Preliminary and Incomplete – Please Do Not Quote
_____________________________________________________________________________________
Abstract
Krusell and Smith (1998) showed that it is possible to construct rational
expectations macroeconomic models with serious microfoundations. We argue
that three modifications to their framework are required to fulfill its promise.
First, we replace their assumption about household income dynamics with a
process that matches microeconomic data. Second, our agents have finite
lifetimes a la Blanchard (1985), which has both substantive and technical
benefits. Finally, we calibrate heterogeneity in time preference rates so
that the model matches the observed degree of inequality in the wealth
distribution. Our model has substantially different, and considerably more
plausible, implications for macroeconomic questions like the aggregate
marginal propensity to consume out of an economic ‘stimulus’ program.
Microfoundations, Wealth Inequality, Marginal Propensity to Consume
D12, D31, D91, E21
PDF: | http://www.econ2.jhu.edu/people/ccarroll/papers/BSinKS.pdf |
Web: | http://www.econ2.jhu.edu/people/ccarroll/papers/BSinKS/ |
Archive: | http://www.econ2.jhu.edu/people/ccarroll/papers/BSinKS.zip |
1Carroll: Department of Economics, Johns Hopkins University, Baltimore, MD, http://www.econ2.jhu.edu/people/ccarroll/, ccarroll@jhu.edu 2Slacalek: European Central Bank, Frankfurt am Main, Germany, http://www.slacalek.com/, jiri.slacalek@ecb.europa.eu 3Tokuoka: International Monetary Fund, Washington, DC, ktokuoka@imf.org.
Macroeconomists have sought credible microfoundations since the dawn of our discipline. Keynes, his critics, and subsequent generations through Lucas (1976) and beyond have agreed on this, if little else.
Since Keynes’s time, consumption modeling has been a battleground between two microfoundational camps. ‘Bottom up’ modelers (e.g. Modigliani and Brumberg (1954); Friedman (1957)) drew wisdom from microeconomic data and argued that macro models should be constructed by aggregation from microeconomic models that matched robust micro facts. ‘Top down’ modelers (e.g., Samuelson (1958); Diamond (1965); Hall (1978)) treated aggregate consumption as reflecting the optimizing decisions of representative agents; with only one such agent (or, at most, one per generation), these models had ‘microfoundations’ under a generous interpretation of the word.
The tractability of representative agent models has made them appealing for business cycle analysis. But such models have never been easy to reconcile with either macroeconomic2 or microeconomic3 evidence on consumption dynamics, nor with microeconomic theory which implies that people who differ from each other (in age, preferences, wealth, liquidity constraints, taxes, and other dimensions) should respond differently to any given shock. If any of these differences matter (and it is hard to see how they could not),4 the aggregate size of a shock is not a sufficient statistic to calculate the aggregate response; information about how the shock is distributed is indispensable.
Bottom-up models, however, also have their problems. Even judged by a sympathetic standard that asks how well they can match measured wealth heterogeneity, bottom-up models have not been as successful as their champions might have initially hoped. For example, bottom-up models calibrated to match the wealth holdings of the median household generally fail to match the large size of the aggregate capital stock, because they seriously underpredict the upper parts of the wealth distribution (Carroll (2000b); Cagetti (2003)). Alternatively, models calibrated to match the aggregate level of wealth greatly overpredict wealth at the median (Hubbard, Skinner, and Zeldes (1994); Carroll (2000b)). A further problem is that (at least until Krusell and Smith (1998)) there has been no common answer to the question of how to analyze systematic macroeconomic fluctuations (business cycles) in bottom-up models.
This paper aims to reconcile the camps. We construct a workhorse model that answers the main objections to both kinds of models by making three modifications to the well-known Krusell–Smith (‘KS’) framework.5 First, we replace KS’s highly stylized assumptions about the nature of idiosyncratic income shocks with a microeconomic labor income process that captures the essentials of the empirical consensus from the labor economics literature about actual income dynamics in micro data (with credibly calibrated transitory and permanent shocks).6 Second, agents in our model have finite lifetimes a la Blanchard (1985), permitting a kind of primitive life cycle analysis and also solving some technical problems created by the incorporation of permanent shocks. Finally, we obtain a necessary extra boost to wealth inequality by calibrating a simple measure of heterogeneity in ‘impatience.’7
The resulting framework differs sharply from the benchmark KS model in its
implications for important microeconomic and macroeconomic questions. A
timely macroeconomic example is the response of aggregate consumption to an
‘economic stimulus payment,’ interpreted here as a one-time lump sum transfer to
households. In response to a $1-per-capita payment, the baseline version of the KS
model implies that the annual marginal propensity to consume (MPC) is about
,8
almost irrespective of how the cash is distributed across households. In contrast,
a version of our model that matches the distribution of liquid financial wealth
implies that if the entire tax cut were directed at households in the bottom half
of the liquid financial-wealth-to-income distribution, the MPC would be
, which counts as a big improvement in realism, given the vast
body of microeconomic evidence that consistently finds MPCs much
greater than the 3–5 percent figure that characterizes representative agent
models.9
Furthermore, the model’s differences with the representative agent framework are
not peculiar to unusual events like a stimulus payment; to the extent that
different kinds of macroeconomic shocks tend systematically to be differently
distributed across the population (for example, labor income shocks may
affect a less wealthy set of households than capital income shocks), this
improvement in realism may also matter for general questions of macroeconomic
dynamics.
Section 2 of the paper begins building the model’s structure by adding microeconomic modeling elements to a benchmark representative agent model. Using this model (without macroeconomic dynamics), the section closes by estimating the degree of heterogeneity in impatience necessary to match the degree of inequality in the U.S. wealth distribution; we find that relatively small differences in impatience substantially affect the model’s fit to the wealth data. Section 3 builds up the full version of the model by adding aggregate shocks of the KS type, and presents detailed comparisons of our model with theirs. Section 4 further improves the model by introducing an aggregate income process that is analytically simpler than the KS ‘toy’ aggregate process, that we believe is more empirically plausible as well, and that simplifies model solution and simulation considerably. We offer this final, simpler version of the model as our preferred jumping-off point for future macroeconomic research.
To establish notation and a transparent benchmark, we begin by briefly sketching a standard perfect foresight representative agent model.
The aggregate production function is
![]() | (1) |
where is aggregate productivity in period
,
is capital,
is time
worked per employee, and
is employment. The representative agent’s goal is
to maximize discounted utility from consumption
![]() |
for a CRRA utility function
.10
The representative agent’s state at the time of the consumption decision is
defined by two variables:
is market resources, and
is aggregate
productivity.
