Eberly (1994) PDF

Adjustment of Consumers' Durables Stocks: Evidence from Automobile Purchases
Author(s): Janice C. Eberly

Source: Journal of Political Economy, Vol. 102, No. 3 (Jun., 1994), pp. 403-436
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/2138617 .
Accessed: 04/10/2013 18:18
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to Journal
of Political Economy.
http://www.jstor.org
This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

All use subject to JSTOR Terms and Conditions
Adjustment of Consumers' Durables Stocks:
Evidence from Automobile Purchases
Janice C. Eberly
Universityof Pennsylvania
This paper tests an optimal (S, s) rule in household durable pur-

chases and examines directly the resulting aggregate expenditure
dynamics. The observed decision rule responds to income uncer-
tainty and growth as predicted by an (S, s) model resulting from
transactions costs. Tests against liquidity constraints find that about
half the households purchase according to an optimal (S, s) rule.
Aggregating the (S, s) rule over households produces a cross-section
distribution of durables holdings. The empirical distribution is simi-
lar to that predicted theoretically, as is its response to aggregate
shocks. Furthermore, simulations of aggregate expenditure based
on the household distribution exhibit dynamics consistent with those
observed in the 1980s.
I. Introduction
Recent work in macroeconomics has explored the aggregate dynam-
ics of expenditure on consumer durables as a result of consumers
facing transactions costs. Little evidence has been offered, however,
to suggest that macroeconomic agents actually behave in this manner
or that the degree of households' slowness to adjust is sufficient to
explain that observed in aggregate data.
This paper takes up the issue of macroeconomic behavior by consid-
ering panel data on household automobile purchases. I consider a
pure transactions cost model and a liquidity-constrained alternative,
I benefited from the comments of the anonymous referee, Andy Abel, Roland Bena-
bou, Ricardo Caballero, David Cutler, Sara Fisher Ellison, John Heaton, Jim Poterba,
Danny Quah, Julio Rotemberg, Stephen Zeldes, and especially Olivier Blanchard. I
thank numerous seminar participants for helpful discussions and the Alfred P. Sloan
Foundation and the National Science Foundation for financial support. Any errors
are my responsibility.
[Journal of Political Economy, 1994, vol. 102, no. 3]
C 1994 by The University of Chicago. All rights reserved. 0022-3808/94/0203-0008$01.50
403

404 JOURNAL OF POLITICAL ECONOMY
and find that about half the households purchase automobiles subject
only to transactions costs.
I then consider in detail the behavior of those facing the transac-
tions cost. Theory predicts that these households should adjust their
durables stock to a target share of their total wealth and then allow
it to depreciate until it reaches a critical share of wealth. At this point,
the household should purchase a new durable good so that the stock
again equals the target share of wealth. I calculate the inaction range
implied by this decision rule for each household and find that autos
fall to about one-half of their optimal value before households adjust
their stocks. I then test the determinants of this decision rule and
find that it is unaffected by the level of income and wealth, but the
inaction range increases when income variability increases.
Finally, I consider explicitly the cross-section distribution of house-
holds according to their durables stocks. This distribution determines
how many households' durables are near the point requiring adjust-
ment and thus provides the link between individual purchase deci-
sions and the behavior of aggregate expenditures. I find this distribu-
tion quite similar to that theoretically derived and its evolution over
time consistent with aggregate changes in the observed growth rate
of income. Simulated aggregate expenditures exhibit the same rapid
acceleration and subsequent slowdown observed in the actual data in
the 1980s.
The next subsection continues with the issues and literature. Sec-
tion II presents the theoretical foundations for the transaction costs
explanation of durables purchases. Section III uses the implications
of the model to separate the households into those facing only trans-
actions costs and those possibly facing a liquidity constraint. This is
done both by exogenously splitting the sample and by estimating
an endogenous switching model. The characteristics of the group
following the transactions cost model are the subject of Section IV. I
explicitly calculate the parameters of the households' decision rule
and measure the effects of wealth, its growth rate and variance, and
other factors on these decision rules. Section V presents the theoreti-
cal ergodic and empirical distributions of households' durables stocks
relative to their wealth. I then simulate the response of the distribu-
tion to aggregate income growth during the 1980s and compare the
resulting estimates of aggregate expenditure to the actual data. Sec-
tion VI concludes with a summary and outlines future work.
A. Previous Evidence
Most empirical studies of durable goods consumption have been
made at the aggregate level. Mankiw (1982) showed that in a repre-

CONSUMERS DURABLES STOCKS 405
tentative agent framework, the stock of durables should follow a ran-
dom walk, and thus durables purchases should follow an IMA(1, 1)
process of the form
Et = Et- -(1 - 8)t- + Et, (1)
where Et is expenditure at time t, 8 is the depreciation rate of durable

goods, and E is a shock to consumption at time t. Estimation of this
equation yields a series much like a random walk: the MA(1) term is
typically much too small and often insignificant (see, e.g., Caballero
1990b).
This persistence of shocks evident in Mankiw's finding was often
modeled in the earlier literature as partial adjustment (see, e.g., Chow
1957). Bernanke motivated the partial adjustment assumption with
a convex costs of adjustment model he first applied to panel data
and subsequently to aggregate data. In panel data, Bernanke (1984)
found that consumers' auto stocks were consistent with a permanent
income model with adjustment costs and did not overrespond to tran-
sitory income. In the later work on aggregate data, Bernanke (1985),
however, concluded that convex adjustment costs are not sufficient
to explain the degree of persistence and sensitivity to transitory in-
come in the aggregate time series.'
A shortcoming of Bernanke's quadratic cost specification is that it
implies that consumers should optimally adjust frequently and in
small increments. As Bar-Ilan and Blinder (1988) point out, this con-
tradicts the observation that durables are typically purchased in large,
lumpy increments and updated only infrequently.2 Such behavior is
consistent with a threshold rule, in which consumers adjust their du-
rables stocks only when their actual stock deviates from the optimal
stock by a critical amount. Lam (1991) estimates a threshold adjust-
ment rule using a subset of Bernanke's automobile panel data set.
He posits a threshold rule in levels, so that households adjust their
automobile stocks, K, when their level deviates from the desired level,
K*, by a fixed amount, which may differ for upward and downward
adjustment. Lam estimates the thresholds and uses their characteris-
tics (asymmetries in upward and downward adjustment and differ-
ences between high- and low-wealth households) to infer the nature
of the adjustment costs.
Threshold behavior can be derived as an optimal rule in the pres-
ence of transactions costs; the rule is then precisely specified, includ-
1Bernanke estimates an adjustment cost parameter of 35 percent and finds that

more than 70 percent of adjustment occurs in the first year.
2 They pursue this argument in aggregate data, confirming the implication that in
such a case, most aggregate expenditure volatility should arise from the number of
buyers rather than from the amount purchased.

ing the parameters that may cause the rule to vary across households.
Transactions costs have been applied to durables by Bertola and Ca-
ballero (1990) and Grossman and Laroque (1990) and produce a
decision rule in which durables are allowed to deviate from their
optimal share of wealth until some threshold is reached, and then
the stock is adjusted to the optimal amount. Such models have been
previously employed in other familiar applications, notably money
holdings and inventories, among others.
Empirical analysis of the aggregate durables series suggests that
much of the observed persistence is consistent with this threshold
behavior at the household level.3 Using a threshold rule optimally
derived in the presence of transactions costs, this paper provides di-
rect evidence of such behavior in household durables purchases and
explicitly characterizes the determinants of these decisions. Linking
the household behavior to that of aggregate expenditure, the paper
then examines the empirical cross-section distribution of durables
stocks relative to wealth. Explicit aggregation shows that the observed
household slowness to adjust explains much of the short-run response
of aggregate expenditure to changes in income growth, as well as the
persistence of such changes in the 1980s.
II. Theoretical Foundation: Durables Purchases

with a Transactions Cost
I consider durables in the context of the lifetime portfolio selection
models of Merton (1969) and Samuelson (1969), using the transac-
tions cost model of Grossman and Laroque (1990). The earlier mod-
els are adjusted to allow consumption of a durable good, and then
lifetime utility is optimized to choose a portfolio among the durable,
a risky asset, and a riskless asset.4 I first develop the solution in the
frictionless case and then show Grossman and Laroque's results in-
cluding a transactions cost. Their results show that at the time of
purchase, the transaction cost model has a proportional stock rule as
in the frictionless case, so purchases track the level of wealth.5
The problem of the consumer is to maximize the present dis-
counted value of expected utility from consumption of a durable
good. The consumption flow is assumed proportional to the stock of
3Bertola and Caballero (1990) and Caballero (1990a) find an R2 of between .60 and
.90 using a threshold model estimated on aggregate durables expenditure and its
components, respectively.
4 The extension to numerous risky assets is straightforward. Here consumers hold
what can be thought of as a market composite.
5The proportion is not the same, however, so the level of purchases is different in
the two models.

the good. The consumer derives income as a return on lifetime
wealth, which can be invested in a portfolio of risky and riskless assets.
The maximization problem is
max Ef U(Kt)eY-tdt} (2)
subject to the dynamic budget constraint

dWt = rWtdt + ([L - r)Atdt + crAtdZt- (r + 8)Ktdt, (3)
where W is total wealth (including durables), A is wealth held in the
risky asset, and K is the stock of durables. The parameter y is the
subjective discount rate, puand oaare the expected rate of return and
standard deviation of the risky asset, r is the riskless rate of return,
8 is the depreciation rate of the durable good, and dZtis the increment
of a standard Brownian motion.
Setting up the problem as in Grossman and Laroque (1990), as-
sume that U(K) takes the constant relative risk aversion (CRRA) form
Kl/o (af = 0, a < 1),6 so that the solution to this problem is a linear
portfolio rule:
At ,Ur (4)
Wt a.2(1 -ao)
and
Kt 1 'y - -
(-
t- r)2a1 (5)
Wt (r + ) (1 - (x) L 2a 2(1 -
t)J
A constant proportion of wealth is held in the risky asset and in

durable goods.7 Since a < 1, as depreciation decreases, more wealth
is held in the durable good.
Suppose now that in addition to equations (2) and (3), the consumer
optimizes subject to the payment of a transactions cost when purchas-
ing the durable good. Following Grossman and Laroque (1990), as-
sume that this cost is proportional to the amount of durable good
sold by the consumer:
WT + =WT_ - XK, (6)
where T+ is the moment immediately following the durable pur-
6 The case of (x = 0 (logarithmic utility) can be addressed along similar lines.

