Journal of Monetary Economics 16 (1985) 309-327.

North-Holland

INDIVISIBLE

LABOR AND THE BUSINESS CYCLE

Gary D. HANSEN*

Unioersify

of California,

Santa

Barbara,

CA 93104,

USA

A growth model with shocks to technology is studied. Labor is indivisible, so all variability in
hours worked is due to fluctuations in the number employed. We find that, unlike previous
equilibrium models of the business cycle, this economy displays large fluctuations in hours worked
and relatively small fluctuations in productivity. This finding is independent of individuals
willingness to substitute leisure across time. This and other findings are the result of studying and
comparing summary statistics describing this economy, an economy with divisible labor, and
post-war U.S. time series.

1. Introduction
Equilibrium theories of the business cycle, such as Kydland and Prescott
(1982) or Lucas (1977), have been criticized for failing to account for some
important labor market phenomena. These include the existence of unemployed workers, fluctuations in the rate of unemployment, and the observation
that fluctuations in hours worked are large relative to productivity fluctuations.
Equilibrium models have also been criticized for depending too heavily on the
willingness of individuals to substitute leisure across time in response to wage
or interest rate changes when accounting for the last observation. This criticism
is based at least partially on the fact that micro studies using panel data on
hours worked by individuals have not detected the intertemporal substitution
necessary to explain the large aggregate fluctuations in hours worked [see
Ashenfelter (1984)].
In this paper, a simple one-sector stochastic growth model with shocks to
technology is constructed in which there is high variability in the number
employed and total hours worked even though individuals are relatively
unwilling to substitute leisure across time. The model differs from similar
models, such as Kydland and Prescott (1982), in that a non-convexity (indivisible labor) is introduced. Indivisible labor is modeled by assuming that individ*This paper is part of my doctoral dissertation written while a student at the University of
Minnesota. I have benefited from conversations with many people including Robert King, Thomas
Sargent, Christopher Sims, Neil Wallace, Sumru Altug, Patrick Kehoe, Ramon Matimon, Ian Bain,
and Rody Manuelli. I owe my greatest debt, however, to my advisor, Edward Prescott. I wish to
also acknowledge the Federal Reserve Bank of Minneapolis which has provided support for this
research. All errors, of course, are mine.
0304-3923/85/.%3.3OSl985,

Elsevier Science Publishers B.V. (North-Holland)

310

G. D. Hunsen.

Indivisible

lubor und the bminess

qde

uals can either work some given positive number of hours or not at all - they
are unable to work an intermediate number of hours. This assumption is
motivated by the observation that most people either work full time or not at
all. Therefore, in my model, fluctuations in aggregate hours are the result of
individuals entering and leaving employment rather than continuously employed individuals adjusting the number of hours worked, as in previous
equilibrium models. This is consistent with an important feature of U.S.
post-war data: most fluctuation in aggregate hours worked is due to fluctuation
in the number employed as opposed to fluctuation in hours per employed
worker. This is a fact that previous equilibrium theories have not tried to
account for.
Existing equilibrium models have also failed to account for large fluctuations
in hours worked along with relatively small fluctuations in productivity (or the
real wage). Prescott (1983), for example, finds that for quarterly U.S. time
series, hours worked fluctuates about twice as much (in percentage terms) as
productivity. In this paper it is shown that an economy with indivisible labor
exhibits very large fluctuations in hours worked relative to productivity. This
stands in marked contrast to an otherwise identical economy that lacks this
non-convexity. In this economy hours worked fluctuates about the same
amount as productivity.2
Equilibrium theories of the business cycle have typically depended heavily
on intertemporal substitution of leisure to account for aggregate fluctuations in
hours worked. The willingness of individuals to substitute intertemporally is
measured by the elasticity of substitution between leisure in different time
periods implied by an individuals utility function. However, the theory developed here is able to account for large aggregate fluctuations in hours worked
relative to productivity without requiring that this elasticity be large. This
follows because the utility function of the representative agent in our model
implies an elasticity of substitution between leisure in different periods that is
inflnite.4 This result does not depend on the elasticity of substitution implied
by the preferences of the individuals who populate the economy. Thus, the
theory presented here is in principle consistent with the low estimates of this
elasticity found from studying panel data [see Altonji (1984) or MaCurdy
(1981)].
The fact that existing equilibrium models are inconsistent with this observation has been
stressed by Heckman (1983) and Coleman (1984).
Kydland and Prescott (1982) attempt to explain the above fact by including past leisure as an
argument in the individuals utility function so as to enhance the intertemporal substitution
response to a productivity shock. However, even after introducing this feature, Kydland and
Prescott were still unable to account for this observation.
3This is true for the technology shock theories, such as Kydland and Prescotts (1982). as well as
the monetary shock theories of Lucas and Barre [see Lucas (1977)].
41n this model there is a crucial distinction between the utility function of the representative
agent and the utility function of an individual or household.