The transition process for is broken up, for clarity of analysis and
consistency with later notation, into three steps. Assets at the end of the period
are market resources minus consumption, equal to
![]() |
while next period’s capital is determined from this period’s assets via
![]() |
The final step can be conceived as the transition from the beginning of period
when capital has not yet been used to produce output, to the middle of
that period, when output has been produced and incorporated into resources but
has not yet been consumed:
After normalizing by the productivity factor
,13
the representative agent’s problem is
Notes: The models are calibrated at the quarterly frequency, and the steady state values are calculated on a quarterly basis.
Except where otherwise noted, our parametric assumptions match
those of the papers in the special issue of the Journal of Economic
Dynamics and Control (2010, Volume 34, Issue 1, edited by den Haan,
Judd, and Julliard) devoted to comparing solution methods for the KS
model (the parameters are reproduced for convenience in the top panel of
Table 1).14
The model is calibrated at the quarterly frequency. When aggregate shocks are
shut down ( and
), the model has a steady-state solution with a
constant ratio of capital to output and constant (gross) interest and wage factors,
which we write without time subscript as
and
and which are reflected in
Table 1.15
Henceforth, we refer to the version of the model solved by the papers in the
special JEDC volume as the ‘KS-JEDC’ model, while we call the original KS
model solved in Krusell and Smith (1998) ‘KS-Orig’ model. (The only effective
difference between the two is the introduction (for realism) of unemployment
insurance in the KS-JEDC version, which does not matter much for any substantive
results.16 17 )
For our purposes, the principal conclusion of the large literature on microeconomic
labor income dynamics is that household income can be reasonably well
described as follows. The idiosyncratic permanent component of labor income
evolves according to
After taking logarithms, this income process is strikingly similar to Friedman (1957)’s characterization of income as having permanent and transitory components. Because this process has been used widely in the literature on buffer stock saving, and though similar to Friedman’s formulation is not identical to it, we henceforth refer to it as the Friedman/Buffer Stock (or ‘FBS’) process.18
Table 2 summarizes the annual variances of log permanent shocks () and
log transitory shocks (
) estimated by a selection of papers from the extensive
literature.19
Some authors have used a process of this kind to describe the labor income or
wage process for an individual worker (top panel) while others have used it to
describe the process for overall household income (bottom panel); it seems to
work reasonably well in both cases (though, obviously, with different estimates of
the variances). (Recent work by Sabelhaus and Song (2010) using newly
available data from Social Security earnings files finds that the variances of both
transitory and permanent shocks have declined during the “Great Moderation”
period at all ages; they also find distinct life cycle patterns of shocks by age, with
young people experiencing higher levels of both kinds of shocks than the
middle-aged).
The second-to-last line of the table shows what labor economists would have found,
when estimating a process like the one above, if the empirical data were generated by
households who experienced an income process like the one assumed by the KS-JEDC
model.20
This row of the table makes our point forcefully: The empirical procedures that
have actually been applied to empirical micro data, if used to measure the
income process households experience in a KS economy, would have
produced estimates of and
that are orders of magnitude
different from what the actual empirical literature finds in actual data.
This discrepancy naturally makes one wonder whether the KS-JEDC
model’s well-known difficulty in matching the degree of wealth inequality is
largely explained by its highly unrealistic assumption about the income
process.21
Notes: MaCurdy (1982) did not explicitly separate
and
, but we have extracted
and
as
implications of statistics that his paper reports. First, we calculate
and
using his estimate (we set
). Then, following Carroll and Samwick (1997) we
obtain the values of
and
which can match these statistics, assuming that the income
process is
and
(i.e., we solve
and
).
Moffitt and Gottschalk (1995) estimated the income process
with random walk plus ARMA. Using income draws generated by their estimated process and following Carroll
and Samwick (1997), we have estimated the variances under the assumption that these income draws were
produced by the process
where
.
Meghir and Pistaferri (2004), Jensen and
Shore (2008), Hryshko (2010), and Blundell, Pistaferri, and Preston (2008) assume that the transitory
component is serially correlated (an MA process), and report the variance of a subelement of the transitory
component. For example, Meghir and Pistaferri (2004) and Blundell, Pistaferri, and Preston (2008) assume
an MA(1) process
and obtain estimates
=
and
, respectively.
for these four articles reported in this table are calculated by
using their estimates.
One might wish to use the FBS income process specified in subsection 2.2 as a complete characterization of household income dynamics, but that idea has a problem: Since each household accumulates a permanent shock in every period, the cross-sectional distribution of idiosyncratic permanent income becomes wider and wider indefinitely as the simulation progresses; that is, there is no ergodic distribution of permanent income in the population.
This problem and several others can be addressed by assuming that the
model’s agents have finite lifetimes a la Blanchard (1985). Death follows a
Poisson process, so that every agent alive at date has an equal probability
of dying before the beginning of period
. (The probability of NOT dying is
the cancelation of the probability of dying:
). Households engage in a
Blanchardian mutual insurance scheme: Survivors share the estates of those who
die. Assuming a zero profit condition for the insurance industry, the
insurance scheme’s ultimate effect is simply to boost the rate of return
(for survivors) by an amount exactly corresponding to the mortality
rate.
In order to maintain a constant population (of mass one, uniformly distributed
on the unit interval), we assume that dying households are replaced by an
equal number of newborns; we write the population-mean operator as
. Newborns, we assume, begin life with a level of idiosyncratic
permanent income equal to the mean level of idiosyncratic permanent income in
the population as a whole. Conveniently, our definition of the permanent shock
implies that in a large population, mean idiosyncratic permanent income
will remain fixed at
forever, while the mean of
is given
by22
Of course for all of this to be valid, it is necessary to impose the parametric
restriction (a requirement that does not do violence to the data, as
we shall see). Intuitively, the requirement is that, among surviving consumers,
income does not spread out so quickly as to overwhelm the compression of the
permanent income distribution that arises because of the equalizing force of
death and replacement.
Since our goal here is to produce a realistic distribution of permanent
income across the members of the (simulated) population, we measure the
empirical distribution of permanent income in the cross section using
data from the Survey of Consumer Finances (SCF), which conveniently
includes a question asking respondents whether their income in the survey
year was about ‘normal’ for them, and if not, asks the level of ‘normal’
income.23
This corresponds well with our (and Friedman (1957)’s) definition of permanent
income (and Kennickell (1995) shows that the answers people give to this
question can be reasonably interpreted as reflecting their perceptions of their
permanent income), so we calculate the variance of
among such
households.24
The results from this exercise are reported in Table 3 (with a final
row that makes the point that both the KS-Orig and KS-JEDC models
assume that permanent shocks did not exist). Substituting these estimates
for into (7) and (8), we obtain estimates of the variance of
.