7 Grossman and Laroque (1990) characterize this solution in their appendix. Note
that the second (bracketed) term in eq. (5) is strictly positive. This follows from the
parameter restriction requiring bounded expected utility (see Merton 1969; Grossman
and Laroque 1990).

chase, T- is the moment before, and X is the proportional transaction

cost.8 This cost may be thought of as a direct sales commission, sales
taxes, a search cost, or the result of a lemons problem that underval-
ues used cars (see Akerlof 1970).
Grossman and Laroque (1990) show that optimal consumer behav-
ior under these conditions is an (S, s) rule governing the state variable,
(WIK) - X y, and the control variable, K. Behavior is characterized
by three critical values of the state variable: a lower bound, an upper
bound, and an internal return point.' The upper and lower bounds
trigger adjustment when they are reached, and the internal return
point is the target value of y chosen at adjustment. Typically, W grows
and K depreciates, so y increases over time. When y reaches the upper
bound, the consumer readjusts his portfolio to increase K, returning
y down to its target value. Then W would continue to grow and K to
depreciate until the upper bound is again reached, and the process
repeats. Occasionally, however, W may fall, since its growth is ran-
dom. If W falls more than K depreciates, then y decreases, and the
consumer may hit the lower band of y, where he has too much of
the durable good. The consumer would then adjust his stock of K
downward, returning the state variable up to its target value. A sam-
ple path of this behavior is shown in figure 1.
This solution precisely characterizes the durables choice at adjust-
ment. Since the target point of the state variable (y) is a constant, at
adjustment (WIK) - X always equals a constant (y*), so K is a fixed
fraction of W. If X and T are times at which adjustment occurs, we
have
in ( I) n (f) = L,TV (7)
where tLT is the realized rate of return on wealth between L and 7.

This realization, of course, depends only on the stochastic process,
the riskless and mean risky rates of return, and the portfolio rule.'0
None of these is history dependent. Grossman and Laroque suggest
the intuition for this result in their conclusion, by noting that the size
8 Two features of the transactions cost are important for the results. First, since the
transactions cost is proportional to the durables stock sold, it increases with wealth.
Thus the cost does not diminish in importance as households become wealthier, as a
purely fixed cost would. Second, the cost does not depend on the amount of the new
purchase, so given the existing durable, it is a fixed cost and thus results in threshold
behavior.
9 Calculation of these critical values requires numerical methods; they will be used
in Sec. IV.
10This assumes that the relative price of durables is fixed.

YH
y(t)
t
FIG. 1.-Sample path of optimal (S, s) behavior. When yt reaches YL or yH, the con-
sumer pays the transaction cost and adjusts K so that y = y*. Then K depreciates and
W grows stochastically until yL or y' is reached again, and the process continues.
of durables purchases will be unpredictable to the extent that wealth

is unpredictable.
This implication can be made more general than the specific case
generated above. For more general stochastic processes, it can be
shown that the expectation of the present value of marginal utility
derived from the next durable purchase (over its holding period,
from X to T) must equal that derived from the current durable over
its holding period. This may be thought of as a generalization of
Hall's (1978) random walk hypothesis to a case in which consumption
derives from a durable good optimally held for an interval in continu-
ous time. Its derivation relies on the value matching and smooth
pasting conditions and holds as long as there is an interior solution
to the optimization problem. In practice this can be mapped into
familiar random walk consumption tests by assuming a CRRA utility
function. Then an equation similar to equation (7) can be derived
without proportionality of wealth and durables stocks at times of pur-
chase. Specifically, this can be written as
/KTi
In fK rr + 'ILTJ (8)
where r is a constant that depends on the parameters governing

preferences and the stochastic driving process and -qhas the familiar

interpretation as an expectational error. This result requires much

less structure than that given in equation (7).1
III. Separating Liquidity-Constrained

Households
Liquidity constraints and transactions costs are not mutually exclu-
sive; constrained and unconstrained households may both face trans-
actions costs. However, household purchase decisions may deviate
from their optimal path when consumers face binding liquidity con-
straints. Work on nondurables has shown these deviations to be em-
pirically relevant for a substantial number of households.'2 Including
constrained households in estimates of a pure transactions cost model
may compromise consistency. I therefore examine the possibility that
some households face binding liquidity constraints (violating eq. [8]
in Sec. II) before turning to the transactions costs model.
A. EmpiricallyDistinguishing Binding Liquidity-

ConstrainedHouseholdsfrom (S, s) Households:
Exogenous Sample Splitting
Section II showed that conditional on a durable having been pur-
chased, we can perform a test equivalent to the random walk tests of
consumption. The martingale property should be true for transac-
tions cost models, as seen in equation (8), since no information known
at the time of the previous purchase (i) should affect the current
slope of the purchase path. However, in the case of a binding liquidity
constraint, the martingale property fails and previously known vari-
ables that affect liquidity will also affect the slope of the purchase
path. I therefore examine whether past liquidity variables or the pre-
dictable component of these variables affects the current slope of the
purchase path.
I use data available in the Survey of Consumer Finances (1983 and
1986), compiled by the Federal Reserve. The data comprise 2,422
(representative sample) households questioned in both 1983 and
1986. They provide information on their asset holdings, as well as
liabilities, income, and major purchases.'3
" With a stochastic process different from that specified in eq. (3), the threshold
levels of durables that trigger adjustment would not necessarily be fixed shares of
wealth.
12 Hall and Mishkin (1982) and Zeldes (1989) both use household food consumption
data and find that a significant share of households in their samples are liquidity
constrained.
13 A more detailed description of the data appears in App. B.

CONSUMERS' DURABLES STOCKS 411
I begin by identifying a priori those individuals I expect to be
liquidity constrained and those I do not, and then test for differences
in their behavior.'4 The survey asks households whether they have
been denied credit or have received less credit than they requested
any time in the past two years. The survey further asks respondents
whether or not they have been discouraged from applying for credit.
I divide the sample into those with affirmative responses to either
question and those with negative responses to both, then estimate the
following equation for each group:
A In Ki (L, T) = P3o+ PI Yi () + I2Ademand, + El- (9)
where A In K(L, T) is the log of the ratio of automobile stocks after
the latest reported purchase (deflated to 1982 by the consumer price
index for new cars) to stocks at the previous purchase, Y(L) is a mea-
sure of liquidity known prior to the first purchase, and Ademand is
a vector of demand shift variables.'5
For the (S, s) households under the assumptions in the previous
section, PI should be zero. For liquidity-constrained households, the
expected sign of PI is ambiguous. If the liquidity constraint depresses
consumption, for given future income, higher current liquidity tends
to decrease the slope of consumption. However, current liquidity may
be correlated with liquidity growth as well. This may actually increase
the slope of the purchase path, if the correlation is large enough.
Therefore, a significant coefficient of either sign on past liquidity is
evidence of a liquidity constraint, and the sign suggests the correla-
tion between levels and growth rates. The observed correlation (for
income) in the sample is - .3, so P1 is expected to be negative if there
are liquidity constraints that decrease durables stocks.'6
Estimation of this equation requires observation of the auto stock
at two purchase times and a measure of liquidity known at the time
of the first purchase observation. While a long history of purchases
can be compiled from the Survey of Consumer Finances, income is
available only from 1982 through 1986. Equation (9) is therefore
estimated using two measures of known liquidity. First, I use a sub-
sample of 550 households that purchased at least two vehicles over
the period 1982-86. Income in 1982 is known at the time of the
first purchase for these households and can be used as the liquidity
14 This is the strategy employed by Zeldes (1989).