G. D. Hansen,

Indivisible

hbor

and the business

cycle

311

The paper is divided as follows: The next section provides a more detailed
explanation and motivation of the indivisible labor assumption. In section 3
the artificial economies to be studied are constructed. The first is a standard
stochastic growth model where labor is divisible, and the second introduces
indivisible labor to that economy. The second economy is a stochastic growth
version of a static general equilibrium model developed by Rogerson (1984).
Lotteries are added to the consumption set (following Rogerson) which makes
it possible to study a competitive equilibrium by solving a representative agent
problem, as in Lucas and Prescott (1971). The addition of the lotteries also
implies that the firm is providing full unemployment insurance to the workers.
The fourth section explains how the equilibrium decision rules and laws of
motion are calculated, as well as how the parameter values used when
simulating the model were chosen. Since the representative agents problem is
not one for which a closed form solution is available, in order to calculate
decision rules a quadratic approximation of this problem is derived using the
method described in Kydland and Prescott (1982). These equilibrium decision
rules are a set of stochastic difference equations from which the statistical
properties of the time series generated by the artificial economies can be
determined. The statistics studied are a set of standard deviations and correlations discussed in section 5. In this section, the statistics computed using the
artificial time series are compared to the same statistics computed using U.S.
time series. Some concluding remarks are contained in section 6.
2. Motivation

Existing equilibrium theories of the business cycle analyze individuals who

are free to adjust continuously the number of hours worked (the intensive
margin) and who are always employed. There are no individuals entering or
leaving employment (the extensive margin). However, the extensive margin
seems important for explaining some aspects of labor supply at both the micro
and macro levels. Heckman and MaCurdy (1980) for example, discuss the
importance of the extensive margin for explaining female labor supply. At the
aggregate level, over half of the variation in total hours worked is due to
variation in the number of individuals employed rather than variation in
average hours worked by those employed. Consider the following decomposition of variance involving quarterly data:
var(logH,)=var(logh,)+var(logN,)+2cov(logh,,logN,),
where H, is total hours worked, h, is average hours worked, and N, is the
number of individuals at work, where all variables are deviations from trend.
The data used for this analysis is available from the Bureau of Labor Statistics Labstat data
tape. The series I used were collected from households using the Current Population Survey. For a
description of the detrending method, see footnote 18.

