.7

The solution to the optimal consumption problem is depicted in Figure 1. The cash-

on-hand ratio x is on the horizontal axis. The optimal consumption ratio for a given

cash-on-hand ratio is on the vertical axis. The solid lines represent the consumption

rules for different time periods, showing how optimal consumption changes as the ratio

of cash-on-hand to labor income increases.

Consumption in the last period cT (x) coincides with the 45 degree line, indicating

11
consumption equal to cash-on-hand. For very low levels of x, consumption in the

second-to-last period cT −1 (x) is fairly close to the 45 degree line; the consumer spends

almost, but not quite, everything. This reflects the precautionary motive: Because

there is a chance the consumer will receive zero income in period T , she will never

spend all of her period-T − 1 resources because of the dire consequences of arriving

at T with nothing and then possibly receiving zero income. Note the contrast with

behavior at high levels of wealth; for example, at an xT −1 of around 10 the figure shows

cT −1 of a bit more than 5 – indicating that at this large level of wealth the consumer

divides remaining lifetime resources roughly evenly between the last two periods of life.

An important feature of this problem is that, if certain conditions hold (in partic-

ular, if consumers are ‘impatient’ in a sense to be described shortly), the successive

consumption rules cT (x), cT −1 (x), cT −2 (x), . . . , cT −n (x) will ‘converge’ as n grows large.

The meaning of convergence is most easily grasped visually: In Figure 1, the rules cT (x)

and cT −1 (x) are very far apart, while the rules cT −10 (x) and the converged consumption

rule c(x) (which can be thought of as cT −∞ (x)) are very close.

The importance of convergence can best be understood by contrasting it with the

alternative. Modigliani (1966) points out that in the certainty equivalent model, opti-

mal behavior is different at every different age, so that one cannot draw many general

lessons about consumption behavior from the rule for any particular age. In the model

solved here, however, behavior is essentially identical for all consumers more than 10

years from the end of life, so analysis of the converged consumption rule yields insights

12
about behavior of most agents in the economy.

What is required to generate convergence? Deaton (1991) and Carroll (1996) show

that the necessary condition is that consumers be impatient, in the sense that if there

were no uncertainty or liquidity constraints the consumer would choose to spend more

than her current income. Technically, the required condition is

(Rβ)1/ρ < G, (2)

where ρ is the coefficient of relative risk aversion and G is the income growth factor.

Consider the version of this equation where G = ρ = 1, so that consumers are

impatient if Rβ < 1. In this case, impatience depends directly on the whether the

reward to waiting, as determined by the interest rate factor R, is large enough to over-

come the utility cost to waiting, β. Positive income growth (G > 1) makes consumers

more impatient (in the sense of wanting to spend more than current income) because

forward-looking consumers with positive income growth will want to spend some of

their higher future income today. Finally, the exponent (1/ρ) on the Rβ term captures

the ‘intertemporal elasticity of substitution,’ which measures the extent to which the

consumer responds to the net incentives for reallocating consumption between periods.

The remainder of the paper will focus almost exclusively on implications of the

converged consumption function. It is natural to wonder, however, whether we should

expect these results to be useful in understanding the behavior of consumers whose

permanent income paths over the lifetime do not resemble the “constant growth at

13
rate G until death” specification used here. For instance, income can be predicted

to decline at retirement! However, Carroll (1997) shows that when a model like this

is solved with an empirically realistic pattern of income growth over the lifetime, the

consumption function resembles the ‘converged’ consumption function examined here

until roughly age 50. After 50, with retirement looming, the consumer begins saving

substantial amounts and behavior begins more and more to resemble that in the per-

fect foresight model. Thus, the results in the remainder of the paper based on the

converged consumption function are most appropriately represented as characterizing

the behavior of moderately impatient households up to about age 50.8

At present, three further observations about the converged consumption function

depicted in Figure 1 are important. (The general shape of the consumption function,

and the validity of the points made here, are robust to alternative assumptions about

parameter values, so long as consumers remain moderately impatient.)

First, the converged consumption function is everywhere well below the perfect fore-

sight solution (the dashed line). Since precautionary saving is defined as the amount

by which consumption falls as a consequence of uncertainty, the difference between the

converged c(x) and the dashed perfect-foresight line measures the extent of precaution-

ary saving. The precautionary effect is large here because under our baseline parameter

values, human wealth is quite large and therefore induces a lot of consumption by the

perfect-foresight consumers. In contrast, consumers with a precautionary motive are

unwilling to spend much on the basis of uncertain future labor income, so the large

14
value of human wealth has little effect on their current consumption.

The second important observation is that as x gets large, the slope of c(x) (which

is to say, the marginal propensity to consume) gets closer and closer to the slope of

the dashed perfect foresight line. That is, as wealth approaches infinity the marginal

propensity to consume approaches the perfect foresight MPC. This happens because as

wealth approaches infinity the proportion of future consumption that will be financed

out of uncertain labor income approaches zero, so the labor income uncertainty becomes

irrelevant to the consumption decision.

The final observation is that for periods before the last one the consumption function

lies everywhere below the 45-degree line; that is, consumers choose never to borrow

(which they would need to do in order to have c > x and to be above the 45-degree

line), even though no liquidity constraint was imposed in solving the problem.

This last result deserves explanation. As noted above, in the second-to-last period,

consumers will always choose to spend less than their cash-on-hand because of the risk

of zero income in the last period of life. If we know that in period T − 1 consumption

will be less than x, then that implies that in period T − 2 the consumer will always

behave in such a way to make sure that he arrives in T − 1 with positive assets, again

out of the fear of a zero-income event in T − 1. Similar logic goes through recursively

to any earlier period.