15 The Ademand vector includes controls for household composition (marital status,
number of children), labor force participation (unemployment, retirement), and
household size.
16 Chah, Ramey, and Starr (1991) point out that if durables can be used as collateral,
a liquidity constraint may tend to increase optimal stocks. The results to follow do not
exhibit this effect.

measure. This subsample selects out infrequent purchasers as well as

first-time buyers. The former will be reincluded in the sample for
later tests, whereas the latter are not the subject of the model and
will be excluded throughout. The second measure of known liquidity
is obtained by taking the component of income correlated with pre-
dictable household characteristics. I use data from the 1985 Con-
sumer Expenditure Survey and regress total income (before tax) on
the age of the household head and dummy variables for the educa-
tion, occupation, gender, and race of the household head (and inter-
actions). I use the coefficients from this regression to estimate the
predictable component of income for households in the Survey of
Consumer Finances.'7 With this measure of known liquidity, histori-
cal stock data can be used to obtain the previous purchase, which
(along with one purchase during the sample period) can be used to
construct the dependent variable, A In K (L, T). This method expands
the sample size to 1,248 purchasing households, or 61 percent of all
the auto owners in the Survey of Consumer Finances.
The results of estimating these equations are reported in columns
1 and 2 of table 1. I exclude the coefficients on the Ademand vector,
but their inclusion or exclusion has little effect on the other results.
For the liquid group, either measure of liquidity has a small negative
effect on the slope of the consumption path, but this effect is not
statistically significant when past income is the measure of liquidity.'8
For the constrained group, known liquidity has a substantial and sta-
tistically significant negative effect on the consumption profile. When
one controls for variation in demand, the slope of the purchase pro-
file (Po + PYi) is larger on average for the constrained group (2.37
with a standard error of 0.53) than for the unconstrained group (1.81
with a standard error of 0. 17).'9 This is consistent with the prediction
of liquidity constraints that, ceteris paribus, constrained households
should have steeper consumption profiles than unconstrained house-
holds.
17
This method avoids possible within-sample bias resulting from estimating the pre-
diction equation on the same households for which prediction is needed. The sample
size in the Consumer Expenditure Survey regression was 4,640, and the R2 was .30.
The procedure is explained in more detail in App. B.
18 The significant negative coefficient on predictable income for the unconstrained
households may arise from misclassification of some constrained households. There
are several reasons to suspect this problem. First, this splitting method uses information
from 1982 to split the sample. If individuals were constrained at the time of purchase
but do not report credit denied or discouraged in 1982, they will be misclassified as
unconstrained.
19 These results correspond to the estimates in col. 2 of table 1. Those for col. 1 are
similar.

TABLE 1
SUBSAMPLE AND SWITCHING ESTIMATION OF THE EULER EQUATION
A in K(L, r) = Po + PlY + R2Ademand + e
SWITCHING
SAMPLE Total
SPLITTING: Income/
Credit History Lifetime Predictable
Income Income Income
SORTING VARIABLE (1) (2) (3) (4) (5)
Liquid Group
Constant 2.20 2.40 1.81 2.04 2.70
(.17) (.36) (.22) (.14) (.81)
Past income ($10,000) - .03 - .00 - .02
(.03) (.04) (.02)
Predictable income ($10,000) -.21 -.21
(.08) (.20)
Constrained Group
Constant 3.9 4.0 3.52 3.37 3.70
(.51) (.56) (.50) (.36) (.21)
Past income ($10,000) -.30 -.47 -.41
(.15) (.25) (.16)
Predictable income ($10,000) -.65 -.78
(.21) (.16)
Switch value Credit denied 30,228 1.06 29,684
(4,703) (.04) (1,083)
Total observations 550 1,248 550 550 1,248
NOTE.-COIS. 1 and 2 report the results of estimating eq. (9). Cols. 3-5 report the results of estimating the system
in eqq. (10) and (11), using switching variables as noted. Lifetime income is the predicted value of the regression
of total income on demographic variables. Predictable income is the predicted value of total income based only on
age, occupation,education,gender,and race.Standarderrorsare in parentheses.All regressionsare heteroscedas-
ticitycorrected.
B. Endogenous Switching Estimation

There are several reasons to question the exogenous sample-splitting
method. First, the 1983 survey followed a period of credit upheaval
in the United States, so individuals denied credit during this period
may not be representative.20 In addition, some households may be-
lieve that they are ineligible for credit and therefore not apply, while
not identifying themselves as deniedor discouraged.On the other hand,
a household might apply for a large amount of credit and receive
less than requested, while remaining sufficiently liquid to purchase
20
The 1980 credit controls and the 1981-82 recession induced a collapse of con-
sumer credit. In aggregate, consumer credit grew only 0.5 percent in 1980, 5 percent
in 1981, and 4 percent in 1982, before jumping to 13 percent in 1983 (EconomicReport
of the President, February 1990, table C75).

an automobile. Finally, this splitting method a priori identifies less

than 15 percent of the sample as constrained; this is lower than most
other estimates. Campbell and Mankiw (1989), for example, estimate
that rule-of-thumbconsumers, who equate consumption and income,
represent between 20 and 50 percent of the aggregate. Zeldes (1989),
on the other hand, splits his sample of households into groups on
the basis of levels of liquid wealth, nonhousing wealth, and total
wealth. He finds evidence of liquidity constraints among the low-
wealth households, constituting 46, 67, and 29 percent (respectively)
of the total sample.
To address these issues, I used an endogenous method of sample
splitting. The constrained and unconstrained households may be
thought of as defining two regimes; the parameters of each can be
estimated simultaneously using equation (8), along with the probabil-
ity of being in each regime. This method is described in Goldfeld
and Quandt (1976) and treated more generally in Maddala (1986).
The switching model is estimated using maximum likelihood includ-
ing the two regimes, with observations assigned to each regime on
the basis of the value of the switching variable.
I split the sample using several criteria, with similar results. The
first variable used is total income in the subsample of 550 for which
income is known before adjustment. Income measures the liquidity
flow available to meet consumption needs. The second measure ac-
counts for the fact that income relative to consumption determines
whether liquidity constraints bind. For example, a household with
relatively high income but even higher expected income might indeed
face a binding constraint. The second measure therefore scales cur-
rent income by a measure of desired consumption. I thus estimate a
value for lifetime income, following Hall and Mishkin (1982), as the
predicted value of a regression of total income on observable charac-
teristics, specifically the age, education, occupation, race, gender, la-
bor force participation, and marital status of the head of household,
and household composition.21 This is not permanent income (since
it excludes the annuity value of transitory income and idiosyncratic
permanent income), but is one component. The ratio of total income
to lifetime income is a measure of the adequacy of current liquidity
to meet desired consumption. For the larger subsample of purchasing
households, I use predictable income as the switching variable. By
this method, households are distinguished using information known
at the time of the earlier purchase.22
21 Lifetime income differs from the "predictable income" estimated earlier because
it may depend on current, not just predetermined, variables.
22 If
a constrained household becomes unconstrained between purchases, its con-
sumption profile would still exhibit the effect of the constraint. Therefore, classification

Equations (10) and (11) compose the estimated system, where the
switching variable is S. Observations are weighted into the first regime
if S is greater than or equal to the critical value S* and to the second
if less than S*. The vectors I and a are then estimated along with
D*:
AIlnK.(L,T) = + P1IYi ()
P3o + I12Ademand, + Ei if S.S*, (10)
AlnK-(LT) = Ot + IYi (L) + 0I2Ademand + Ei if S <S*. (11)
The results of this estimation are reported in columns 3-5 of table
1. Column 3 gives the results when income is used as the switching
variable and the level of past income is used as an independent vari-
able. Column 4 shows the results when the ratio of income to lifetime
income is used as the switching variable. Column 5 repeats the estima-
tion that uses predictable income rather than actual income and the
larger sample of purchasing households.
The switching results reinforce those of the exogenous sample-
splitting exercise. When the level of income is used as the trigger, the
endogenous switch occurs at a household income of $30,228, which
sorts 58 percent of the sample into the constrained group. Within
this group, the coefficient on past income is negative and signifi-
cant, and 50 percent larger than in column 1. For the households
with high income, the coefficient on past income is indistinguishable
from zero.
Column 4 of table 1 reports the results of the switching estimation
using the ratio of income to lifetime income as the trigger variable.
The switching value is estimated at 1.06. The parameter estimates
are qualitatively the same as those for the level of income switching
and are quantitatively bounded by the earlier estimates. Column 5 of
table 1 repeats the estimation procedure using predictable income
rather than past income and the sample of 1,248 purchasing house-
holds. The coefficient on predictable income for the unconstrained
households is negative but not statistically different from zero. For
the constrained households, the coefficient on predictable income is
very large and statistically significant. Again, the slope of the purchase
profile for constrained households is higher (2.65) than for the un-
constrained households (1.97).23
These results suggest that there is a substantial group of house-
holds that purchase automobiles facing only a transaction cost rather
than liquidity constraints. The rejection of equation (8) for the other
into constrained and unconstrained households for this test should be made at the
time of the earlier purchase.
23 These values correspond to the estimates in col. 4 of table 1. The results from
cols. 3 and 5 are similar.