312

G. D. Hansen.

Indioisihle

l&or

and rhe business

cycle

Using this decomposition, 55% of the variance of H, is due to variation in N,,

while only 20% of this variance can be directly attributed to h,. The remainder
is due to the covariance term.6
Most people either work full time or not at all. This might be ascribed to the
presence of non-convexities either in individual preferences for leisure or in the
technology. For example, the technology may be such that the marginal
productivity of an individuals work effort is increasing during the first part of
the workday or workweek, and then decreasing later on. That is, the individual
faces a production function which is convex at first and then becomes concave.
This could be due to individuals requiring a certain amount of warm up time
before becoming fully productive. Such a technology could induce individuals
to work a lot or not at all.
Another possibility is that the non-convexity is a property of individuals
preferences. If the utility function exhibited decreasing marginal utility of
leisure at low levels of leisure and increasing marginal utility at higher levels,
individuals would tend to choose a low level of leisure (work a lot) or use their
entire time endowment as leisure (not work at all). These preferences may be
interpreted as indirect preferences which reflect costs associated with working
each period, such as driving a long distance to work or enduring the hassle of
putting on a suit and tie. Bearing these hxed costs makes an individual less
likely to choose to work only half a day.
In this paper the non-convexity is assumed to be a property of preferences.7
However, to make the model tractable, the non-convexity introduced - indivisible labor - is an extreme version of the non-convexity described above.
Individuals are assumed to have preferences that are defined only at two levels
of leisure - one level corresponding to working full time and the other corresponding to not working at all. This is modeled by assuming that the
consumption possibilities set consists of only two levels of leisure. This
assumption implies that an individual can only adjust along the extensive
margin.
Of course fluctuations along both the extensive and intensive margins are
observed in the actual economy, as the above evidence indicates. However, by
studying two economies - one that exhibits fluctuations only along the intensive margin and another with fluctuations only along the extensive margin - we
can determine the importance of non-convexities for explaining labor variability in business cycles. If it turns out that both economies exhibit the same
cyclical behavior, then it seems likely that a model that incorporated both
margins would also exhibit similar behavior. In fact, non-convexities of this
6Coleman (1984) comes to a similar conclusion using establishment data.
One advantage of modeling the non-convexity as a feature of the technology is that it would
likely explain why part-time workers are paid less than full-time workers, in addition to accounting
for features of the data discussed in this paper.

G. D. Hansen,

Indivisible

labor and rile business

cycle

313

sort could probably be safely abstracted from when studying business cycle
phenomena. However, it happens that the two models have very different
implications and that the non-convexity improves our ability to account for
U.S. aggregate time series data.
3. Two economies
3.1. A one-sector

stochastic growth model with divisible labor

The economy to be studied is populated by a continuum of identical

infinitely lived households with names on the closed interval [0, 11. There is a
single firm with access to a technology described by a standard Cobb-Douglas
production function of the form
jh,

k,, h,) =h,k:hf-,

where labor (h,) and accumulated capital (k,) are the inputs and h, is a
random shock which follows a stochastic process to be described below. Agents
are assumed to observe h, before making any period t decisions. The assumption of one firm is made for convenience. Since the technology displays
constant returns to scale - implying that firms make zero profit in
equilibrium - the economy would behave the same if there were many firms.
Output, which is produced by the firm and sold to the households, can either
be consumed (c,) or invested (i,), so the following constraint must be satisfied:
c, + i, sf(A,,

k,, h,).

(2)

The law of motion for the capital stock is given by

k ,+*=(l-6)k,+i,,

0<6_<1,

(3)

where 6 is the rate of capital depreciation. The stock of capital is owned by the
households who sell capital services to the firm.
The technology shock is assumed to follow a first-order Markov process. In
particular, h, obeys the following law of motion:
h r+l

=Yh,+Et+lr

(4)

where the E,S are iid with distribution function F. This distribution is assumed
to have a positive support with a finite upper bound, which guarantees that
output will always be positive. By requiring F to have mean 1 - y, the
unconditional mean of A, is equal to 1.
This technology shock is motivated by the fact that in post-war U.S. time
series there are changes in output (GNP) that can not be accounted for by

314

changes in the inputs (capital and labor). We follow Solow (1957) and Kydland
and Prescott (1982) in interpreting this residual as reflecting shocks to technol%Y*
Households in this economy maximize the expected value of C$J3u( c,, I,),
where 0 < fi < 1 is the discount factor and c, and I, are consumption and
leisure in period t, respectively. The endowment of time is normalized to be
one, so I, = 1 - h,. Utility in period f is given by the function
A>O.

u(c,,I,)=logc,+AlogI,,

(5)

We now have a complete specification of the preferences, technology, and

stochastic structure of a simple economy where individuals are able to supply
any level of employment in the interval [O,l]. Each period three commodities
are traded: the composite output commodity, labor, and capital services. It is
possible to consider only this sequence of spot markets since there is no
demand for intertemporal risk sharing which might exist if households were
heterogeneous.
Households solve the following problem, where w, is the wage rate at time t
and r, is the rental rate of capital:
maxEfflu(c,,l
r=O
subject to
c,+i,~

w,h,+r,k,

-h,),

and

given k, and A,,

(6)

(3).