This mechanism for preventing borrowing may seem rather implausible, relying as

it does on the slight possibility of disastrous zero-income events. However, essentially

15
the same logic works as long as income has a well-defined lower bound. For example,

suppose the worst possible outcome were that income might fall to, say, 30 percent of

its permanent level. In this case the recursive logic outlined above would not prohibit

borrowing. But it would prevent the consumer from borrowing more than the amount

that could be repaid with certainty out of the lowest possible future income stream.

In this case, consumers would define their precautionary target in terms of the size of

their wealth holdings in excess of the lowest feasible level. The distinctive features of

the model discussed below would all go through, with the solitary difference that the

average level of wealth would be lower (perhaps even negative).

This logic provides the simplest intuition for a fundamental conclusion: The pre-

cautionary saving motive can generate behavior that is virtually indistinguishable from

that generated by a liquidity constraint,9 because the precautionary saving motive es-

sentially induces self-imposed reluctance to borrow (or borrow too much).

2 Implications

2.1 Concavity of the Consumption Function and Buffer Stock

Saving

Perhaps the most striking feature of the converged consumption function c(x) depicted

in figure 1 is that the marginal propensity to consume (the slope of the consumption

function) is much greater at low levels of cash-on-hand than at high levels. In other

16
words, the converged consumption function is strongly concave.10 Thus, the first intu-

itive result that comes out of the analysis is that, as Keynes (1935) argued long ago,

rich people spend a smaller proportion of any transitory shock to their income than do

poor people.

Carroll (1996) shows that concavity of the consumption function also implies that

impatient consumers will engage in ‘buffer-stock’ saving behavior. That is, there will

be some target level of the cash-on-hand ratio x∗ such that, if actual cash-on-hand is

greater than the target, impatience will outweigh prudence and wealth will fall, while if

cash-on-hand is below the target the precautionary saving motive will outweigh impa-

tience and the consumer will try to build wealth up back toward the target. As usual,

this result is something that Friedman grasped intuitively: He refers repeatedly to the

role of wealth as an ‘emergency reserve’ against uncertainty or a ‘balancing resource’;

indeed, Mayer (1972), p. 70 summarizes Friedman’s version of the PIH succinctly: ‘It

is basic to [Friedman’s] permanent income theory that households attempt to maximize

utility by using savings as a buffer against income fluctuations.’

Buffer-stock saving behavior is a qualitative implication of the model. In order to

determine the model’s quantitative implications (for example, what it predicts about

the average value of the MPC), it is necessary to simulate a population of consumers

behaving according to the converged consumption rule. The first row of Panel A of

Table 1 provides a variety of statistics about average behavior when consumers are

distributed according to the steady-state distribution generated by the baseline para-

17
metric assumptions. Columns two and three indicate that the mean and median of the

wealth ratio are both about 0.4, or equal to about five months’ worth of permanent

noncapital income (remember that the time unit is a year). The average marginal

propensity to consume is 0.33, in the ballpark of both empirical estimates and Fried-

man’s (1957) statement of his conception of the permanent income hypothesis, but a

long way from the approximately 0.04 implied by the perfect foresight model under

our baseline parameter values.

The second row of Panel A presents results under the assumption that household

noncapital income growth is 2 percent a year, rather than the baseline of 3 percent.

Lower income growth makes people more “patient,” in the sense that the contrast

between tomorrow’s and today’s income – and thus the temptation to borrow against

future income – is not as great. The table shows that greater patience leads to a higher

mean wealth ratio a lower average MPC.

The final row of Panel A presents results when predictable income growth is zero.11

With these extremely patient consumers, who cannot rely on future income gains at

all, average wealth is much higher, and the average MPC is only about 0.06, not much

greater than in the perfect foresight model.

These results confirm that if consumers are moderately impatient, their behavior in

the modern model with uncertainty resembles Friedman’s conception of the permanent

income hypothesis. Neither liquidity constraints nor myopia is necessary to generate

the high average marginal propensity to consume that has repeatedly been found in

18
empirical studies and that Friedman (1957) deemed consistent with his conception of

the permanent income hypothesis. Impatience plus uncertainty will do the trick.

The reason precautionary saving increases the MPC is because the precautionary

motive relaxes as the level of wealth rises. To put it another way, an extra unit of

cash-on-hand today means that one has a better ability to buffer consumption against

income shocks in the future, and so there is less need to depress consumption to build

up one’s precautionary assets. Thus, the decline in the intensity of the precautionary

motive as cash-on-hand rises allows consumption to rise faster than it would in the

absence of a precautionary motive – which is to say, the MPC out of cash-on-hand

(and therefore the MPC out of transitory shocks to income) is higher.

Recall that another difference between Friedman and the subsequent models was in

the rate at which consumers were assumed to discount future income. In the subsequent

models, the mean expectation of future labor income was discounted to the present

at a market interest rate (say, 4 percent). Friedman (1963) insisted that future labor

income was discounted at a rate of around 33 percent. (A substantial body of empirical

evidence confirms that the actual reaction of consumption to information about future

income is much smaller than the perfect foresignt and certainty equivalent models

imply; see Campbell and Deaton (1989); Viard (1993); Carroll (1994); and the large

literature that finds that saving responds much less than one-for-one to expected future

pension benefits (Samwick (1995)).)