households may be due to liquidity constraints or to the failure of

other assumptions in Section II, but in either case this rejection
should not enter our assessment of (S, s) behavior. Throughout the
remainder of the paper, I therefore use the switching results from
the ratio of income to predictable income (table 1, col. 4) to split the
sample into (S, s) and liquidity-constrained households.
Table 2 describes the characteristics of the constrained and uncon-
strained households.24 The income of the constrained group is, on
average, less than 50 percent of the income of the unconstrained
group. The net worth of constrained households is 40 percent lower
than that of the unconstrained, and the median constrained house-
hold has less than $1,000 in checking and savings accounts. Approxi-
mately equal shares of the two groups report being denied credit.
Regarding their automobile purchases, constrained households, on
average, spend about 25 percent less than the unconstrained. They
are 30 percent less likely than the average to obtain a loan, but when
successful they borrow about the same amount (80 percent of pur-
chase price) as their liquid counterparts. The constrained household
heads are slightly older and are more likely to be headed by non-
whites, women, and single individuals.
IV. Behavior of Consumers Facing a

Transactions Cost for Automobiles
Having identified those households facing only a transactions cost, I
now focus explicitly on describing their behavior. First, I calculate
directly the (S, s) bands of these households. I then estimate the deter-
minants of the width of the bands. The results indicate that increasing
the level of wealth or income does not reduce the waiting time be-
tween purchases. On the other hand, higher variability of movement
through the bands widens the bands since households want to avoid
a premature purchase that is costly to reverse.
A. Location of the (S, s) Bands

The primary characteristic of the (S, s) model is the range over which
the consumer does not adjust his durables stock. The limits of this
"range of inaction," the consumer's (S, s) bands, can be calculated by
observing the value of WIK before and after the consumer adjusts
his durables stock.25 The observation before adjustment provides the
24
Table 2 reports the characteristics of the entire representative sample, not just the
auto purchasers examined in table 1.
25
I neglect the term - X in y in what follows so that the calculation does not depend
on an assumed transactions cost.

TABLE 2
CHARACTERISTICS OF THE SUBSAMPLES, 1983
Constrained
Liquid Group Group Total
(N = 1,348) (N = 1,074) (N = 2,422)
Financial Characteristics
Income:
Mean 36,767 17,459 28,205
Median 28,920 12,565 22,000
(35,783) (15,792) (30,249)
Net worth:
Mean 128,572 78,570 106,399
Median 46,732 34,811 41,018
(523,828) (153,327) (404,462)
Checking and saving:
Mean 4,691 3,192 4,026
Median 1,476 953 1,200
(12,039) (6,386) (9,963)
Report credit denied 182 152 334
Automobile Purchases
Months since last purchase 30 36 33
(27) (41) (34)
Vehicle purchase if > 0 11,986 9,306 10,990
(9,728) (7,511) (9,058)
Share of new car owners with loans .46 .32 .41
Share financed if loan obtained .78 .79 .78
Demographics
Age in years 42.3 51.3 46.3
(14.3) (17.6) (16.5)
Share nonwhite .12 .14 .13
Share female .05 .48 .24
Share married .85 .42 .66
Share professional/technical .14 .19 .16
Share managerial/administrative .12 .13 .12
Share sales/clerical .11 .17 .13
NOTE.-The statistics are computed from the 1983 Survey of Consumer Finance. The statistics represent the
means (unless otherwise noted) of each series for the unconstrained subsample, the constrained subsample, and
the entire representative sample. Standard deviations are in parentheses. Demographic variables refer to the head
of household.
trigger point, and the observation afterward provides the target

point.26 The durables stock, K, is directly observable.27 A complica-
tion, however, is the difficulty of measuring total lifetime wealth, W,
26 The total width of the range of inaction in this model is (y* _ YL) + (YH _ Y*)
yH _ yL since households jump to y* from either trigger. This may not be true in
other specifications in which households choose different target stocks depending on
the trigger they hit.
27 For households with multiple automobiles, K is calculated as the sum of their
values.

which includes both financial and unobservable human wealth. Net

worth is clearly the first component of total wealth, but human wealth
is more problematic. In order to calculate human wealth, I assume
that a household's horizon is the difference between the age of the
head and age 65; for household heads over 65 with labor income, I
take the horizon to be 1 year. Human wealth is then the present
discounted value of current income over this horizon, discounted at
5 percent annually. Define rras the ratio of human wealth to current
income. The state variable, y, should then be the following sum:
net worth income
=- durables 'durables (12)
In the theoretical model, there is only one durable good. If the
denominator includes only automobiles, we shall underestimate dura-
bles and overestimate y. A simple adjustment is to divide automobile
stocks by their share in total durables, which in the sample is 10
percent. This produces an estimate of durables based only on auto-
mobiles and forms the denominator of y.28
Table 3 reports the means of these ratios and the resulting values
of y. The first row reports the means of y calculated as above, taken at
the trigger and target points.29 The second row reports the theoretical
values for these points.30 Theory indicates that at adjustment about
13 percent of total wealth should be held in durables. This would be
allowed to drift down to about 6 percent before adjusting again, or
up to 30 percent if wealth falls. The empirical values are close to
those theoretically predicted, and while the standard errors are large
(this is discussed further below), the model seems to do well on aver-
age. The means of the components of y are reported in the third and
fourth rows, and they follow the same general pattern as the theoreti-
cal prediction.3'
28 This forms an estimate of automobiles as though they were the only durable
consumption good available.
29 The empirical trigger and target values of y are calculated by taking the values of
durables, net worth, and income for households that purchased a new automobile and
calculating y*. Income and net worth and the value of their automobile before the
purchase are used to calculate yH or yL, depending on whether the household adjusted
its durables stock up or down. Since y is measured at a fixed interval and not at the
moment of adjustment, band width is underestimated. The magnitude of this bias will
vary with the speed of movement through the bands. This will be discussed further
in Sec. IVB.
30 As noted earlier, the trigger and target values of y are calculated using numerical
methods. The table indicates the parameter values (corresponding to those in Gross-
man and Laroque [1990]) used in the calculations.
31 The last two rows in both the upper and lower panels of table 3 give the band
calculations for the components of y separately. These values follow the same pattern
as those for the calculated y, suggesting that the band widths found for y do not depend
on the particular formulation chosen for human wealth and autos.

TABLE 3
OBSERVED VALUES OF THE (S, s) BANDS: CALCULATED y*, yH, AND yL
TRIGGERS TO ADJUST
STATE VARIABLE TARGET: y* yH yL

y + X = W/K (1) (2) (3)
Empirical 7.7 16.9 7.6

(6.1) (18.5) (5.4)
Theoretical 7.7 17.1 3.4
Net worth/automobile stock 19.4 37.8 14.7
(26.5) (85.9) (13.9)
Income/automobile stock 4.9 11.3 4.7
(3.5) (11.3) (3.2)
Normalized by Target Value
Empirical 1.0 2.2 .9
Theoretical 1.0 2.2 .5
Net worth/automobile stock 1.0 2.0 .8
Income/automobile stock 1.0 2.3 .9
NOTE-The upper panel gives the observed values for the target point y* and the trigger points, yH and yL, as
well as predictions from a parameterization of the model. Standard deviations are in parentheses. The lower panel
reports band width measured as the ratio of the trigger points to the target point. The bands are calculated by taking
the ratios of wealth to durables preceding adjustment of the automobile stock and then again after adjustment. The
theoretical values are calculated for the following parameter values in the Grossman and Laroque (1990) model:
transactions cost of .05, annual depreciation rate of .10, mean risky returns of .059 with a standard deviation of
.22, and a riskless rate of .01. Asset returns and standard deviations are taken from R. G. Ibbotson Associates
(1989). The depreciation rate corresponds to an average time to scrappage of 10 years in Ward's (1988) and the
estimate from App. B., sec. D.
The lower panel of table 3 indicates observed band width. Theory

predicts that y doubles from target to trigger (halving the ratio of
durables to wealth) before adjusting, corresponding to values of two
in the second column. Column 2 of the table shows that this is true
for both the calculated value of y and its components. Column 3 in
the lower panel of table 3 suggests that if wealth is falling, consumers
wait until the stock is 10 percent too large before adjusting it down-
ward. This is smaller than the theoretical value, which is closer to 50
percent. The number of consumers who adjust downward, however,
in the sample is relatively small.
These band widths are substantially wider than those estimated by
Bertola and Caballero (1990) in the aggregate data.32 They find that
consumers adjust their durables stocks by 26 percent when adjusting.
This may result from their wider classification of durable goods: all
durables rather than just automobiles. However, in subsequent work,
Caballero (1990a) estimates that automobile stocks should increase
32
Bertola and Caballero construct aggregate durables stocks and wealth from the
series for durables purchases and income. Setting the theoretically optimal durables
stock to a share of wealth, they then estimate the (S, s) bands necessary to generate
the observed deviations from this stock.

by a factor of 2.8 at adjustment, which is closer to the estimates in

table 3.
Note that the high standard errors in table 3 are not a rejection of
the model; they may simply indicate cross-section heterogeneity in
the individual parameters that determine band width.33 Using the
model of how the thresholds are chosen, I can exploit this heteroge-
neity to estimate how these parameters affect the choice of band
width.34
B. Determinantsof Band Width