Agents are assumed to make period t decisions based on all information

available at time t (which includes r, and w,). They have rational expectations
in that their forecasts of future wages and rental rates are the same as those
implied by the equilibrium laws of motion. The first-order conditions for the
firms profit maximization problem imply that the wage and rental rate each
period are equal to the marginal productivity of labor and capital, respectively.
Since there are no externalities or other distortions present in this economy,
the equal-weight Pareto optimum can be supported as a competitive equilibrium. Since agents are homogeneous, the equal-weight Pareto optimum is
the solution to the problem of maximizing the expected welfare of the
representative agent subject to technology constraints. This problem is the
following:
maxEfPu(c,,l

-h,),

given k, and h,,

r-0

subject to
(l)-(4)

and

E, - c.d.f. F.

(7)

G. D. Hunsen.

Indivisible

bhor

and the business

qcle

315

The state of the economy in period t is described by k, and A,. The decision
variables are h,, c,, and i,.
This problem can be solved using dynamic programming techniques. This
requires finding the unique continuous function V: S + R (where S is the state
space) that satisfies Bellmans equation (primes denote next period values)
V(k,h)=max(u(c,l-h)+/3E[V(k,A)]h]},

03)

where the maximization is over c and h and is subject to the same constraints
as (7). The value function, V(k, A), is the maximum obtainable expected return
over all feasible plans. It turns out that since the utility function is concave and
the constraint set convex, the value function is also concave. This implies that
the problem (8) is a standard finite-dimensional concave programming problem.
Unfortunately,
this problem is not one which can be solved analytically.
There is no known explicit functional form for the value function, V. In
principle this problem could be solved using numerical methods [see Bertsekas
(1976)], but a cheaper method - which does enable one to solve for closed
form decision rules - is to approximate this problem by one which consists of
a quadratic objective and linear constraints, as in Kydland and Prescott (1982).
This method will be explained briefly in section 4.
3.2. An economy with indivisible labor

The assumption of indivisible labor will now be added to the above

stochastic growth model. This will give rise to an economy where all variation
in the labor input reflects adjustment along the extensive margin. This differs
from the economy described above where all variation in the labor input
reflects adjustment along the intensive margin. In addition, the utility function
of the representative agent for this economy will imply an elasticity of
substitution between leisure in different periods that is infinite and independent of the elasticity implied by the utility function of the individual households.
Indivisibility of labor is modeled by restricting the consumption possibilities
set so that individuals can either work full time, denoted by h,, or not at all9
For
(1984).

a detailed

presentation

of dynamic

programming

methods,

see Lucas,

Prescott

and Stokey

This
is consistent
with the interpretation
given in section 2. An alternative
interpretation
of
indivisible
labor assumes that households
can work one of two possible (non-zero)
number
of
hours,
/I, or /I,. This interpretation
is consistent
with an environment
where each household
consists of two individuals,
at least one of whom works at all times. When only one member works,
the household
is working
h, hours, and when both members
work the household
is working
11s
hours.