Reassuringly, we can interpret the variances of
thus obtained as
being easily in the range of the estimated variances of
in
Table 2.25
Such a correspondence, across two quite different methods of measurement,
suggests there is considerable robustness to the measurement of the size of
permanent shocks. (Below, we will choose a calibration for
that is in the
middle range of estimates from either method.)
We now examine how wealth would be distributed in the steady-state equilibrium of an economy with wage rates and interest rates fixed at the steady state values calibrated in Table 1 of subsection 2.1, an income process like the one described in subsection 2.2, and finite lifetimes per subsection 2.3.
The process of noncapital income of each household follows
where Following the assumptions in the JEDC volume, the distribution of is:
The decision problem for the household in period can be written using
normalized variables; the consumer’s objective is to choose a series of
consumption functions
between now and the end of the horizon that satisfy:
Since households die with a constant probability between periods, the effective
discount factor is
(in (13)); the effective interest rate is
(combining (14)
and (15)).27
Aside from heterogeneity in impatience (introduced below), three parameters
characterize our modifications to the KS-JEDC model: ,
, and
.
implies the average length of working life is
quarters
=
years (dating from entry into the labor force at, say, age 25). The
variance of log transitory income shocks
is the value advocated in
Carroll (1992) (based on the Panel Study of Income Dynamics (PSID)
data),28
as is
(but note that this value also matches the median value in
Table 3).29
30
Other parameter values (
,
,
, and
) are from the JEDC volume
(Table 1).
The one remaining unspecified parameter is the time preference factor. As a preliminary theoretical consideration, note that Carroll (2011) (generalizing Deaton (1991) and Bewley (1977)) has shown that models of this kind do not have a well-defined solution unless the condition holds:
where Carroll (2011) dubs this inequality the ‘Growth Impatience Condition’ because
it guarantees that consumers are sufficiently impatient to prevent the indefinite
increase in the ratio of net worth to permanent income when income is growing
(see also Szeidl (2006)). This condition is an amalgam of the pure time
preference factor, expected growth, the relative risk aversion coefficient,
and the real interest factor. Thus, a consumer can be ‘impatient’ in the
required sense even if , so long as expected income growth is
positive.31
We begin by searching for the time preference factor such that if all households
had an identical
the steady-state value of the capital-to-output ratio (
)
would match the value that characterized the steady-state of the perfect foresight
model.32
turns out to be
(recall that this is at a quarterly, not an annual,
rate).
We now ask whether this model with realistically calibrated income and finite
lifetimes (henceforth, the model is referred to as the ‘-Point’ model) can
reproduce the degree of wealth inequality evident in the micro data. An
improvement in the model’s ability to match the data is to be expected, since in
buffer stock models agents strive to achieve a target ratio of wealth to permanent
income. By assuming no dispersion in the level of permanent income across
households, KS’s income process disables a potentially vital explanation for
variation in the level of target wealth (and, therefore, on average, actual wealth)
across households.
Notes:
.
, which implies
.
The results are from Krusell and Smith (1998) who solved the
models with aggregate shocks.
U.S. data is the SCF reported in Castaneda, Diaz-Gimenez, and
Rios-Rull (2003). Bold quantiles are targeted.
Table 4 shows that compared to the distribution of net worth
implied by our solution of the KS-JEDC model solved without an
aggregate shock (or the results of the KS-Orig model from Krusell and
Smith (1998)),33
our -Point model does indeed yield a substantial improvement
(compare the first, third and fourth columns to the last
column).34
For example, in our
-Point model, the fraction of total net worth held by the
top 1 percent is about 10 percent, while the corresponding statistic is
only 3 percent in our solution of the KS-JEDC model (or the KS-Orig
model).
The KS-JEDC model’s failure to match the wealth distribution is not confined
to the top. In fact, perhaps a bigger problem is that the model generates a
distribution of wealth in which most households’ wealth levels are not very far
from the wealth target of a representative agent in the perfect foresight
version of the model. For example, in steady state about percent
of all households in the KS-JEDC model have net worth between 0.5
times mean net worth and 1.5 times mean net worth; in the SCF data
from 1992–2004, the corresponding fraction ranges from only 20 to 25
percent.
But while the -Point model fits the data better than the original KS model,
it still falls short of matching the empirical degree of wealth inequality. The
proportion of net worth held by households in the top 1 percent of the
distribution is three times smaller in the model than in the data (compare the
first and last columns in the table). This failure reflects the fact that, empirically,
the distribution of wealth is considerably more unequal than the distribution of
permanent income.
As the simplest method to address this defect, we introduce heterogeneity in impatience: Each household is now assumed to have an idiosyncratic (but fixed) time preference factor.35 We do not think of this assumption as only capturing actual variation in pure rates of time preference across people (though such variation surely exists). Instead, we view discount-factor heterogeneity as a shortcut that captures the essential consequences of many other kinds of heterogeneity (e.g., heterogeneity in age, income growth expectations, investment opportunities, tax schedules) as well. To be more concrete, take the example of age. A robust pattern in most countries is that income grows much faster for young people than for older people. According to (16), young people should therefore tend to act, financially, in a more ‘impatient’ fashion than older people. In particular, we should expect them to have lower target wealth-to-income ratios. Thus, what we are capturing by allowing heterogeneity in time preference factors is probably also some portion of the difference in behavior that (in truth) reflects differences in age instead of in time preference factors, and that would be introduced into the model if we had a more complex specification of the life cycle that allowed for different growth rates for households of different ages.36
One way of gauging a model’s predictions for wealth inequality is to ask how
well it is able to match the proportion of total net worth held by the
wealthiest ,
,
, and
percent of the population. Because
these statistics have been targeted by other papers (e.g., Castaneda,
Diaz-Gimenez, and Rios-Rull (2003)), we adopt a goal of matching
them.37
We replace the assumption that all households have the same time
preference factor with an assumption that, for some , time preference
factors are distributed uniformly in the population between
and
(for this reason, the model is referred to as the ‘
-Dist’ model
below). Then, using simulations, we search for the values of
and
for which the model best matches the fraction of net worth held
by the top
,
,
, and
percent of the population, while
at the same time matching the aggregate capital-to-output ratio from
the perfect foresight model. Specifically, defining
and
as the
proportion of total aggregate net worth held by the top
percent in our
model and in the data, respectively, we solve the following minimization
problem:
![]() | (17) |
subject to the constraint that the aggregate wealth (net worth)-to-output ratio in the
model matches the aggregate capital-to-output ratio from the perfect foresight model
():38
The introduction of even such a relatively modest amount of time preference heterogeneity sharply improves the model’s fit to the targeted proportions of wealth holdings (second column of the table). The ability of the model to match the targeted moments does not, of course, constitute a formal test, except in the loose sense that a model with such strong structure might have been unable to get nearly so close to four target wealth points with only one free parameter.39 But the model also sharply improves the fit to locations in the wealth distribution that were not explicitly targeted; for example, the net worth shares of the top 10 percent and the top 1 percent are also included in the table, and the model performs reasonably well in matching them.