Band width is determined by three characteristics. First, factors that
systematically increase the speed of movement through the bands will
increase band width, since the consumer wants to avoid frequent and
costly adjustment. Intuitively, the state variable, WIK, can be easily
transformed into a measure of desired versus actual durables hold-
ings, for example, K*/K = OW/K. The triggers and targets can then
be thought of as critical values in the preferred level of durables
relative to its actual value. In this transformation, drift in the state
variable (movement through the bands) occurs because of deprecia-
tion in K and also because of predictablechanges in K*. In the model
here, the only source of such changes is wealth. However, a more
general formulation would account for changes in household size,
number of drivers, desire for quality or reliability, and relative prices.
Therefore, in addition to the mean growth rate of wealth, these vari-
ables would affect band width.35
Second, risk has two effects. The risk aversion effect tends to widen
the bands since the consumer wants to avoid adjusting often in re-
sponse to jumpy realizations of y. The portfolio effect, however, re-
duces the speed of movement by reducing investment in the riskier
high-return asset and therefore tends to narrow the bands. Thus
increases in the variance of returns may increase band width, but this
effect will be damped if the consumer substitutes away from this asset.
Third, an increase in the adjustment cost also broadens the bands.
To quantify these effects, I estimate the following equation:
width- = Po + PI Ldum. + r2InYi + r33InNW, (13)
+ P4 R(AYlfei) + P5a(Yi) + P6 In taxi + mi
33 Lam (1991) also finds a very high idiosyncratic variance of the thresholds used by
the households.
34 Caballero and Engel (1991) consider analytically the effect of such heterogeneity
on aggregation with a uniform cross-section distribution.
35 Increased depreciation also increases band width, but cross-section variation is
difficult to identify.

where Ldum is a dummy for downward adjustment of the durables

stock, in Y is the natural logarithm of 1985 income, In NW is the
natural logarithm of 1985 net worth, ,u(AYlife) is the growth rate of
lifetime income (as estimated in Sec. III), u(Y) is the standard devia-
tion of non-lifetime income from 1982 to 1985,36 and tax is state
auto licensing fees per household. The dependent variable is the
share that the state variable y is adjusted. I include both upward and
downward adjustment in the sample since the distance to the lower
bound is typically a constant fraction of the distance to the upper
band; a dummy variable for adjustment from the lower band is in-
cluded in the regression.37
Estimation of equation (13) requires data on purchases for house-
holds that previously owned automobiles. The subsample of 1,248
automobile owners who purchased new vehicles is then divided into
an unconstrained group (642 (S, s) households) and a constrained
group (606 households) using the switching results from column 4
of table 1.
Column 1 of table 4 gives the estimation results for the (S, s) house-
holds. The first two rows give the coefficients on the constant term
and the dummy variable for downward adjustment. The next two
rows give the results for income and net worth, neither of which
should affect band width. The last three variables should all increase
band width. The first is growth in lifetime income. Since this depends
on demography, such changes should be largely predictable and
should be factored into the choice of band width. Higher growth
should therefore increase band width. The next variable is income
variance, which should increase the variability of growth in the state
variable and therefore also increase band width.38 The final variable
is (log) license fees, which increase the cost of adjusting automobile
stocks and increase band width.
Column 1 of table 4 is generally supportive of the hypotheses
above. Income and net worth do not affect the choice of band width;
36 Non-lifetime income is the difference between total income and lifetime income.
The standard deviation of this variable, as a share of 1982 income, is used as o(Y) to
eliminate potentially foreseeable income changes from the measure of uncertainty.
Note that this is income uncertainty rather than the pure portfolio uncertainty of
Grossman and Laroque. Potentially nondiversifiable (and idiosyncratic) labor income
uncertainty is a component of risk unaccounted for in the model.
37 If there is drift in the state variable due to depreciation, e.g., the bands will be
asymmetric about the target point, and upward and downward adjustment will have
different magnitudes.
38 Recall that increases in variability have two opposing effects on band width. The
tendency of variability to reduce band width arises because households substitute away
from risky assets, damping this effect on band width. The increase in incomevariability
noted above is an increase in total variability and therefore already incorporates any
portfolio change. Any remaining effect should be to increase band width.

TABLE 4
DETERMINANTS OF BAND WIDTH
+ I35UY
width = Poo+ PILdum + IB2In Y + 0B3 In NW + I34NW(AYlife)
+ IP6ln(tax) + -q
(S, s) Households Constrained Households

Independent Variables (1) (2)
Constant 2.65 1.96
(.79) (.68)
Dummy for downward adjustment .20 - .23
(.08) (.09)
Income ($10,000) -.11 -.12
(.07) (.05)
Net worth ($10,000) -.02 -.10
(.03) (.03)
Real lifetime income growth .42 .15
(.15) (.14)
Year-to-year income variance .22 .33
(.03) (.06)
ln(license fees) (per household) -.15 .27
(.10) (.11)
R2 .10 .14
Number of observations 642 606
NOTE.-Standard errors are in parentheses. Width is calculated as f(y*/trigger) - 1|. The (S, s) households
and the constrained households are those endogenously sorted into the unconstrained and constrained groups,
respectively, by the switching regression of Sec. IIIB. The switching variable is income/lifetime income, as reported
in col. 4 of table 1. All regressions are heteroscedasticity corrected.
income has a marginally significant effect, but both are insignificant

at conventional levels. Lifetime income growth is significant but has
a small effect on band width. Increasing lifetime income growth by
one percentage point increases band width by only 0.42 percentage
points.39 Income variance has a significant and very large effect. In-
creasing income variability by one standard deviation (0.50) increases
band width by 11 percent (both are calculated at their means). This
implies that the durables stock would be allowed to deviate from
optimal by 55 percent rather than 50 percent before adjusting. The
fees variable enters with the wrong sign but is insignificant at conven-
tional levels.
Column 2 of table 4 reports the results of the same regression
for the liquidity-constrained households. These results show much
different behavior than for the (S, s) group. Income and net worth
both enter negatively and significantly, implying that households with
39 The downward bias in measured band width increases with the speed of movement
through the bands. This biases downward the coefficient on lifetime income growth,
since it is correlated with the growth rate of y. The result reported above is therefore
a lower bound.

higher income and assets keep their durables stocks closer to the
optimal level. Lifetime income growth does not have a significant
effect for this group. Income variance again enters significantly, in-
creasing the size of adjustment by 50 percent more than for the un-
constrained group. License fees have a significant positive effect on
band width for the constrained households. The magnitude of the
coefficient is moderate, however, suggesting that doubling fees (from
their average value of $90 per household) increases band width by
15 percent.40
These results suggest that income and net worth levels do not affect
how closely the (S, s) households maintain their auto stocks relative
to their optimal stocks. As income rises, therefore, this implies that
households will not increase the frequency of adjustment. The (S, s)
consumers' durables stocks will deviate from their optimal level to a
greater extent when income variability is high and also when income
growth is high, though the latter effect is smaller. The former is
consistent with Romer's (1990) evidence on consumer uncertainty
and postponement of durables purchases during the Great Depres-
sion. For the liquidity-constrained consumers, however, income and
net worth are important determinants of how closely they are able to
track their optimal auto stocks.
V. Distribution of Durables Stocks Relative

to Wealth
The critical step in moving from the household behavior described in
Section IV to the macroeconomic behavior of durables expenditures
involves the cross-section distribution of households. This distribu-
tion determines how many households are near the trigger points
and, therefore, how many will adjust their durables stocks during a
given period. I first present the theoretical limiting distribution of
households and find that it closely matches the observed empirical
distribution. I then examine how the distribution responds to aggre-
gate shocks and show that it can explain the magnitude and persis-
tence of the response of durables expenditure to income growth in
the 1980s.
40 It is also interesting to note that the sign of the dummy for downward adjustment
changes for the liquidity-constrained households. While the (S, s) households allow a
large deviation from optimal before adjusting downward, the liquidity-constrained
households allow a smaller deviation (compared to when they adjust their durables
upward). This is consistent with the presence of sales for liquidity reasons.

A. The Ergodic Distributionof Durables Relative

to Wealth
The ergodic distribution of y for a single individual gives the probabil-
ity density function for his value of y at any point in time. In this
case under very general conditions, the ergodic distribution can be
precisely characterized. With some simplifying assumptions about the
portfolio chosen by the household, the ergodic distribution can be
solved analytically (see App. A). This distribution is shown in figure
2. The shape, piecewise exponential and discontinuous at the target
point, is similar to that found by Tsiddon (1991) and Bertola and
Caballero (1990) in other models.
Under simple regularity conditions, when there is no aggregate
uncertainty, this distribution is also the cross-section stationary distri-
bution. This may be interpreted as the distribution to which the cross
section converges in the long run, as long as shocks to wealth are not
correlated across individuals.
Taken as the cross-section stationary distribution, the shape is very
intuitive. Households in the tails tend to hit the triggers and adjust
to the target in the interior of the distribution. Since this point can
be reached from either the triggers or the interior of the distribution,
it has the largest mass. Transitions to nearby states are then more
likely than transitions to distant states, so the density near the target
point remains high and falls as y nears the triggers.
This distribution is the stable ergodic distribution if there are no
aggregate shocks. However, if there are also aggregate shocks, there
is no single stationary distribution, and we would not expect to actu-
ally observe this shape at any point in time.41 The density would
tend toward the ergodic distribution between aggregate shocks, but
it would not be stable.42
Note that the distribution in figure 2 is calculated for a single set
of parameters, including the drift and variance of wealth; Section IV
showed cross-section variation in these values. The intent of this sec-
tion is to determine the effects of transactions costs on aggregate
durables expenditures ceteris paribus. In order to focus on this effect
in isolation from other sources of heterogeneity, I use a common
set of parameters for all the households in constructing the ergodic
distribution and in the simulations in the next subsection.
41 There would be an ergodic "distribution of distributions."