316

G. D. Hunsetr.

lodivisible

lubor

atrd the business

cycle

In order to guarantee [using Theorem 2 of Debreu (1954)] that the solution of

the representative agents problem can be supported as a competitive equilibrium, it is necessary that the consumption possibilities set be convex.
However, if one of the commodities traded is hours worked (as in the above
model), the consumption possibilities set will be non-convex. To circumvent
this problem, we convexify the consumption possibilities set by requiring
individuals to choose lotteries rather than hours worked, following Rogerson
(1984). Thus, each period, instead of choosing manhours, households choose
a probability of working, (Y,.l1 A lottery then determines whether or not the
household actually works. After changing the economy in this manner, we
make it possible for the competitive equilibrium to be derived by solving a
concave programming problem, just as for the economy with divisible labor.
The new commodity being introduced is a contract between the firm and a
household that commits the household to work ho hours with probability (Y,.
The contract itself is being traded, so the household gets paid whether it works
or not. Therefore, the firm is providing complete unemployment insurance to
the workers. Since all households are identical, all will choose the same
contract - that is, the same LY,. However, although households are ex ante
identical, they will differ ex post depending on the outcome of the lottery: a
fraction ~1,of the continuum of households will work and the rest will not.*
Using (5), expected utility in period t is given by cy,(log c, + A log(1 - ho)) +
(1 - a,)(log c, + Alogl). l3 This simplifies to the following function U: [w +

X[O,ll+R
U(c,,ol,)=logc,+Aa,log(l-ho).

(9)

In
Rogersons
paper, a static economy
with indivisible
labor is studied
and lotteries
are
introduced
to solve the problem
introduced
by this non-convexity.
Readers may wish to consult
Rogersons
paper for a rigorous general equilibrium
formulation
of this type of model.
Adding
lotteries
to the consumption
set increases
the choices available
to households
when
labor is indivisible.
If lotteries were not available, households
would only be able to choose to not
work (corresponding
to a = 0) or to work ho (corresponding
to a = 1). Therefore,
adding lotteries
can only make individuals
better off.
*The
lottery
involves
drawing
a realization
of a random
distribution
on [O,l]. Each individual
i E [0, l] is now renamed
x,(i,z)=i+r,
if i+z,Sl,
= i + z, - 1 otherwise.
The amount

worked
h,(x)

by agent x in period t is equal

if x,(i.z)ll-a,,
if s,(i,z)>
l-a,.

variable
according

z, from
the uniform
to the following
rule:

to

=o
= ha

This provides
a mechanism
for dividing
the continuum
of agents into two subsets, one where each
individual
works zero hours and another where individuals
work ho. The first will have measure
(1 - a,) and the other measure a,. This follows from the easily verified fact that Prob[s,(
i, z) 5
1 - a,] is equal to 1 - a, for each i.
This
uses the fact that, since preferences
are separable
in consumption
and leisure,
the
consumption
level chosen in equilibrium
is independent
of whether
the individual
works or not.

G. D. Hansen,

Indivisible

labor and Ihe business

qcle

317

Since a fraction OL,of households will work h, and the rest will work zero,
per capita hours worked in period t is given by

The other features of this economy are exactly the same as for the economy
with divisible labor. These include the technology and the description of the
stochastic process for the technology shock. These features are described by
eqs. (1) through (4).
Firms in the economy, as in the previous economy, will want to employ
labor up to the point where f,,(h,, k,, h,) = w,. However, due to the fact that
lottery contracts are being traded, households are not paid for the time they
actually spend working, but are instead paid for the expected amount of time
spent working. This implies that each worker is paid as if he worked h, [as
defined in (lo)] rather than for the amount he actually does work. Therefore,
the budget constraint of a typical household differs from the budget constraint
for the economy where labor is divisible (6) and is given by
c, + i, 4 wp,h, + r,k,.

(11)

Thus, the problem solved by a typical household is

maxEfPU(c,,a,),

given k, and A,,,

02)

r-0

subject to
(11) and (3).
This problem is equivalent to the problem solved by households in a slightly
different economy where agents trade man-hours and actuarially fair insurance
contracts, rather than the type of contracts traded in the economy studied here.
In this alternative economy, which is described in more detail in the appendix,
households only get paid for the time they actually spend working. However, if
a household has purchased unemployment insurance, it will receive compensation if the lottery determines that the household does not work. In the
appendix it is shown that households will choose to insure themselves fully.
Therefore, in equilibrium, the households will have full unemployment insurance, just like the households populating the economy described in this
section. This implies that the equilibrium allocations for these two economies
are the same.