We can examine this controversy in the new model by determining how average con-

19
sumption changes when expectations about the future path of income change. Suppose

we have a population of consumers who have received their period t income and are

distributed according to the steady-state distribution of xt that obtains under the base-

line parameter values. Now consider informing these consumers that henceforth growth

will be G = 1.02 rather than 1.03. It turns out that under the baseline parameter val-

ues, consumers react to the news of the change in income growth as though they are

discounting future noncapital income at a 39 percent rate - even higher than Fried-

man’s estimate of 33 percent!12 The reason for the high discount rate is that prudent

consumers know it would be unwise to spend today on the basis of future income that

might not actually materialize.

2.2 The Consumption Euler Equation

Robert Hall (1978) provided the impetus for a large empirical literature over the past

two decades by pointing out that in the certainty equivalent model, the predictable

change in consumption in a given period should be unrelated to any information that

the consumer possessed in earlier periods; consumption should follow a ‘random walk.’

The logic was simple: forward-looking consumers who want to smooth their consump-

tion should react immediately and fully to any information they possess about the fu-

ture, because if consumption doesn’t react fully and immediately, it will have to finish

reacting at some future date, which implies a failure to smooth consumption between

the present and that future date. Thus the only reason for a change in consumption is

20
the arrival of new, previously unknown information. So changes in consumption will

not be related to predictable changes in income, or indeed to any other predictable fac-

tor. The only exception to the unpredictability of consumption growth comes from the

interplay between tastes and opportunities represented by the discount factor and the

time preference factor: If interest rates are predictably high, consumption growth will

be predictably faster, and vice versa for the time preference rate. A similar analysis

shows that there is no relationship between the optimal rate of consumption growth

and the average rate of income growth in the perfect foresight model.

Because the formal statement of Hall’s result relied on the mathematical optimality

condition known as the Euler equation, tests of the kind Hall proposed became known as

Euler equation tests. Many papers in the subsequent literature (though not all) found

that consumption growth was strongly related to the predictable component of income

growth. This apparent violation of the Euler equation was typically considered evidence

either of myopia or of binding liquidity constraints, since the Euler equation does not

apply to consumers who are constrained; constrained consumers may set consumption

equal to income in every period, even if changes in income are predictable.

The fourth column of Panel A in Table 1 shows that the growth rate of aggregate

consumption for our simulated consumers is equal to the predictable underlying growth

rate of permanent income for all three values for the rate of income growth. This

obviously conflicts with the implication of the perfect foresight and certainty equivalent

models that consumption growth should be unrelated to predictable income growth.

21
 The resolution can be found in the analogy between precautionary saving and liq-

uidity constraints. The constrained consumer cannot borrow against future income to

finance current consumption; the consumer with a precautionary saving motive chooses

not to borrow for fear of the consequences of borrowing and then experiencing negative

shocks to income. The ultimate effect is the same: consumption growth can be strongly

tied to predictable income growth. Note also that, in both cases, the connection be-

tween consumption growth and predictable income growth arises only if consumers are

impatient enough; for patient consumers, consumption is less than income anyway, and

neither external nor self-imposed constraints have much effect.

Another way of understanding the equality of consumption growth and income

growth comes from thinking about the implications of target saving behavior. If con-

sumption growth were forever below income growth then eventually consumption would

constitute a vanishingly small proportion of income, and the wealth/income ratio would

grow without bound – a contradiction of the target saving proposition. Conversely, if

consumption growth were forever above income growth, then consumption would even-

tually exceed income by an arbitrarily large amount, sending the wealth/income ratio

toward negative infinity – again a violation of target saving.

Note further that the stability of wealth around the target level implies that the level

of consumption must generally be close to the level of total income, capital and labor.

Once again, the prediction of the modern buffer-stock model is close to Friedman’s

original idea that people set their consumption to approximately the average value of

22
expected income over the next few years, and not very close to the implication of the

perfect foresight or certainty equivalent models that consumption equals some small

MPC multiplied by total remaining lifetime wealth, human and nonhuman.

2.3 Precautionary Saving and Liquidity Constraints

The results of Panel A suggest that a model with precautionary saving and impatient

consumers can produce behavior that resembles behavior in a liquidity constrained

version of the perfect foresight or certainty equivalent models. This raises the question

of whether there are any important differences in behavior between constrained and

unconstrained consumers.

The simplest form of liquidity constraint is one in which all borrowing must be

collateralized by some marketable asset (as Friedman (1963) points out, human wealth

cannot serve as collateral because it cannot be seized and sold - unless anti-slavery

laws are repealed!) Appending a constraint requiring consumption to stay below total

market wealth, however, has no effect in the model specified above, since the possibility

of the dreaded zero-income-events means that these consumers never want to borrow

anyway. However, one could plausibly argue that in modern developed countries, the

social safety net prevents consumption from falling all the way to zero, mitigating the

impact of unemployment spells. To capture a social safety net, suppose now that the

worst possible event is an unemployment spell in which income drops to 50 percent of

its usual level, an event that occurs with probability p = 0.05 representing a 5 percent

23
unemployment rate. What does optimal behavior look like with such a social safety

net if consumers are strictly prohibited from borrowing against future labor income?

Panel B of Table 1 presents some answers. The mean and median amount of buffer-

stock wealth are both now around 0.25, or about two months’ worth less of income

than in the unconstrained case. Precautionary wealth is lower because the risk of zero

income events has now been replaced with a comparatively generous unemployment

insurance system. Note, however, that the average MPC is roughly the same as in

Panel A. Furthermore, the effect on the MPC of making consumers more patient (by

reducing income growth to G = 1.00) is also virtually identical to that in Panel A: for

patient consumers, the MPC drops to about 6 percent. A final result is that at any

given time about 7 percent of households have exactly zero wealth; these are households

who have recently experienced bad income shocks and have not had time to rebuild

their buffer stocks.