42 With continuous shocks, the distribution tends toward the ergodic when idiosyn-
cratic uncertainty is large relative to aggregate shocks. This is the "attractor effect"
noted in Caballero (1990a).

0.05
0.04 -
0.03 -
0.02 -
0.01
0
0.1 1.6 3.1 4.6 6.1 7.6 9.1 10.6 121 13.6 15.1 16.6 18.1
FIG. 2.-Ergodic distribution of WIK. The parameter values used to calculate the
ergodic distribution are given in the note to table 3.
B. The Empirical Distributionof Durables Stocks

To compare the ergodic density in figure 2 with the empirical density,
I first purge the household data of observable heterogeneity. I regress
durables stocks on a vector of demographic characteristics and the
time elapsed since the household's last purchase, and then adjust each
household's current durables stocks to reflect an average set of the
demographic characteristics.43 This measure of durables stocks then
forms the denominator of y. I then construct a kernel density estimate
of the distribution of y. Intuitively, this constructs a distribution
around each of the data points and uses them as weights in a moving
average. This systematically smooths the discrete histogram using a
weighted moving average.44
The resulting estimate is shown in figure 3 as the empirical density,
plotted together with the theoretical ergodic distribution. The empiri-
cal peak is very near the optimal adjustment point, with similar asym-
metries. A Kolmogorov-Smirnov goodness-of-fit test confirms that
this similarity cannot be statistically rejected (the p-value is .2). In
economic terms, the difference between the two distributions is more
substantial. If both distributions are subjected to a 3 percent aggre-
43The method of adjustment is described in App. B.

4'This was done using a Gaussian kernel and optimal bin widths from Silverman
(1990).

00n5
- 1983
0.04-
ERGODIC
0.03
01
0.1 1.6 3.1 4. 6.1 7. 9.1 10. 17.1 13. 15.116. 15.1 19.621.1 22.6 24.1 25.6 27.1 Z3.431.0
FIG. 3.-Cross-section distributions of WIK: 1983 kernel density estimate and er-
godic distribution.
gate shock, the share of households adjusting in the first year is 50

percent higher for the empirical distribution than for the ergodic
distribution. After 3 years, however, the accumulated share of house-
holds adjusting differs by 7.5 percent; the difference falls to 5.2 per-
cent after 5 years.45
If the cross-section density were stable, however, the same number
of households would hit the triggers each period; this would not
provide a basis for aggregate dynamics. In fact, however, the distribu-
tion does shift over time. Figure 4 shows that between 1983 and 1986,
the peak becomes more pronounced, implying that large numbers of
households adjusted their automobile stocks relative to wealth. This
is consistent with the fact that during this period real personal con-
sumption expenditures for automobiles averaged annual growth of
15 percent per year, whereas income and net worth growth averaged
3.2 and 4.1 percent, respectively (1990 EconomicReport of the President
and data from Flow of Funds Accounts, 1946-87). The 1986 distribu-
tion also differs significantly from the ergodic distribution. When
subjected to a 3 percent aggregate shock, the share of households
adjusting is 25 percent higher than that associated with the ergodic
distribution in the first year. However, accumulated adjustments
starting from the 1986 distribution are 19.7 percent and 30.4 percent
15These results were calculated using the simulation procedure and parameters
described below.

Om
FIG.4. Cros-sctin dstrbutonsof 193ad186ernldnsiyetimte1983

1986
0.01
0
0.1 1.6 3.1 4.6 6.1 7.6 9.1 10.6 12.113.6 15.1 16.6 15.1 19.621.1 22.6 24.1 25.6 27.1 28.4 31.0
FIG. 4.-Cross-section distributions of WIK: 1983 and 1986 kernel density estimates
lower than would be implied by the ergodic distribution after 3 and

5 years, respectively.
To see whether dynamics in the cross-section distribution can ac-
count for the dynamics in aggregate automobile expenditures, I use
a series of simulations starting from the 1983 observed cross section.46
I assume aggregate shocks corresponding to actual growth in real per
capita disposable income (assuming expected growth of 2 percent).
Idiosyncratic shocks have a zero mean with a standard deviation of
18.6 percent.47 I then simulate the response of the households to
these shocks and calculate the kernel density estimate corresponding
to each year. The simulated density estimate for 1986 is plotted in
figure 5 along with its empirical counterpart. The densities corre-
spond closely, suggesting that much of the shift in the empirical distri-
46 The simulation procedure takes the household cross section of y in 1983 and adds
a common shock and an idiosyncratic shock for each household. Households hitting
the upper or lower trigger are then moved to the return point, and the process is
repeated for the next year.
47 This implies that the ratio of the idiosyncratic standard deviation to the total
standard deviation is 0.85. The idiosyncratic standard deviation was obtained by assum-
ing that labor income has a standard deviation of 0.4 (see Hubbard, Skinner, and
Zeldes 1994). The average ratio of human wealth to total wealth is 0.71, and predictable
income movements account for approximately one-third of the income variation in
the sample. If the household is better able to predict income movements than the
econometrician, the idiosyncratic variance would be reduced. If, however, human
wealth is not the only source of total wealth uncertainty, then the idiosyncratic variance
would be increased.

0
U)
0
0
______ 1986 distribution
1V)/
0
To
CD G 5.te Siuaedadacul186cos-eto distributions
CDI
0
0
o -5 0 5 10 15 20 25 50 35
W/K
FIG. 5.-Simulated and actual 1986 cross-section distributions
bution from 1983 to 1986 can be explained by the increase in income

growth in the early 1980s. Over this period, rapid income growth
caused households in the upper tail of the distribution to adjust to
the target point.
Figure 6 shows the evolution of the density over the entire period.
Year 1 is 1982, and the series continues through 1988. Year 2 is the
second contour on the surface and shows little change in the distribu-
tion. Year 3 shows that the peak moves slightly toward the origin
and becomes more pronounced in response to doubling the growth
rate of income. When the growth rate returns to 2 percent in the
subsequent years, the peak stabilizes near the target point. This im-
plies that more households are adjusting each year, so aggregate
expenditure has increased in response to higher income growth.
The implied increase in aggregate expenditures is quantified in
table 5. Panel A of table 5 reports the growth in automobile expendi-
tures corresponding to the simulation above. The first row gives the
assumed income growth. The second row gives the growth in aggre-
gate expenditures, and the third and fourth rows break this into the
growth in number of buyers versus growth in average purchase.
These simulations imply that the elasticity of expenditure with respect
to income exceeds five in the short run, but the effect is depleted
within 4 years. Most of the variation occurs in the number of buyers,
but growth in the average purchase also contributes to expenditure
growth. Growth in the average purchase differs from growth in ag-

FIG. 6.-Simulated series of cross-section distributions
gregate wealth since, with idiosyncratic shocks, households receiving

"good" idiosyncratic realizations have higher wealth and are more
likely to adjust.
Panel B of table 5 reports the actual income, wealth, and expendi-
ture aggregates over the period of the simulation. The first row gives
the growth rates in real disposable income, which motivate the as-
sumptions for the simulations. The second row gives the growth in
real net worth. The third row reports the growth rate of real expendi-
tures on motor vehicles. The first row shows that real income surged
in 1984 and then tapered off to less than 1 percent by 1987. Net
worth grew faster than income in 1983 but slowed in 1984 before
increasing in 1985 and tapering off thereafter. Expenditures on mo-
tor vehicles increased over 16 percent in 1983 and 1984; this corre-
sponds to an average contemporaneous income elasticity of 5.0. The
growth rate of expenditures then fell slightly in 1985 and again in
1986, and became negative in 1987.
Qualitatively, the simulations match the actual data, with several
caveats noted below. Simulated expenditure rises sharply in the first
two years as the higher growth of income causes more households to
hit the trigger point in a given period. The growth rate of expendi-

TABLE 5
SIMULATED AND ACTUAL AGGREGATE AUTO EXPENDITURES (Percentage Changes)
A. SIMULATIONS
YEAR
1 2 3 4 5
Income 2.0 4.0 2.0 2.6 2.0
Expenditure 11.7 29.0 8.1 8.4 -7.8
Number of buyers 3.9 22.6 2.0 6.6 -.9
Average purchase 7.8 6.4 6.1 1.8 -6.9
B. AGGREGATE DATA
1983 1984 1985 1986 1987

Real disposable income 2.0 4.9 2.0 2.6 .6
Real net worth 6.5 2.4 5.8 4.1 2.5
Real auto expenditure 16.4 16.7 11.1 7.2 -4.2
C.
Frictionless prediction of 17.0 22.3 - 19.8 6.9 - 16.4

auto expenditure
NOTE.-The simulations were run starting from the observed empirical distribution of y for 1983. The optimal
return and trigger points were used from table 3. Idiosyncratic shocks have a standard deviation of 18.6 percent,
and the aggregate shocks are those to income noted in the first row. The aggregates for real per capita disposable
income and real personal consumption expenditure on motor vehicles are annual National Income and Product
Accounts data. The wealth data are nominal from the Federal ReserveBulletin (table C.9), deflated by the consumer
price index for new cars. The frictionless prediction of auto expenditure is the prediction of the optimal consump-
tion model with a representative consumer, setting the transaction cost X equal to zero.
ture declines when the growth rate of income stabilizes, since house-
holds are no longer hitting the trigger at an increasing rate. Expendi-
ture declines in year 5, when virtually all households have adjusted
to the surge in the growth rate. After this, expenditure stabilizes with
low frequency cycles. Overall, the temporary increase in the growth
of income boosts expenditure for 4 years-most strongly after the
initial shock-and then tapers off.
This is in contrast to the predictions of a model with zero transac-
tion cost. Panel C of table 5 shows that the frictionless model predicts
the immediate response to higher income growth in 1983 and 1984,
but then implies a rapid reduction in expenditure when growth slows
in 1985 and 1987. The reason is that the model implies that expendi-
ture should rise immediately in response to higher wealth but then
fall the next period since consumers will need to replace only the
previous period's depreciation.
Simulated expenditures differ from the aggregate data primarily
because the response to the aggregate shock in period 2 is "too large."