Of course, Krusell and Smith (1998) were well aware that their baseline model
match the wealth distribution well. They, too, examined whether inclusion of a
form of discount rate heterogeneity could improve the model’s match to
the data. Specifically, they assumed that the discount factor takes one
of the three values (,
, and
), and that agents
anticipate that their discount factor might change between these values
according to a Markov process. As they showed, the model with this
simple form of heterogeneity (henceforth ‘KS-Orig Hetero’ model) did
improve the model’s ability to match the wealth holdings of the top
percentiles (see KS-Orig Hetero column in the table). Indeed, as inspection of
the long-dashing locus in Figure 1 shows, their model of heterogeneity
went a bit too far: It concentrated almost all of the net worth in the
top 20 percent of the population (though rather evenly among that top
20 percent). By comparison, the figure shows that our model does a
notably better job matching the data across the entire span of wealth
percentiles.
The reader might wonder why we do not simply adopt the KS specification of heterogeneity in time preference factors, rather than introducing our own novel form of heterogeneity. The principal answer is that our purpose here is to define a method of explicitly matching the model to the data via statistical estimation of a parameter of the distribution of heterogeneity, letting the data speak flexibly to the question of the extent of the heterogeneity required to match model to data. A second point is that, having introduced finite horizons in order to yield an ergodic distribution of permanent income, it would be peculiar to layer on top of the stochastic death probability a stochastic probability of changing one’s time preference factor within the lifetime; Krusell and Smith motivated their differing time preference factors as reflecting different preferences of alternative generations of a dynasty, but with our finite horizons assumption we have eliminated the dynastic interpretation of the model. Having said all of this, the common point across the two papers is that a key requirement to make the model fit the data is a form of heterogeneity that leads different households to have different target levels of wealth.
In this section, we examine a model with an FBS household income process
that also incorporates KS aggregate shocks, and investigate the model’s
performance in replicating aggregate statistics. Krusell and Smith (1998)
assumed that the level of aggregate productivity alternates between
if the aggregate state is good and
if it is bad;
similarly,
where
if the state is good and
if
bad. (For reference, we reproduce their assumed parameter values in
Table 5.)
The decision problem for an individual household in period can be written
using normalized variables and the employment status
:
There are more state variables in this version of the model than in the model
with no aggregate shock: The aggregate variables and
, and the
household’s employment status
whose transition process depends on the
aggregate state. Solving the full version of the model above with both aggregate
and idiosyncratic shocks is not straightforward; the basic idea for the solution
method is the key insight of Krusell and Smith (1998). See Appendix C for
details about our solution method.
We now report the results of simulations, both for the model in which all
households have the same time preference factor (-Point model) and for the
version with a uniform distribution of time preference factors (
-Dist model).
While the
-Point model uses
estimated in Section 2, the
-Dist model
uses parameter values reestimated by solving the minimization problem (17)
with the KS aggregate shocks (
). Results using
our solution of the KS-JEDC model (with the KS aggregate shocks,
,
for all
, and no death (
)) are also reported for
comparison.
Table 6 shows some aggregate statistics that we think are useful for macroeconomic
analysis: The serial correlation of consumption growth, and correlation between
consumption growth, income growth, and interest rates at several frequencies.
The results are generally similar across the -Point,
-Dist, and KS-JEDC
models. They all produce positive
, and high correlation
of consumption growth with (current) income growth or (current) interest rates.
The serial correlation of consumption growth in our solution of the
KS-JEDC model is similar to that reported by Maliar, Maliar, and
Valli (2008) who also solved the KS-JEDC model (fourth column of the
table).40
41
We also report results for the representative agent model with the KS aggregate
income shock parameters (last column), the results of which are very close to
those of our solution of the KS-JEDC model.
Notes: and
are one-year and two-year growth rates, respectively.
The classic reference point for consumption growth measurement is the random walk model of Hall (1978), and the large literature that rejects the random walk proposition in favor of models that either contain some ‘rule-of-thumb’ consumers who set spending equal to income in every period (Campbell and Mankiw (1989)) or, more popular recently, models with habit formation or ‘sticky expectations’ (Carroll, Slacalek, and Tokuoka (2008)) that imply serial correlation in consumption growth (see Carroll, Sommer, and Slacalek (2011) for evidence).
The KS-JEDC model produces a relatively high correlation coefficient
, which is closer to the U.S. data (where the statistic is about
one-third) than that produced by standard consumption models stemming from
Hall (1978).42
As noted already, our
-Point and
-Dist models also imply positive
, although not as high as that predicted by the
KS-JEDC model. At first blush, it seems puzzling that the KS-JEDC model,
which includes neither habits nor sticky expectations, generates a substantial
violation of the random walk proposition. This puzzle does not seem to have
been noticed in the previous literature on the KS-JEDC model, but after
some investigation we determined that the KS-JEDC model’s sticky
consumption growth is produced by the high degree of serial correlation in
interest rates in the model, which results from the assumption about the
process of aggregate productivity shocks (see Appendix D for details). The
interesting questions, in a model with time-varying interest rates, are,
first, whether one can reliably estimate an intertemporal elasticity of
substitution (IES) from the coefficient in a regression of consumption on the
predictable component of interest rates (as Hall (1988) attempts to do),
and, second, whether consumption growth is serially correlated after
accounting for the predictable component related to interest rates (no random
walk).43
A macroeconomic question of perhaps even greater interest is whether a model that manages to match the distribution of wealth has similar, or different, implications from the KS-JEDC or representative agent models for the reaction of aggregate consumption to an economic ‘stimulus’ payment.