Of course, a complete inability to borrow is unrealistic. But Ludvigson (1996)

presents evidence that lenders do attempt to limit the ratio of the borrower’s debt to

income. Panel C therefore summarizes behavior in an economy in which lenders restrict

consumers’ debt to a maximum of 30 percent of their permanent noncapital income.

The effect of a constraint of this type is essentially just to shift the no-borrowing con-

sumption function and wealth distribution to the left by almost exactly 0.3. Note that

the steady-state average MPC essentially the same as when consumers were prohibited

from borrowing at all. This goes against the grain of intuition, since one might think

24
that consumers who can borrow should be better able to shield their consumption

against income shocks. But remember that precautionary motives are the only reason

these impatient consumers do any saving at all. The “buffering capacity” of a given

level of wealth depends on how much lower wealth could potentially be driven in the

case of a bad shock, so allowing borrowing just shifts the whole consumption function

and wealth distribution left, without changing steady-state consumption behavior.

Collectively, the results in Panels A through C of the table demonstrate that liq-

uidity constraints are neither necessary nor sufficient to generate a high MPC. What

is both necessary and sufficient is impatience, whether there are constraints or not.

(Another point these panels demonstrate is that if the entire population is impatient,

the model generates far less aggregate wealth than is observed empirically; see the

discussion of this problem in the ‘limitations’ section right before the conclusion).

The point that the average MPC depends on impatience rather than on the pres-

ence or absence of constraints means that many traditional tests of liquidity constraints

are questionable at best. For example, Campbell and Mankiw (1991) argue that differ-

ences across countries in the sensitivity of consumption growth to predictable income

growth may reflect differences in the degree of liquidity constraints, while Jappelli and

Pagano (1989) suggest that constraints may be stronger in countries in which consump-

tion growth exhibits excess sensitivity to lagged income growth. It is not clear that

either of these interpretations is valid. Instead, the warranted conclusion is probably

that countries in which consumption exhibits excess sensitivity to lagged or current

25
income may have more households who are more impatient, and consequently inhabit

the portion of the consumption function where the MPC is high, whether they are

formally constrained or not.

If empirical evidence on the MPC is not informative about the importance of liq-

uidity constraints, what kind of evidence would be informative? One example is work

by Gross and Souleles (2000), who have obtained a database containing comprehensive

credit information on a representative sample of consumers. They show that exoge-

nous increases in households’ credit limits result in a substantial increase in actual

total debt burdens; in fact, the observed behavior appears to be qualitatively similar

to the simulation results presented in Panels B and C of table 1, in the sense that the

data show that the debt load after an exogenous credit expansion appears to stabilize

at a point that provides roughly the same amount of unused credit capacity as before

the expansion in the credit line.

Another way to distinguish precautionary behavior from liquidity constraints is to

start with the point, noted above, that the wealth distribution under strict constraints

contains a mass of households at exactly zero wealth. Figure 2 presents the cumulative

distribution function for data from the 1995 US Survey of Consumer Finances on the

ratio of nonhousing wealth to permanent income for US consumers between the ages of

25 and 50 – the age range for which the baseline buffer-stock model has been claimed

as a plausible description of behavior.13

Although it is hard to see in the figure, there is indeed a small concentration of

26
households (about 2.5 percent of the population, as indicated in Table 1) at exactly the

zero-wealth point, and a total of about 10 percent have net worth in the range from

zero to two weeks’ worth (one paycheck) of their permanent income (not in table).

(‘Permanent income’ is defined, a la Friedman, as actual income for the subset of

households who said that their income in the survey year was ‘about normal.’)

However, about 20 percent of households in the figure actually have negative finan-

cial net worth (uncollateralized loans greater than total financial assets); these people

obviously have not been completely constrained from borrowing. Furthermore, a strict

prohibition against borrowing flies in the face of daily experience in modern America,

where even household pets receive unsolicited offers of credit cards (and sometimes

accept them! see Bennett (1999)).

For comparison to the empirical data, therefore, figure 3 presents the theoretical

cumulative distribution function (CDF) from a final version of the model in which

households are allowed to borrow, but only up to a maximum debt/income ratio of 30

percent and only at a real interest rate of 15 percent (which roughly matches typical

credit card interest rates). The qualitative shape of the theoretical CDF is a nice

match with the empirical CDF; indeed, Panel D of Table 1 shows that under baseline

parameters the proportion of households with negative net worth in the simulations

(about 30 percent) is actually larger than in the empirical data (20 percent). This

indicates, perhaps surprisingly, that for these moderately impatient consumers, the

optimal consumption plan involves borrowing (even at a 15 percent annual rate) for a

27
substantial fraction of the time.

However, one big problem for the model is evident from a closer look at the upper

part of the empirical CDF (Figure 2). Although the empirical median wealth/income

ratio, at about 0.3, is in the vicinity of the small values predicted by all the models in

Table 1 under baseline parameter values, the upper part of the empirical distribution

contains vastly more wealth than is implied by the model; Panel E of Table 1 indicates

that the mean value of the empirical wealth-to-permanent-income ratio w is much

greater than its median, indicating the strong skewness of the distribution. Thus,

while the presence of substantial numbers of impatient consumers may be essential for

reproducing the empirical finding of a high average marginal propensity to consume,

the presence of some patient consumers is also necessary if the model is to match

the overall amount of wealth in the US. Whether a life cycle version of the model

with patient and impatient consumers can match the entire distribution of wealth is

a matter of ongoing debate; my own view is that the model certainly cannot match

the behavior of the richest few percent in the distribution (unless a bequest motive is

added), but may be able to match much of the rest.14

3 Limitations

I have argued here that the modern version of the dynamically optimizing consump-

tion model is able to match many of the important features of the empirical data on

28
consumption and saving behavior. There are, however, several remaining reasons for

discomfort with the model.