There are several possible explanations for this result. First, growth
in net worth is negatively correlated with that of income over this
period. Total wealth growth is therefore smoother than that used in
the experiment, smoothing the response of expenditure. Second, the
aggregate data also include the liquidity-constrained households. For
these households, expenditure is determined partly by the timing of
increases in income. If these increases are distributed throughout the
sample period, then the measured growth rate is less volatile than
the simulated rate, reducing the initial response and prolonging the
adjustment period. Finally, in a general equilibrium framework, the
surge in demand would produce a price response, damping initial
demand and increasing it later.
VI. Conclusions
This paper considers the automobile purchases of households facing
a transactions cost. It shows that about half the households in the
sample behave in this manner, and the others also exhibit behavior
suggestive of liquidity constraints. The decision rules of the house-
holds facing the transactions cost are not affected by the level of
income or wealth but are determined largely by their growth rates
and variability. Higher variability broadens the (S, s) bands, so dura-
bles deviate more from their optimal levels before households adjust.
Higher growth of income or wealth, on the other hand, speeds house-
holds through the bands, increasing the number of buyers per period
and stimulating aggregate expenditure. The empirical distribution of
consumers' auto stocks relative to wealth fits that predicted by theory
but is not stationary. Simulations show that the changes in the distri-
bution are consistent with observed growth in income over the period.
Further, the model explains the actual acceleration and subsequent
slowing of aggregate durables expenditure in the 1980s.
These results may provide an explanation for the persistence of
shocks in aggregate durables expenditures. With transactions costs,
disturbances to aggregate expenditures persist until all households
have adjusted their durables stocks. Those close to the trigger point
will react immediately to an aggregate disturbance, incorporating it
into their purchases. Households further away from the trigger point
will take longer to respond, incorporating the aggregate shock only
when they purchase. Thus the aggregate effect is complete only when
all households have adjusted. While theoretically these effects can be
negated if the economy is at a steady state, as in Caplin and Spulber
(1987), the cross-section evidence rejects this condition. Simulations
of the macroeconomic adjustment show that complete adjustment
should take 3-4 years, consistent with the aggregate finding of Caba-

Hero (1990a). The simulations further show that the adjustment pro-
cess is not smooth, but instead exhibits the rapid initial response and
slow tapering off observed in the aggregate data.
Further, aggregate durables expenditures respond vigorously to
changes in income relative to nondurables, even while the transac-
tions cost produces "sluggish" dynamics. An increase in the growth
of income moves households through the (S, s) bands more rapidly,
initially increasing the number of purchasers per period.48 This tem-
porarily increases aggregate expenditure, potentially producing a
large income elasticity of aggregate expenditure, although each indi-
vidual purchase is still of the optimal amount.
Finally, these aggregate effects will arise even when individual pur-
chases are consistent with life cycle-permanent income behavior. In-
dividual purchases are selected as an optimal share of wealth, but the
number of buyers also responds to innovations. The product of these
two effects produces a dichotomy between observations of a house-
hold and the aggregate, with the difference explained by the failure
of the representative agent assumption and the dynamics of the cross-
section distribution of households.
Appendix A
Derivation of the Ergodic Distribution
The derivationand implicationsof the ergodic distributionof y (--[WIK] -
X)are found in Eberly (1991). The following provides an overview.
Nondurable assets are held in a risky asset (A) and a risklessasset (B). The
risky asset has expected rate of return puand standarddeviation cr.The rate
of depreciationof the durable good is 8. Define x as AIK,one component of
y. I first derive the distributionof x and then use it to characterizethat of y.
The equation of motion governing x is
dxt= (p + B)xtdt+ axtdZt, (A1)

where dZtis the increment of a standard Brownian motion.
It is convenient in what follows to use z, defined as the natural log of x.
This variablefollows a standard Brownian motion with drift g and standard
deviation crwhen no adjustmentoccurs, so that
ZtIn xt, (A2)
dzt= gdt + crdZ.
The movement of z is regulated by the (S, s) bounds in the state space of
y. When g is large relativeto cr,the relationshipbetweenx and y is monotonic,
and therefore the bounds in the state space of y (y", y*, and y1) can be
48A jumpin the levelof incomewouldhavethe sameeffect.

transformedinto the state space of x and then to z. Call these bounds zL, Z*,
and zH. The variablez then follows a regulated Brownian motion described
by equation (A2) and the bounds.
The limiting distributionof this process is p(z):
Pr(TL< TH) exp [a2 (z ] 11
Pr(TH < TL) (zH - ZL) - (Z* - ZL) ifz < z

(P(Z)= (A3)
I - exp [
[2g
Pr(TH < TL) ( )
if Z > Z*,
Pr(TH < TL) (zH - ZL) _ (Z* _ ZL) i
where TL and TH are the hitting times for the lower and upper bounds,
respectively. Then Pr(TL < TH) is the probability of hitting the lower bound
before the upper. This distribution is piecewise exponential with a disconti-
nuity at the internal return point z*.
The ratios x and y are related by the portfolio rule of the consumer, which
determines how much of the risky asset is held according to
x(y) - r)h'(y)
(= (A4)
ur2h"f(y)
where r is the riskless rate of return and h(y) is a transformation of the
consumer's value function. Without the transaction cost, this would be the
linear portfolio rule of Merton (1969):
x(y)= (~ - r)yY (A5)

2(I- t)'
where 1 - axis the coefficient of relative risk aversion.

Using equations (A2)-(A4), I use a change of variable to characterize the
limiting distribution of y:
=P(Z)
_ (A6)
Y (Z)
This requires that y'(z) be a single-valued function. This is satisfied over the
region [y*, y'] for all values of the parameters. Over the region [yL,y*), this
is most likely to be satisfied when the drift in y is large relative to its variance.
I assume that this condition holds for automobiles, which have high deprecia-
tion and therefore high drift in y. Calculations in Eberly (1991) suggest that
the condition should be satisfied for most parameter values over the entire
state space [yL, y*].
Appendix B
Data
A. HouseholdData
The household data used are taken from the Survey of Consumer Finances,
conducted in 1983 and 1986. The 1983 survey contacts a representative

sample of noninstitutionalhouseholds, and the 1986 survey recontacts the

same households. Households that were divided between surveys are elimi-
nated from the tests in the paper. I do not utilize the high-income sample
that was also collected with these data.
The 1983 survey was conducted by personal interview.Of the 3,824 inter-
views completed, 3,665 provided sufficientinformationon income and assets
to remain in the sample. The 1986 survey recontacted 2,822 of the same
households and conducted telephone interviews. Sample weights are avail-
able in the survey to overcome potentialsample selectionbias in the recontact
process. A complete list of variables used in the paper is found in Eberly
(1993).
B. Aggregateand StateData
Statelicense fees were taken from SignificantFeaturesofFiscalFederalism
(Advi-
sory Commissionon IntergovernmentalRelations,1987 ed., table59). Aggre-
gate price indexes were taken from the 1990 EconomicReportof thePresident
(tablesC58, C59).
C. PredictedIncomefrom theConsumer
Expenditure
Survey
An estimate of the predictable component of income was obtained by re-
gressing income (before tax) on age, education and occupationdummy vari-
ables, the race and gender of the head of household, and interactionterms
for age, education, and occupation using data from the 1985 Consumer
Expenditure Survey. The sample size was 4,640 households, and the R2 of
the regression was .295. The estimatedcoefficientsfrom the regression were
then used to obtain an estimate of predictableincome for households in the
Survey of Consumer Finances. This variable is called "predictableincome"
in the text and tables.
D. Demography-Adjusted
DurablesStocks
Suppose that households differ in the stock of durables that they choose
when adjusting such that ln(Kj*) = Zjyl + ln(Wj)Y2 + vp, where Kj is the
stock chosen at adjustment, ZJincludes household characteristics,and v; is
an idiosyncraticshock. The term Zj1ycaptures observablehousehold differ-
ences in the target wealth to durables ratio. Since Kj = Kj*(l - Wj'i,where
Tj is the time since the household's last purchase, taking logs we can write
In(K1)= In(K* ) + TjIn(1 - r). Substitutein the expression for In Kj*,and
regress ln(K1)on characteristicsZj, In Wj, and Tp.The coefficient on time
since last purchase is then an estimate of ln(1 - 8), providing an implied
annual depreciation rate of .105, which is used in the numerical solutions
and simulations. The heterogeneity-adjustedmeasure of durables stocks is
exp[ln(Kj) - (Zj - Z)11], where j' is the vector of regressioncoefficientson
Z and Z is the mean of these characteristicsover households. The R2 of this
regression was .445, and the Z vector includes the following variables:
* age of head by date of birth (enters quadratic)
* sex of head