Specifically, we pose the question as follows. The economy has been in its
steady-state equilibrium leading up to date . Before the consumption decision
is made in that period, the government announces the following plan: Effective
immediately, every household in the economy will receive a ‘stimulus check’
worth some modest amount $
(financed by a tax on unborn future
generations).44
In theory, the distribution of wealth across recipients of the stimulus checks has
important implications for aggregate MPC out of transitory shocks to income.
To see why, the solid line of Figure 2 plots our -Point model’s individual
consumption function in the good (aggregate) state, with the horizontal axis
being cash on hand normalized by the level of (quarterly) permanent
income. Because the households with less normalized cash have higher
MPC,45
the average MPC is higher when a larger fraction of households has less
(normalized) cash on hand.
There are many more households with little wealth in our -Point model
than in the KS-JEDC model, as illustrated by comparison of the short-dashing
and the long-dashing lines in Figure 1. The greater concentration of wealth at
the bottom in the
-Point model, which is the case in the data (see the
histogram in Figure 2), should produce a higher average MPC, given the
concave consumption function.
Indeed, the average MPC out of the transitory income (‘stimulus
check’) in our -Point model is
in annual terms (first column of
Table 7),46
about double the value in the KS-JEDC model
(the fourth column of the
table) or the perfect foresight partial equilibrium model (0.04). Our
-Dist model
(second column of the table) produces an even higher average MPC
,
since in the
-Dist model there are more households who possess less wealth,
are more impatient, and have higher MPCs (Figure 1 and dashed lines in
Figure 2).47
However, this is still at best only at the lower bound of typical empirical MPC
estimates which are typically between
–
or even higher (see Table 13 in
the Appendix E).
Notes: Annual MPC is calculated by quarterly MPC
.
: Discount factors are uniformly
distributed over the interval
.
Thus far, we have been using total household net worth as our measure of wealth. Implicitly, this assumes that all of the household’s debt and asset positions are perfectly liquid and that, say, a household with home equity of $50,000 and bank balances of $2,000 (and no other balance sheet items) will behave in every respect similarly to a household with home equity of $10,000 and bank balances of $42,000. This seems implausible. The home equity is more illiquid (tapping it requires, at the very least, obtaining a home equity line of credit, which requires an appraisal of the house and some paperwork).
Otsuka (2003) formally analyzes the optimization problem of a consumer with a FBS income process who can invest in an illiquid but higher-return asset (think housing), or a liquid but lower-return asset (cash), and shows, unsurprisingly, that the marginal propensity to consume out of shocks to liquid assets is higher than the MPC out of shocks to illiquid assets. Her results would presumably be even stronger if she had allowed that households hold so much of their wealth in illiquid forms (housing, pension savings), for example, as a mechanism to overcome self-control problems (see Laibson (1997) and many others).48
These considerations suggest that it may be more plausible, for purposes of extracting a predictions about the MPC out of stimulus checks, to focus on matching the distribution of liquid financial assets across households (that is, assets which are of the same kind as represented by the stimulus check, once it has been deposited into a bank account).
When we ask the model to estimate the time preference factors that allow it
to best match the distribution of liquid financial assets (instead of net
worth),49
estimated parameter values are and the average
MPC is
(third column of the table), which lies in the upper part of the
range typically reported in the literature (see Table 13), and is considerably
higher than when we match the distribution of net worth. This reflects the fact
that matching the more skewed distribution of liquid financial assets than that of
net worth (Table 8) requires a wider distribution of the time preference factors,
which produces even more households with little wealth. The estimated
distribution of discount factors lies below that obtained by matching net worth
and is considerably more dispersed because of substantially lower median
and more unevenly distributed liquid financial wealth (compared to net
worth).
Notes: Survey of Consumer Finances, : From Castaneda, Diaz-Gimenez, and Rios-Rull (2003).
Figure 3 shows the cumulative distribution functions of MPCs for the -Dist
models estimated to match the empirical distribution of net worth and liquid
financial assets. The Figure illustrates the high values of implied MPCs obtained
for both models, especially the latter.
MPCs are generally higher among low wealth/income households and
the unemployed in both our -Point and
-Dist models (rest of the
rows in Table 7). These results provide the basis for a common piece of
conventional wisdom about the effects of economic stimulus mentioned in our
introduction: If the purpose of the stimulus payments is to stimulate
consumption, it makes much more sense to target those payments to low-wealth
households than to distribute them uniformly to the population as a
whole.
The KS process for aggregate productivity shocks has little empirical foundation; indeed, it appears to have been intended by the authors as an illustration of how one might incorporate business cycles in principle, rather than a serious candidate for an empirical description of actual aggregate dynamics. In this section, we introduce an aggregate income process that is considerably more tractable than the KS aggregate process and is also a much closer match to the aggregate data. We regard the version of our model with this new income process as the ‘preferred’ version for use as a starting point for future research.
The aggregate production function is the same as equation (1), but following
Carroll, Slacalek, and Tokuoka (2008), the aggregate state (good or bad) no
longer exists in this model (). Aggregate productivity is instead captured
by
. Specifically,
;
is aggregate permanent productivity, where
;
is the aggregate permanent shock; and
is the
aggregate transitory shock (note that
is the capitalized version of the Greek
letter
used for the idiosyncratic permanent shock; similarly (though less
obviously),
is the capitalized
). Both
and
are assumed to be log
normally distributed with mean one, and their log variances are from Carroll,
Slacalek, and Tokuoka (2008), who have estimated them using U.S. data
(Table 5).
The assumption that the structure of aggregate shocks resembles the structure
of idiosyncratic shocks is valuable not only because it matches the data better,
but also because it makes the model easier to solve. In particular, the elimination
of the ‘good’ and ‘bad’ aggregate states reduces the number of state variables to
two ( and
) after normalizing the model by
(as elaborated in
Carroll, Slacalek, and Tokuoka (2008)). As in Section 2, employment status is
not a state variable (in eliminating the aggregate states, we also shut down
unemployment persistence, which depends on the aggregate state in the
KS-JEDC or KS-Orig model). As before, the main thing the household
needs to know is the law of motion of
, which can be obtained by
following essentially the same method as described in the Appendix
C.