One problem is the spectacular contrast between the sophisticated mathematical

apparatus required to solve the optimal consumption problem and the mathematical

imbecility of most actual consumers. We can turn, again, to Milton Friedman for a

potentially plausible justification for such mathematical modelling. Friedman (1953)

argued that repeated experience in attempting to solve difficult problems could build

good intuition about the right solution. His example was an experienced pool player

who does not know Newtonian mechanics, but has an excellent intuitive grasp of where

the balls will go when he hits them. This parable may sound convincing, but some

recent work I have done with Todd Allen (2001) suggests that it may sound more

convincing than it should. We examine how much experience it would take for a

consumer who does not know how to solve dynamic optimization problems to learn

nearly optimal consumption behavior by trial and error. Under our baseline setup,

we find that it takes about a million ‘years’ of model time to find a reasonably good

consumption rule by trial and error. This result may sound preposterous, but we

are fairly confident that our qualitative conclusion will hold up, because if there were

some trial-and-error method of finding optimal consumption behavior without a large

number of trials (and errors), such a method would also constitute a fundamental

breakthrough in numerical solution methods for dynamic programming problems. We

suspect that the total absence of trial-and-error methods from the literature on optimal

29
solution methods for dynamic optimization problems indicates that such methods are

very inefficient, even compared to the enormous computational demands of traditional

dynamic programming solution methods. We conclude by speculating that there may

be more hope of consumers finding reasonably good rules in a “social learning” context

in which one can benefit from the experience of others. However, even the social

learning model will probably take considerable time to converge on optimal behavior,

so this model provides no reason to suppose that consumers will react optimally in the

short- or medium-run to the introduction of new elements into their environment.

As an example of such a change in the consumption and savings environment,

consider the introduction of credit cards. In a trial-and-error economy, many consumers

would need to try out credit cards, discover that their heavy use can yield lower utility if

they lead to high interest payments, and communicate this information to others before

there would be any reason to expect the social use of credit cards to approximate their

optimizing use. This social learning process could take some time, and even the passage

of a recession or two.

There certainly seems to be strong evidence that many American households are

now using credit cards in nonoptimal ways. The optimal use of credit cards (at least

as implied by solving the final optimizing model discussed above) is as an emergency

reserve to be drawn on only rarely, in response to a particularly bad shock or series

of shocks. However, the median household with at least one credit card holds about

$7,000 in debt on all cards combined; that $7,000 is the balance on which interest is

30
paid, not just the transactions use (Gross and Souleles (2000)). Laibson, Repetto, and

Tobacman (1999) argue that this pattern results from time-inconsistent preferences

in which consumers have a powerful preference for immediate consumption. Their

approach is discussed further in the paper in this symposium.

Another set of empirical findings that are very difficult to reconcile with the mod-

ern model of consumption presented here comes in the relationship between saving and

income growth, either across countries or across households. A substantial empirical

literature has found that much and perhaps most of the strong positive correlation

between saving and growth across countries reflects causality from growth to saving

rather than the other way around (see Carroll, Overland, and Weil (2000) for a sum-

mary). This is problematic because the model implies that consumers expecting faster

growth should save less, not more (cf. the model simulations in Table 1). Carroll,

Overland, and Weil (2000) suggest that the puzzle can be explained by allowing for

habit formation in consumption preferences, but as yet, there is no consensus answer

to this puzzle.

A final problem for the standard model is its inability to explain household port-

folio choices. The “equity premium puzzle” over which so much ink has been spilled

(for a summary see Siegel and Thaler (1997)) remains a puzzle at the microeconomic

level, where standard models like the ones presented here imply that consumers should

hold almost 100 percent of their wealth in the stock market (for simulation results,

see, e.g., Fratantoni (1998), Cocco, Gomes, and Maenhout (1998), Gakidis (1998),

31
Hochguertel (1998), Bertaut and Haliassos (1997)).

4 Conclusion

We shall never cease from exploration

And the end of all our exploring

Will be to arrive where we started

And know the place for the first time.

- T.S. Eliot, “Four Quartets”

Few consumption researchers today would defend the perfect foresight or certainty

equivalent models as adequate representations either of the theoretical problem facing

consumers or of the actual behavior consumers engage in. Most would probably agree

that Milton Friedman’s original intuitive description of behavior was much closer to the

mark, at least for the median consumer. It is tempting therefore to dismiss most of the

work between Friedman (1957, 1963) and the new computational models of the 1980s

and ‘90s as a useless diversion. But a more appropriate view would be that solving and

testing those first formal models was an important step on the way to obtaining our

current deeper understanding of consumption theory, just as (in a much grander way)

the development of Newtonian physics was a necessary and important predecessor to

Einstein’s general theory.

Understanding of the quantitative implications of the new computational model

32
of consumption behavior is by no means complete. As techniques for solving and

simulating models of this kind disseminate, the coming decade promises to produce

a flood of interesting work that should define clearly the conditions under which ob-

served consumption, portfolio choice, and other behavior can or cannot be captured

by the computational rational optimizing model. Indeed, one purpose of this paper

is to encourage readers to join in this enterprise - a process that I hope will be made

considerably easier by the availability on the author’s website (see the address on the

first page) of a set of Mathematica programs capable of solving and simulating quite

general versions of the computational optimal consumption/saving problem described

in this paper.