The following is the representative agents problem that must be solved to

derive the equilibrium decision rules and laws of motion:
m

maxEx/3U(c,,(Y,),
r-o
subject to
(l)-(4),

(10)

given k, and A,,

03)

and E, - c.d.f. F.

Like problem (7), this is a standard concave discounted dynamic programming

problem. The state of the economy in period I is described by k, and X,. The
decision variables are (Y,, c,, and i,.
A key property of this economy is that the elasticity of substitution between
leisure in different periods for the representative agent is infinite. To understand this result, first substitute h, = 1 - I, into (10) and solve for (Y,. After
substituting this expression for a, into (9) one obtains the following utility
function for the representative agent (ignoring the constant term):
U(c,,I,)=logc,+BI,,

(14)

where B = - A(log(1 - h,))/h,. Since this utility function is linear in leisure it

implies an infinite elasticity of substitution between leisure in different periods.
This follows no matter how small this elasticity is for the individuals populating the economy. Therefore, the elasticity of substitution between leisure in
different periods for the aggregate economy is infinite and independent of the
willingness of individuals to substitute leisure across time.14
4. Solution method and calibration

The problems (7) and (13) are not in the class of problems for which it is
possible to solve analytically for decision rules. This special class of problems
includes those with quadratic objectives and linear constraints, as well as some
other structures. For this reason, approximate economies are studied for which
the representative agents problem is linear-quadratic [see Kydland and
Prescott (1982)]. It is then possible to obtain explicit decision rules for these
approximate economies.
By making appropriate substitutions, one can express problems (7) and (13)
as dynamic optimization problems with decision variables i, and h, and state
variables X, and k,. The constraints for these problems are linear although the
14The fact that in this type of model the representative agents utility function is linear in leisure
was originally shown by Rogerson (1984) for his model. This result depends, however, on the
utility function being additively separable across time.

G. D. Hansen.

Indivisible

labor and the bwiness

cycle

319

objective functions are non-linear. For each of these problems, Kydland and
Prescotts procedure is used to construct a quadratic approximation of the
objective function to be accurate in a neighborhood of the steady state for the
appropriate model after the technology shock has been set equal to its
unconditional mean of 0ne.l The reader may consult Kydland and Prescott
(1982) for details on the algorithm used for forming these approximations.6
To actually compute these quadratic approximations, solve for an equilibrium, and generate artificial time series, it is necessary to choose a distribution function, F, and specific parameter values for 19, 6, /3, A, y, and he.
Kydland and Prescott (1982,1984) follow a methodology for choosing parameter values based on evidence from growth observations and micro studies. This
methodology will also be followed here. In fact, since they study a similar
economy, some of the above parameters (19,8, /3) also appear in their model.
This enables me to draw on their work in selecting values for these parameters,
thereby making it easier to compare the results of the two studies.
The parameter 8 corresponds to capitals share in production. This has been
calculated using U.S. time series data by Kydland and Prescott (1982,1984)
and was found to be approximately 0.36. The rate of depreciation of capital, 6,
is set equal to 0.025 which implies an annual rate of depreciation of 10 percent.
Kydland and Prescott found this to be a good compromise given that different
types of capital depreciate at different rates. The discount factor, /3, is set equal
to 0.99, which implies a steady state annual real rate of interest of four percent.
The parameter A in the utility function (5) is set equal to 2. This implies that
hours worked in the steady state for the model with divisible labor is close to
l/3. This more or less matches the observation that individuals spend l/3 of
15Let the steady states for the certainty version of these models be denoted by the variables
symbol without any subscript. Eq. (3) implies that investment in the steady state is given by i = 6k.
Expressions for k and h can be determined by deriving the Euler equations for the appropriate
representative agent problem and setting h, = h, k, = k, and i, = i= Sk for all f. For both
economies, the steady state capital stock is given by
k= [(p+S)/O]-
h where p=(l/p)-1.
Hours worked in the steady state for the economy with divisible labor is given by h = (1 - 6) X
(p + S)/(3(p + 6) - f?(p + 36)]; and for the economy with indivisible labor, h = (1 - O)(p + S)/
[#(p + 6 - M)] where $= -A[log(l - h,)]/h,.
l6 Kydland and Prescotts method for approximating this problem requires choosing a vector of
average deviations, z E R4, which determines the size of the neighborhood around the steady state
within which the approximation is accurate. The four components of z are average deviations from
trend of the four variables, x, = (X,, k,, i,, h,). as found in U.S. time series data. This implies that
along those dimensions where there is more variability, the approximation will be accurate in a
larger neighborhood around the steady state (.Y). For the exercise carried out in this paper
( 'i/jzt
I:- 1 = (0.012.0.006,0.08,0.017), reflecting the average standard deviations of these series as
reported in the next section. Although attention was paid to specifying this vector in a reasonable
way, it turns out that the results are not altered when the zi components are decreased by a factor
of ten.