When matching the distribution of net worth, aggregate statistics produced by
the -Dist model with our preferred (Friedman/Buffer Stock) aggregate process
are relatively similar to those under the KS aggregate process, despite the
difference in the aggregate process (first and third columns of Table 9). Given
that there is no aggregate state in the economy, we are using
and
estimated in Section 2 and assuming that the unemployment rate
is fixed at
0.07 (same as in Section 2). Our preferred version of the
-Dist model
maintains positive
, and high correlation of consumption
growth with income growth or interest rates. We have obtained similar results by
matching the distribution of liquid financial assets (second column of the
table).
More importantly, the preferred version of the -Dist model can
produce high MPCs. For example, in the net worth case, the average MPC
is
, which is very close to the estimate under the KS aggregate
process (compare second and last but one columns of Table 7). In the
liquid financial assets case, the average MPC is higher at
(last
column).50
This paper found that the performance of a KS-type model in replicating wealth distribution can be improved significantly by introducing i) a microfounded income process, ii) finite lifetimes, and iii) heterogeneity in time preference factors. Moreover, such modifications improve macroeconomic characteristics of the model by substantially boosting the MPC out of transitory income.
Appendix
The evolution of the square of is given by
Because and
, we have
This appendix estimates the annual income process à la Moffitt and
Gottschalk (1995) using quarterly income draws generated by our income
process (Section 2) with parameter values from Table 1. Moffitt and
Gottschalk (1995) assume log permanent income follows a
random walk and log transitory income
) an ARMA process:
Interestingly, even though our true quarterly transitory shock process is just
white noise, if we estimate the process on an annual basis we obtain
positive AR () and negative MA (
) coefficients, reflecting time
aggregation. This suggests that the positive
and negative
reported
in Moffitt and Gottschalk (1995) may be (at least) partly due to time
aggregation.
Broadly speaking, the literature takes one of the following two approaches in solving the KS problem in Section 3:
Table 11 lists some existing articles that solve the KS-JEDC model according to this categorization. All articles in the table except Kim, Kim, and Kollmann (2010) solve the exact KS-JEDC model using various methods.51
The advantage of the first approach is that simulation performed to obtain the law of motion generates micro data, which can be used directly to investigate issues such as wealth distribution. The disadvantage is that this approach is generally subject to cross-sectional sampling variation, because this approach typically performs simulation using a finite number of households. Young (2010) and Den Haan (2010b)’s approaches can also be categorized in the first approach but avoid cross-sectional sampling variation by running nonstochastic simulation that approximates the density of wealth with a histogram.
The advantages of the second approach are: i) there is no cross-sectional sampling variation; ii) it is generally faster than the first approach. Using the second approach, Algan, Allais, and Den Haan (2008) and Reiter (2010) find a wealth distribution function of various moments,52 while Reiter (2010) calculates a matrix for the transition probabilities of individual wealth (see Appendix ?? for details about his technique). Kim, Kim, and Kollmann (2010) use a perturbation method that linearizes the system. Although they are not able to solve the exact same KS-JEDC model and thus modify the form of the utility function, they can solve a related problem very quickly.
We use the first approach because it directly generates various micro data (e.g., individual wealth and MPC), which can be used to examine wealth distribution and the aggregate MPC. Details about our algorithm are in the next subsection.
In solving the problem in section 3 we closely follow the stochastic simulation
method of Krusell and Smith (1998). Krusell and Smith find that per capita
capital today () is sufficient to predict per capita capital tomorrow (
).
Our procedure is as follows:
![]() |
if the aggregate state in period is good (
), and
![]() |
if the aggregate state is bad ().
We repeat this process until with a given degree of
precision.53
From the second iteration and thereafter, we use the terminal distribution of wealth
(and permanent component of income ()) in the previous iteration as the initial
one. For the case of the
-Dist model, the number of households is multiplied
by 10 in the final two (or three) iterations to reduce cross-sectional simulation
error.54
While we can eventually obtain some solution whatever the initial
is, we use
obtained using the representative agent model as the
starting point. This can significantly reduce the time needed to obtain the
solution.
Parameter values to solve the model are from Table 1 (except for the
unemployment rate ) and Table 5. The time preference factors are imposed
to be those estimated in Section 2.
In obtaining the aggregate law, we introduce the following tricks to reduce simulation errors (or to speed up the solution given a degree of estimate precision):
The estimated laws of motions for -Point,
-Dist and KS-JEDC models are
given in Table 12. The fit measured with
in all specifications exceeds
0.9999.56
Notes: The coefficients for the KS-JEDC model are very close to those estimated in Maliar, Maliar, and Valli (2010).
Although reported in subsection 3.1 may not be high
enough relative to that observed in the U.S. data, it is still not clear why
simulations produce such a high value.
Previous studies on KS type models have not investigated this issue. Using the KS-JEDC model, we performed an experiment to understand the phenomenon better. In this experiment we assume that the aggregate state switches from good to bad (or from bad to good) every eight quarters.57
Figure 4 plots 24 quarters of simulated observations (the state is
bad for the first eight quarters, good for the next eight quarters, and bad for the
final eight quarters). The figure shows that
is very persistent (it is
negative in the bad state and positive in the good state), resulting in a relatively
high
.
It is easy to understand that is higher when the state is good (and
vice versa) given the following facts:
![]() | (20) |
where ,
,
is the coefficient of
relative risk aversion,
is the interest rate,
is the time
preference factor, and
is the depreciation rate. Indeed, when
we conduct an IV regression of equation (20) using
as the
instrument,58
which effectively means estimating
, the
estimate of
is
(with a standard deviation of
) and
relatively close to the actual value of
(
). This suggests
that using the predictable component of interest rates (
),
we can obtain a reasonable estimate of intertemporal elasticity of
substitution.
While in typical simulation one state does not generally last for exactly eight
quarters, we observe sticky aggregate consumption growth (and a relatively high
) because the same mechanisms are at work as in the
experiment above.
In sum, a relatively high in the KS-JEDC model can
be interpreted as a consequence of the persistent behavior of the interest rate
. Indeed, denoting
the residual after
controlling for the predictable component of consumption growth related
to interest rates, we find that
0.02 is much lower than
.59
Table 13 summarizes the estimates of MPCs obtained using household-level data on various recent fiscal stimulus measures in the U.S.
AGARWAL, SUMIT, CHUNLIN LIU, AND NICHOLAS S. SOULELES (2007): “The Response of Consumer Spending and Debt to Tax Rebates – Evidence from Consumer Credit Data,” Journal of Political Economy, 115(6), 986–1019.
ALGAN, YANN, OLIVER ALLAIS, AND WOUTER J. DEN HAAN (2008): “Solving Heterogeneous-agent Models with Parameterized Cross-sectional Distributions,” Journal of Economic Dynamics and Control, 32(3), 875–908.