33
References

Allen, Todd M., and Christopher D. Carroll (2001): “Individ-

ual Learning About Consumption,” Forthcoming, Macroeconomic Dynamics,

http://www.econ.jhu.edu/people/ccarroll/IndivLearningAboutC.pdf .

Bennett, Lennie (1999): “Platinum Pooch,” St. Petersburg Times.

Bertaut, Carol C., and Michael Haliassos (1997): “Precautionary portfolio

behavior from a life-cycle perspective,” Journal of Economic Dynamics And Control,

21(8-9), 1511–1542.

Bodkin, Ronald (1959): “Windfall Income and Consumption,” American Economic

Review, 49(4), 602–614.

Cagetti, Marco (1999): “Wealth Accumulation Over the Life Cycle and Precau-

tionary Savings,” Manuscript, University of Chicago.

Campbell, John Y., and Angus S. Deaton (1989): “Why Is Consumption So

Smooth?,” Review of Economic Studies, 56, 357–74.

Campbell, John Y., and N. Gregory Mankiw (1991): “The Response of Con-

sumption to Income: A Cross-Country Investigation,” European Economic Review,

35, 723–67.

34
Carroll, Christopher D. (1992): “The Buffer-Stock Theory of Saving: Some

Macroeconomic Evidence,” Brookings Papers on Economic Activity, 1992(2), 61–

156.

(1994): “How Does Future Income Affect Current Con-

sumption?,” The Quarterly Journal of Economics, CIX(1), 111–148,

http://www.econ.jhu.edu/people/ccarroll/howdoesfuture.pdf .

(1996): “Buffer Stock Saving: Some Theory,” Manuscript, Johns Hopkins

University, January, http://www.econ.jhu.edu/people/ccarroll/BufferTheory.pdf .

(1997): “Buffer-Stock Saving and the Life Cycle/Permanent In-

come Hypothesis,” Quarterly Journal of Economics, CXII(1), 1–56,

http://www.econ.jhu.edu/people/ccarroll/BSLCPIH.pdf .

(2000a): “Requiem for the Representative Consumer? Aggregate Implications

of Microeconomic Consumption Behavior,” American Economic Review, Papers and

Proceedings, pp. 110–115, http://www.econ.jhu.edu/people/ccarroll/RequiemFull.pdf,

Table 1 Data: http://www.econ.jhu.edu/people/ccarroll/requiemfiles.zip. .

(2000b): “Why Do the Rich Save So Much?,” in Does Atlas Shrug? The

Economic Consequences of Taxing the Rich, ed. by Joel B. Slemrod, pp. 463–485.

Harvard University Press, http://www.econ.jhu.edu/people/ccarroll/Why.pdf .

35
Carroll, Christopher D., and Wendy E. Dunn (1997): “Unem-

ployment Expectations, Jumping (S,s) Triggers, and Household Balance

Sheets,” in NBER Macroeconomics Annual, 1997, ed. by Benjamin S.

Bernanke, and Julio Rotemberg, pp. 165–229. MIT Press, Cambridge,

MA, http://www.econ.jhu.edu/people/ccarroll/macroann.pdf; Methodological Appendix:

http://www.econ.jhu.edu/people/ccarroll/methods3.pdf; Empirical Results and Simula-

tion Programs: http://www.econ.jhu.edu/people/ccarroll/cdfiles.html; .

Carroll, Christopher D., and Miles S. Kimball (1996): “On the

Concavity of the Consumption Function,” Econometrica, 64(4), 981–992,

http://www.econ.jhu.edu/people/ccarroll/concavity.pdf .

(2001): “Liquidity Constraints and Precautionary Saving,” Manuscript, Johns

Hopkins University.

Carroll, Christopher D., Jody R. Overland, and David N. Weil (2000):

“Saving and Growth with Habit Formation,” American Economic Review, 90(3),

341–355, http://www.econ.jhu.edu/people/ccarroll/AERHabits.pdf .

Cocco, Joao, Francisco J. Gomes, and Pascal J. Maenhout (1998): “Con-

sumption and Portfolio Choice Over the Life Cycle,” Manuscript, Harvard Univer-

sity.

36
Deaton, Angus S. (1991): “Saving and Liquidity Constraints,” Econometrica, 59,

1221–1248.

Dynan, Karen E., Jonathan S. Skinner, and Stephen P. Zeldes (1996):

“Do the Rich Save More?,” Manuscript, Board of Governors of the Federal Reserve

System.

Engen, Eric, William Gale, and Cori Uccello (1999): “The Adequacy of

Retirement Saving,” Brookings Papers on Economic Activity, 1999(2).

Fratantoni, Michael C. (1998): “Income Uncertainty and the Equity Premium

Puzzle,” Manuscript, Johns Hopkins University.

Friedman, Milton A. (1953): Essays in Positive Economics. University of Chicago

Press, Chicago.

(1957): A Theory of the Consumption Function. Princeton University Press.

(1963): “Windfalls, the ‘Horizon,’ and Related Concepts in the Permanent

Income Hypothesis,” in Measurement in Economics, ed. by Carl Christ, et al., pp.

1–28. Stanford University Press.

Gakidis, Haralobos E. (1998): “Stocks for the Old? Earnings Uncertainty and

Life-Cycle Portfolio Choice,” Manuscript, Massachusetts Institute of Technology.

Gourinchas, Pierre-Olivier, and Jonathan Parker (1999): “Consumption

Over the Life Cycle,” Manuscript, Princeton University.