* race of household (race of respondent)
* maritalstatus of respondent
* education of head: dummy = 1 if at least some college
* occupation of head:
dummy = 1 if professional, technical, or kindred workers
dummy = 1 if managers and administrators
dummy = 1 if self-employed managers
dummy = 1 if sales, clerical, and kindred workers
dummy = 1 if crafts, protective service, and kindred workers
dummy = 1 if operatives, laborers,and service workers
* total number of persons in household
* total number of persons in household under 18
* 1982 total household income
References
Akerlof, George A. "The Market for 'Lemons': Quality Uncertainty and the
Market Mechanism." QJ.E. 84 (August 1970): 488-500.
Bar-Ilan, Avner, and Blinder, Alan S. "Consumer Durables and the Opti-
mality of Usually Doing Nothing." Working Paper no. 2488. Cambridge,
Mass.: NBER, January 1988.
Bernanke, Ben S. "Permanent Income, Liquidity, and Expenditure on Auto-
mobiles: Evidence from Panel Data." QJ.E. 99 (August 1984): 587-614.
. "Adjustment Costs, Durables, and Aggregate Consumption."J. Mone-
tary Econ. 15 (January 1985): 41-68.
Bertola, Giuseppe, and Caballero, Ricardo J. "Kinked Adjustment Costs and
Aggregate Dynamics." In NBER MacroeconomicsAnnual 1990, edited by
Olivier J. Blanchard and Stanley Fischer. Cambridge, Mass.: MIT Press
(for NBER), 1990.
Caballero, Ricardo J. "Durable Goods: An Explanation for Their Slow Ad-
justment." Manuscript. New York: Columbia Univ., Dept. Econ., 1990. (a)
. "Expenditure on Durable Goods: A Case for Slow Adjustment."
QJ.E. 105 (August 1990): 727-43. (b)
Caballero, Ricardo J., and Engel, Eduardo M. R. A. "Dynamic (S, s) Econo-
mies." Econometrica59 (November 1991): 1659-86.
Campbell, John Y., and Mankiw, N. Gregory. "Consumption, Income, and
Interest Rates: Reinterpreting the Time Series Evidence." In NBER Macro-
economicsAnnual 1989, edited by Olivier J. Blanchard and Stanley Fischer.
Cambridge, Mass.: MIT Press (for NBER), 1989.
Caplin, Andrew S., and Spulber, Daniel F. "Menu Costs and the Neutrality
of Money." Q.J.E. 102 (November 1987): 703-25.
Chah, Eun Young; Ramey, Valerie; and Starr, Ross M. "Durable Goods,
Liquidity Constraints, and Intertemporal Consumer Optimization: Theory
and Evidence." Manuscript. San Diego: Univ. California, Dept. Econ.,
1991.
Chow, Gregory C. Demandfor Automobilesin the United States: A Study in Con-
sumerDurables. Amsterdam: North-Holland, 1957.
Eberly, Janice C. "Durable Goods and Transactions Costs: Theory and Evi-
dence." Ph.D. dissertation, Massachusetts Inst. Tech., 1991.
. "Adjustment of Consumers' Durables Stocks: Evidence from Auto-

mobile Purchases." Manuscript. Philadelphia: Univ. Pennsylvania, Whar-

ton School, 1993.
Goldfeld, Stephen M., and Quandt, Richard E., eds. Studies in Nonlinear Esti-
mation. Cambridge, Mass.: Ballinger, 1976.
Grossman, Sanford J., and Laroque, Guy. "Asset Pricing and Optimal Portfo-
lio Choice in the Presence of Illiquid Durable Consumption Goods." Econo-
metrica58 (January 1990): 25-51.
Hall, Robert E. "Stochastic Implications of the Life Cycle-Permanent Income
Hypothesis: Theory and Evidence."J.P.E. 86 (December 1978): 971-87.
Hall, Robert E., and Mishkin, Frederic S. "The Sensitivity of Consumption
to Transitory Income: Estimates from Panel Data on Households." Econo-
metrica50 (March 1982): 461-81.
Hubbard, R. Glenn; Skinner, Jonathan; and Zeldes, Stephen P. "The Impor-
tance of Precautionary Motives for Explaining Individual and Aggregate
Saving." In On the State of Macroeconomics,edited by Allan H. Meltzer and
Charles Plosser. Carnegie-Rochester Conference Series on Public Policy,
vol. 40. Amsterdam: North-Holland, 1994 (in press).
Lam, Pok-sang. "Permanent Income, Liquidity, and Adjustments of Automo-
bile Stocks: Evidence from Panel Data." QJ.E. 106 (February 1991):
203-30.
Maddala, G. S. "Disequilibrium, Self-Selection, and Switching Models." In
Handbook of Econometrics,vol. 3, edited by Zvi Griliches and Michael D.
Intriligator. New York: North-Holland, 1986.
Mankiw, N. Gregory. "Hall's Consumption Hypothesis and Durable Goods."
J. MonetaryEcon. 10 (November 1982): 417-25.
Merton, Robert C. "Lifetime Portfolio Selection under Uncertainty: The
Continuous-Time Case." Rev. Econ. and Statis. 51 (August 1969): 247-57.
R. G. Ibbotson Associates. Stocks,Bonds, Bills, and Inflation: Yearbook.Chicago:
Ibbotson Assocs. and Capital Market Res. Center, 1989.
Romer, Christina D. "The Great Crash and the Onset of the Great Depres-
sion." QJ.E. 105 (August 1990): 597-624.
Samuelson, Paul A. "Lifetime Portfolio Selection by Dynamic Stochastic Pro-
gramming." Rev. Econ. and Statis. 51 (August 1969): 239-46.
Silverman, B. W. Density Estimationfor Statistics and Data Analysis. London:
Chapman and Hall, 1990.
Tsiddon, Daniel. "On the Stubbornness of Sticky Prices." Internat. Econ. Rev.
32 (February 1991): 69-75.
Zeldes, Stephen P. "Consumption and Liquidity Constraints: An Empirical
Investigation."J.P.E. 97 (April 1989): 305-46.


Eberly (1994) PDF

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Eberly (1994) PDF

Загружено:

Авторское право:

Доступные форматы

Adjustment of Consumers' Durables Stocks: Evidence from Automobile Purchases

Author(s): Janice C. Eberly

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

This paper tests an optimal (S, s) rule in household durable pur-

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

Et = Et- -(1 - 8)t- + Et, (1)

where Et is expenditure at time t, 8 is the depreciation rate of durable

1Bernanke estimates an adjustment cost parameter of 35 percent and finds that

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

II. Theoretical Foundation: Durables Purchases

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

max Ef U(Kt)eY-tdt} (2)

subject to the dynamic budget constraint

A constant proportion of wealth is held in the risky asset and in

6 The case of (x = 0 (logarithmic utility) can be addressed along similar lines.

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

chase, T- is the moment before, and X is the proportional transaction

in ( I) n (f) = L,TV (7)

where tLT is the realized rate of return on wealth between L and 7.

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

of durables purchases will be unpredictable to the extent that wealth

where r is a constant that depends on the parameters governing

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

interpretation as an expectational error. This result requires much

III. Separating Liquidity-Constrained

A. EmpiricallyDistinguishing Binding Liquidity-

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

14 This is the strategy employed by Zeldes (1989).

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

measure. This subsample selects out infrequent purchasers as well as

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

A in K(L, r) = Po + PlY + R2Ademand + e

B. Endogenous Switching Estimation

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

an automobile. Finally, this splitting method a priori identifies less

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

households may be due to liquidity constraints or to the failure of

IV. Behavior of Consumers Facing a

A. Location of the (S, s) Bands

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

trigger point, and the observation afterward provides the target

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

which includes both financial and unobservable human wealth. Net

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

STATE VARIABLE TARGET: y* yH yL

Empirical 7.7 16.9 7.6

The lower panel of table 3 indicates observed band width. Theory

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

by a factor of 2.8 at adjustment, which is closer to the estimates in

B. Determinantsof Band Width

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

where Ldum is a dummy for downward adjustment of the durables

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

(S, s) Households Constrained Households

income has a marginally significant effect, but both are insignificant

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

V. Distribution of Durables Stocks Relative

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM

A. The Ergodic Distributionof Durables Relative

41 There would be an ergodic "distribution of distributions."

This content downloaded from 146.155.94.33 on Fri, 4 Oct 2013 18:18:48 PM