BEWLEY, TRUMAN (1977): “The Permanent Income Hypothesis: A Theoretical Formulation,” Journal of Economic Theory, 16, 252–292.
BLANCHARD, OLIVIER J. (1985): “Debt, Deficits, and Finite Horizons,” Journal of Political Economy, 93(2), 223–247.
BLUNDELL, RICHARD, LUIGI PISTAFERRI, AND IAN PRESTON (2008): “Consumption Inequality and Partial Insurance,” Manuscript.
BRAUN, RICHARD ANTON, HUIYU LI, AND JOHN STACHURSKI (2009): “Computing Densities and Expectations in Stochastic Recursive Economies: Generalized Look-Ahead Techniques,” Manuscript.
CAGETTI, MARCO (2003): “Wealth Accumulation Over the Life Cycle and Precautionary Savings,” Journal of Business and Economic Statistics, 21(3), 339–353.
CAMPBELL, JOHN Y., AND N. GREGORY MANKIW (1989): “Consumption, Income, and Interest Rates: Reinterpreting the Time-Series Evidence,” in NBER Macroeconomics Annual, 1989, ed. by Olivier J. Blanchard, and Stanley Fischer, pp. 185–216. MIT Press, Cambridge, MA, http://ideas.repec.org/p/nbr/nberwo/2924.html.
__________ (1991): “The Response of Consumption to Income: A Cross-Country Investigation,” European Economic Review, 35, 723–67.
CARROLL, CHRISTOPHER D. (1992): “The Buffer-Stock Theory of Saving: Some Macroeconomic Evidence,” Brookings Papers on Economic Activity, 1992(2), 61–156, http://www.econ2.jhu.edu/people/ccarroll/BufferStockBPEA.pdf.
__________ (2000a): “Requiem for the Representative Consumer? Aggregate Implications of Microeconomic Consumption Behavior,” American Economic Review, 90(2), 110–115.
__________ (2000b): “Why Do the Rich Save So Much?,” in Does Atlas Shrug? The Economic Consequences of Taxing the Rich, ed. by Joel B. Slemrod. Harvard University Press, http://www.econ2.jhu.edu/people/ccarroll/Why.pdf.
__________ (2011): “Theoretical Foundations of Buffer Stock Saving,” Manuscript, Department of Economics, Johns Hopkins University, http://www.econ2.jhu.edu/people/ccarroll/papers/BufferStockTheory.
CARROLL, CHRISTOPHER D., AND ANDREW A. SAMWICK (1997): “The Nature of Precautionary Wealth,” Journal of Monetary Economics, 40(1), 41–71.
CARROLL, CHRISTOPHER D., JIRI SLACALEK, AND KIICHI TOKUOKA (2008): “Sticky Expectations and Consumption Dynamics,” Manuscript.
CARROLL, CHRISTOPHER D., MARTIN SOMMER, AND JIRI SLACALEK (2011): “International Evidence on Sticky Consumption Growth,” Review of Economics and Statistics, 93(4), 1135–1145, http://www.econ2.jhu.edu/people/ccarroll/papers/cssIntlStickyC/.
CASTANEDA, ANA, JAVIER DIAZ-GIMENEZ, AND JOSE-VICTOR RIOS-RULL (2003): “Accounting for the U.S. Earnings and Wealth Inequality,” Journal of Political Economy, 111(4), 818–857.
CORONADO, JULIA LYNN, JOSEPH P. LUPTON, AND LOUISE M. SHEINER (2005): “The Household Spending Response to the 2003 Tax Cut: Evidence from Survey Data,” FEDS discussion paper 32, Federal Reserve Board.
DEATON, ANGUS S. (1991): “Saving and Liquidity Constraints,” Econometrica, 59, 1221–1248.
DEN HAAN, WOUTER J. (2010a): “Assessing the Accuracy of the Aggregate Law of Motion in Models with Heterogeneous Agents,” Jouranl of Economic Dynamics and Control, 34(1), 79–99.
__________ (2010b): “Comparison of Solutions to the Incomplete Markets Model with Aggregate Uncertainty,” Journal of Economic Dynamics and Control, 34(1), 4–27.
DEN HAAN, WOUTER J., AND PONTUS RENDAHL (2010): “Solving the Incomplete Markets Model with Aggregate Uncertainty Using Explicit Aggregation,” Journal of Economic Dynamics and Control, 34(1), 69–78.
DIAMOND, PETER A. (1965): “National Debt in a Neoclassical Growth Model,” American Economic Review, 55, 1126–1150.
FRIEDMAN, MILTON A. (1957): A Theory of the Consumption Function. Princeton University Press.
FUHRER, JEFFREY C. (2000): “Habit Formation in Consumption and its Implications for Monetary Policy Models,” American Economic Review, 90(3), 367–390, http://www.jstor.org/stable/117334.
GORMAN, WILLIAM M. (1953): “Community Preference Fields,” Econometrica, 21(1), 63–80.
GUVENEN, FATIH (2007): “Learning Your Earning: Are Labor Income Shocks Really Very Persistent?,” American Economic Review, 97(3), 687–712.
__________ (2009): “An Empirical Investigation of Labor Income Processes,” Review of Economic Dynamics, 12.
HALL, ROBERT E. (1978): “Stochastic Implications of the Life-Cycle/Permanent Income Hypothesis: Theory and Evidence,” Journal of Political Economy, 96, 971–87, Available at http://www.stanford.edu/~rehall/Stochastic-JPE-Dec-1978.pdf.
__________ (1988): “Intertemporal Substitution in Consumption,” Journal of Political Economy, XCVI, 339–357, Available at http://www.stanford.edu/~rehall/Intertemporal-JPE-April-1988.pdf.
HRYSHKO, DMYTRO (2010): “RIP to HIP: The Data Reject Heterogeneous Labor Income Profiles,” Manuscript, University of Alberta.
HUBBARD, R. GLENN, JONATHAN S. SKINNER, AND STEPHEN P. ZELDES (1994): “The Importance of Precautionary Motives for Explaining Individual and Aggregate Saving,” in The Carnegie-Rochester Conference Series on Public Policy, ed. by Allan H. Meltzer, and Charles I. Plosser, vol. 40, pp. 59–126.
JENSEN, SHANE T., AND STEPHEN H. SHORE (2008): “Changes in the Distribution of Income Volatility,” Manuscript.