37
Gross, David B., and Nicholas S. Souleles (2000): “Consumer Response to

Credit Supp[ly: Evidence from Credit Card Data,” Manuscript, University of Penn-

sylvania.

Hall, Robert E. (1978): “Stochastic Implications of the Life-Cycle/Permanent In-

come Hypothesis: Theory and Evidence,” Journal of Political Economy, 96, 971–87.

Hall, Robert E., and Frederic Mishkin (1982): “The Sensitivity of Consump-

tion to Transitory Income: Evidence from PSID Households,” Econometrica, L,

461–81.

Hochguertel, Stefan (1998): “A Buffer Stock Model with Portfolio Choice: Impli-

cations of Income Risk and Liquidity Constraints,” Manuscript, Uppsala University.

Huggett, Mark (1996): “Wealth Distribution in Life Cycle Economies,” Journal of

Monetary Economics, 38(3), 469–494.

Keynes, John Maynard (1935): The General Theory of Employment, Interest, and

Money. Harvest/HBJ, San Diego, New York, London, 1964.

Kreinin, Mordecai E. (1961): “Windfall Income and Consumption: Additional

Evidence,” American Economic Review, 51, 388–390.

Krusell, Per, and Anthony A. Smith (1998): “Income and Wealth Heterogeneity

in the Macroeconomy,” Journal of Political Economy, 106(5), 867–896.

38
Laibson, David, Andrea Repetto, and Jeremy Tobacman (1999): “A Debt

Puzzle,” Manuscript, Harvard University.

Landsberger, M. (1966): “Windfall Income and Consumption: Comment,” Ameri-

can Economic Review, 56, 534–540.

Ludvigson, Sydney (1996): “Consumption and Credit: A Model of Time-Varying

Liquidity Constraints,” Ph.D. thesis, Princeton University.

Mayer, Thomas (1972): Permanent Income, Wealth, and Consumption. University

of California Press, Berkeley.

Merton, Robert C. (1969): “Lifetime Portfolio Selection under Uncertainty: The

Continuous Time Case,” Review of Economics and Statistics, 50, 247–257.

Modigliani, Franco (1966): “The Life Cycle Hypothesis of Saving, the Demand for

Wealth, and the Supply of Capital,” Social Research, 33, 160–217.

Modigliani, Franco, and Richard Brumberg (1954): “Utility Analysis and

the Consumption Function: An Interpretation of Cross-Section Data,” in Post-

Keynesian Economics, ed. by Kenneth K. Kurihara, pp. 388–436. Rutgers University

Press, New Brunswick, N.J.

Quadrini, Vincenzo, and José-Víctor Ríos-Rull (1997): “Models of the Dis-

tribution of Wealth,” Manuscript, University of Pennsylvania.

39
Samuelson, Paul A. (1969): “Lifetime Portfolio Selection By Dynamic Stochastic

Programming,” Review of Economics and Statistics, 51, 239–46.

Samwick, Andrew A. (1995): “The Limited Offset Between Pension Wealth and

Other Private Wealth: Implications of Buffer-Stock Saving,” Manuscript, Depart-

ment of Economics, Dartmouth College.

Siegel, Jeremy J., and Richard H. Thaler (1997): “Anomalies: The Equity

Premium Puzzle,” Journal of Economic Perspectives, 11(1), 191–200.

Tullio Jappelli, Marco Pagano (1989): “Consumption and Capital Market Im-

perfections: An International Comparison,” The American Economic Review, 79(5),

1088–1105.

Viard, Alan (1993): “The Productivity Slowdown and the Savings Shortfall: A

Challenge to the Permanent Income Hypothesis,” Economic Inquiry, 31, 549–564.

Zeldes, Stephen P. (1984): “Optimal Consumption with Stochastic Income,” Ph.D.

thesis, Massachusetts Institute of Technology.

40
Footnotes

1
The uncertainty considered here is explicitly labor income uncertainty. Samuel-

son (1969) and Merton (1969) found explicit solutions long ago in the case where there

is rate-of-return uncertainty but no labor income uncertainty, and showed that rate-of-

return uncertainty does not change behavior much compared to the perfect-foresight

model.

2
For a rigorous analysis of the relationship between constraints and precautionary

behavior, see Carroll and Kimball (2001).

3
My definitions of ‘transitory’ and ‘permanent’ shocks (spelled out explicitly in the

next subsection) correspond to usage in much of the modern consumption literature,

but differ from Friedman’s (1957) usage. In fact, Friedman (1957) actually states that

the MPC out of ‘transitory income shocks’ is zero, but Friedman (1963) was very

clear that in his conception of the PIH, first-year consumption out of windfalls was

about 0.33. The reconcilation is that such windfalls were not ‘transitory’ shocks in

Friedman’s terminology. Terminology aside, Friedman’s quantitative predictions for

how consumption should change, for example in response to a windfall, are clear,

41
so I will simply translate the Friedman model’s predictions into modern terminology

without further remark, e.g. by stating that Friedman’s model implies that the MPC

out of (my definition of) transitory shocks is a third.

4
The evidence included aggregate time series regressions of consumption on current

and lagged income; a comparison of the income elasticity of consumption in microdata

with the times-series correlation properties of household income; an examination of

time series paterns in microdata on saving rates and income; and several other tests.

5
For an excellent summary of these studies by Bodkin (1959), Kreinin (1961), Lands-

berger (1966), and others see Mayer (1972).