JOHNSON, DAVID S., JONATHAN A. PARKER, AND NICHOLAS S. SOULELES (2006): “Household Expenditure and the Income Tax Rebates of 2001,” American Economic Review, 96(5), 1589–1610.
__________ (2009): “The Response of Consumer Spending to Rebates During an Expansion: Evidence from the 2003 Child Tax Credit,” working paper, The Wharton School.
KAPLAN, GREG, AND GIOVANNI L. VIOLANTE (2011): “A Model of the Consumption Response to Fiscal Stimulus Payments,” NBER Working Paper Number W17338.
KENNICKELL, ARTHUR B. (1995): “Saving and Permanent Income: Evidence from the 1992 SCF,” Board of Governors of the Federal Reserve System.
KIM, JINILL, SUNGHYUN HENRY KIM, AND ROBERT KOLLMANN (2010): “Solving the Incomplete Market Model with Aggregate Uncertainty Using a Perturbation Method,” Journal of Economic Dynamics and Control, 34(1), 50–58.
KIRMAN, ALAN P. (1992): “Whom or What Does the Representative Individual Represent?,” The Journal of Economic Perspectives, 6(2), 117–136.
KRUSELL, PER, AND ANTHONY A. SMITH (1998): “Income and Wealth Heterogeneity in the Macroeconomy,” Journal of Political Economy, 106(5), 867–896.
LAIBSON, DAVID (1997): “Golden Eggs and Hyperbolic Discounting,” Quarterly Journal of Economics, CXII(2), 443–477.
LOW, HAMISH, COSTAS MEGHIR, AND LUIGI PISTAFERRI (2005): “Wage Risk and Employment Over the Life Cycle,” Manuscript, Stanford University.
LUCAS, ROBERT E. (1976): “Econometric Policy Evaluation: A Critique,” in The Phillips Curve and Labour Markets, ed. by Karl Brunner, and Allan H. Meltzer, vol. 1, pp. 19–46. Journal of Monetary Economics (Supplement).
LUSARDI, ANNAMARIA (1996): “Permanent Income, Current Income, and Consumption: Evidence from Two Panel Data Sets,” Journal of Business and Economic Statistics, 14(1), 81–90.
MACURDY, THOMAS (1982): “The Use of Time Series Processes to Model the Error Structure of Earnings in a Longitudinal Data Analysis,” Journal of Econometrics, 18(1), 83–114.
MALIAR, LILIA, SERGUEI MALIAR, AND FERNANDO VALLI (2008): “Comparing Numerical Solutions of Models with Heterogeneous Agents (Models B): a Grid-based Euler Equation Algorithm,” Manuscript.
__________ (2010): “Solving the Incomplete Markets Model with Aggregate Uncertainty Using the Krusell Smith Algorithm,” Journal of Economic Dynamics and Control, 34(1), 42–49.
MEGHIR, COSTAS, AND LUIGI PISTAFERRI (2004): “Income Variance Dynamics and Heterogeneity,” Journal of Business and Economic Statistics, 72(1), 1–32.
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.
MOFFITT, ROBERT A., AND PETER GOTTSCHALK (1995): “Trends in the Covariance Structure of Earnings in the U.S.: 1969-1987,” Manuscript.
NIELSEN, HELENA SKYT, AND ANNETTE VISSING-JORGENSEN (2006): “The Impact of Labor Income Risk on Educational Choices: Estimates and Implied Risk Aversion,” Manuscript.
OTSUKA, MISUZU (2003): “Household Portfolio Choice with Illiquid Assets,” manuscript, Johns Hopkins University.
PARKER, JONATHAN A. (1999): “The Reaction of Household Consumption to Predictable Changes in Social Security Taxes,” American Economic Review, 89(4), 959–973.
PARKER, JONATHAN A., NICHOLAS S. SOULELES, DAVID S. JOHNSON, AND ROBERT MCCLELLAND (2011): “Consumer Spending and the Economic Stimulus Payments of 2008,” NBER Working Paper Number W16684.
REITER, MICHAEL (2010): “Solving the Incomplete Markets Model with Aggregate Uncertainty by Backward Induction,” Journal of Economic Dynamics and Control, 34(1), 28–35.
SABELHAUS, JOHN, AND JAE SONG (2010): “The Great Moderation in Micro Labor Earnings,” Journal of Monetary Economics, 57(4), 391–403, http://ideas.repec.org/a/eee/moneco/v57y2010i4p391-403.html.
SAHM, CLAUDIA R., MATTHEW D. SHAPIRO, AND JOEL B. SLEMROD (2009): “Household Response to the 2008 Tax Rebate: Survey Evidence and Aggregate Implications,” NBER Working Paper Number W15421.
SAMUELSON, PAUL A. (1958): “An exact consumption loan model of interest with or without the social contrivance of money,” Journal of Political Economy, 66, 467–482.
SHAPIRO, MATTHEW W., AND JOEL B. SLEMROD (2009): “Did the 2008 Tax Rebates Stimulate Spending?,” American Economic Review, 99(2), 374–379.
SOLOW, ROBERT M. (2003): “Dumb and Dumber in Macroeconomics,” Manuscript, MIT.
SOULELES, NICHOLAS S. (1999): “The Response of Household Consumption to Income Tax Refunds,” American Economic Review, 89(4), 947–958.
__________ (2002): “Consumer Response to the Reagan Tax Cuts,” Journal of Public Economics, 85, 99–120.
STORESLETTEN, KJETIL, CHRIS I. TELMER, AND AMIR YARON (2004a): “Consumption and Risk Sharing Over the Life Cycle,” Journal of Monetary Economics, 51(3), 609–633.
__________ (2004b): “Cyclical Dynamics in Idiosyncratic Labor-Market Risk,” Journal of Political Economy, 112(3), 695–717.
SZEIDL, ADAM (2006): “Invariant Distribution in Buffer-Stock Saving and Stochastic Growth Models,” Manuscript, University of California at Berkeley.
TOPEL, ROBERT H. (1990): “Specific Capital, Mobility and Wages: Wages Rise with Job Seniority,” NBER Working Paper Number W3294.
TOPEL, ROBERT H., AND MICHAEL P. WARD (1992): “Job Mobility and the Careers of Young Men,” Quarterly Journal of Economics, 107(2), 439–479.
YOUNG, ERIC R. (2010): “Solving the Incomplete Markets Model with Aggregate Uncertainty Using the Krusell-Smith Algorithm and Non-Stochastic Simulations,” Journal of Economic Dynamics and Control, 34(1), 36–41.