6
This choice of ρ implies that a consumer would be indifferent between consuming

$66,666 with certainty or consuming $50,000 with probability 1/2 and $100,000 with

probability 1/2. For ρ = 0, the consumer is not risk averse at all and would be

indifferent between $75,000 with certainty and $50,000 with probability .5 and $100,000

with probability .5. For ρ = ∞, the consumer is infinitely risk averse, and would choose

$50,000.01 with certainty over equal probabilities of $50,000 and $100,000.

7
For more details of some of these tricks, see the lecture notes on solution methods

42
for dynamic optimization problems on the author’s website.

8
Recent work by Gourinchas and Parker (1999) finds the switchpoint to be between

40 and 45 rather than 50, but Cagetti’s (1999) similar work suggests a later switching

age.

9
In fact, Carroll and Kimball (2001) show that as the probability of the zero-income

events approaches zero, behavior in the model with zero-income events becomes math-

ematically identical to behavior in the liquidity-constrained model.

10
Carroll and Kimball (1996) provide a proof that uncertainty induces a concave

consumption function for a very broad class of utility functions, including the constant

relative risk aversion form used here.

11
In this case the consumer is on the edge of failing the impatience condition (but

the condition does hold because (Rβ)1/ρ = 0.9992 < 1.00 under the baseline values for

{R, β, ρ} = {1.04, 0.96, 2}).

12
The procedure for calculating an average ‘effective’ interest rate is as follows. First,

determine what aggregate consumption would be in period t if consumers continued to

43
expect G = 1.03; call the result Ct.03 . Next, find the converged consumption rule under

the expectation that G = 1.02, and use it to determine how much consumption would

be done if consumers’ expectations were suddenly switched to G = 1.02 permanently;

call that result Ct.02 . Finally, find the value of the interest factor R such that, in the

perfect foresight model, if growth expectations changed from G = 1.03 to G = 1.02

then consumption would change by Ct.03 − Ct.02 . Unfortunately, the answer that one

gets from this methodology for the “effective” interest rate depends very much on how

the change in income is distributed over time, its stochastic properties, the level of

current wealth, and all of the other parameters of the model.

13
Housing and vehicle wealth have been excluded on the grounds that the model does

not pretend to capture the complexities associated with durable goods investment. See

Carroll and Dunn (1997) for simulation results showing that even when durable goods

are added to the model, buffer stock saving behavior emerges with respect to liquid

asset holdings.

14
See Huggett (1996), Dynan Skinner and Zeldes (1996), Quadrini and Ríos-Rull (1997),

Engen, Gale, and Uccello (1999), and Carroll (2000b) for several perspectives on this

44
question. For general equilibrium macro models which attempt to match both mi-

cro and macro data using mixed populations of patient and impatient consumers, see

Krusell and Smith (1998) and Carroll (2000a).

45
 cHxL
7.
¨ cT HxL= 45 Degree Line
6.
cT-1 HxL
5.
cPF HxL
Ø
4.

3.
cT-5 HxL
2. cT-10 HxL
cHxL

1.

x
2. 4. 6. 8. 10. 12.
Figure 1: Convergence of Consumption Functions cT −n (x) as n Rises

46
 1

.8

.6

.4

.2

0
-1 0 1 2 3
nwXhOy

Figure 2: Empirical CDF of Ratio of Net Worth to Permanent Income, 1995 SCF

47
CDFHwL
1.0

0.8

0.6

0.4

0.2

w
-0.25 0.00 0.25 0.50 0.75 1.00 1.25 1.50
Figure 3: Steady-State Wealth Distribution with Credit Card Borrowing

48
 Table 1: Steady-State Statistics For Alternative Consumption Models

Income Aggregate
Growth Mean Median Consumption Mean Frac With Frac With
Factor w w Growth MPC w<0 w=0
Panel A. Baseline Model, No Constraints
G=1.03 0.43 0.40 1.030 0.330 0.000 0.000
G=1.02 0.52 0.48 1.020 0.276 0.000 0.000
G=1.00 2.26 2.06 1.000 0.064 0.000 0.000
Panel B. Strict Liquidity Constraints
G=1.03 0.28 0.24 1.030 0.361 0.000 0.070
G=1.02 0.36 0.32 1.020 0.301 0.000 0.051
G=1.00 2.28 2.06 1.000 0.065 0.000 0.000
Panel C. Borrowing Up To 0.3 Allowed
G=1.03 −0.03 −0.06 1.030 0.361 0.611 0.000
G=1.02 0.06 0.01 1.020 0.299 0.478 0.000
G=1.00 1.94 1.71 1.000 0.064 0.023 0.000
Panel D. Borrowing Up to 0.3 at R = 1.15 Allowed
G=1.03 0.11 0.07 1.030 0.327 0.320 0.058
G=1.02 0.21 0.16 1.020 0.274 0.210 0.046
G=1.00 2.11 1.89 1.000 0.064 0.007 0.002
Panel E. Statistics from the 1995 SCF
– 1.02 0.29 – – 0.205 0.025
Notes: Results in Panels A through D reflect calculations by the author using simulation programs
available at the author’s website, http://www.econ.jhu.edu/people/carroll/ccaroll.html. In Panel A,
no constraint is imposed, but income can fall to zero, which prevents consumers from borrowing. In
Panels B through D, the worst possible event is for income to fall to half of permanent income. For
comparison, Panel E presents the mean and median values of the ratio of nonhousing wealth to
permanent income from the 1995 Survey of Consumer Finances for non-self-employed households
whose head was aged 25-50; the measure of permanent income is actual measured household income
for households who reported that their income over the past year was ‘about normal’, and whose
reported income was at least $5000; other households are dropped. The program that generates
these statistics (and figure 2) is also available at the author’s website.