320

G. D. Hansen,

Indivisible

labor

and fire bminess

cycle

their time engaged in market activities and 2/3 of their time in non-market
activities.
To determine the parameter h,, I set the expressions for hours of work in
the steady state for the two models equal to each other. Since steady state
hours worked in the model with divisible labor is fully determined by the
parameters 8, 6, A, and fi for which values have already been assigned (see
footnote 15) it is possible to solve for ha. This implies a value for h, of 0.53.
The distribution function F along with the parameter y determine the
properties of the technology shock, h,. The distribution of E, is assumed to be
log normal with mean (1 - y), which implies that the unconditional mean of X,
is 1. The parameter y is set equal to 0.95 which is consistent with the statistical
properties of the production function residual. The standard deviation of
E,, us, is difficult to measure from available data since this number is significantly affected by measurement error. A data analysis suggests that u, could
reasonably be expected to lie in the interval [0.007,0.01]. A value of 0.007, for
example, would imply that a little over half of the variability in E, is being
attributed to measurement error, which is probably not unreasonable. The
actual value used for the simulations in this paper is 0.00712. This particular
value was chosen because it implies that the mean standard deviation of output
for the economy with indivisible labor is equal to the standard deviation of
GNP for the U.S. economy (see next section).
All parameters of the two models have now been determined. We are now
ready to study and compare the statistical properties of the time series
generated by these two models.
5. Results

For the purposes of this study, the statistical properties of the economies
studied are summarized by a set of standard deviations and correlations with
output that are reported in table 1.
The statistics for the U.S. economy are reported in the first two columns of
the table. Before these statistics were calculated, the time series were logged
and deviations from trend were computed. Detrending was necessary because
the models studied abstract from growth. The data were logged so that
standard deviations can be interpreted as mean percentage deviations from
The production, function residual is measured, using U.S. time series, by
logX,-logy,-elogk,-(1-e)logh,,
where data on GNP, capital stock (nonresidential equipment and structures). and hours worked is
obtained from a standard econometric data base. The first-order autocorrelation coefficient for h,
is about 0.95, indicating high serial correlation in this series. The parameter 0 was assumed to be
equal to 0.36 for calculating this residual. A more detailed study of the statistical properties of this
technology shock is planned but has not yet been carried out.

G. D. Hansen,

Indivisible

labor

Table
Standard

deviations

in

Quarterly

Scrics
OulpuI
Consumption
lnvestmcnl
Capital stock
Hours
Productivily

percent

(a)

and

(a)

(b)
1.00
0.85
0.92
0.04
0.76
0.42

321

with output (II) for

Economy with
divisible labo?
(a
1.35
0.42
4.24
0.36
0.70
0.68

cycle

correlations
economies.

U.S. rime series

(55.3-84.1)

1.76
1.29
8.60
0.63
1.66
1.18

and the business

(0.16)
(0.06)
(0.51)
(0.07)
(0.08)
(0.0X)

and

artificial

Economy with
indivisible laborh
(b)

1.00
0.89
0.99
0.06
0.98
0.98

U.S.

(0.00)
(0.03)
(0.00)
(0.07)
(0.01)
(0.01)

(a)
1.76
0.51
5.71
0.47
1.35
0.50

(b)

(0.21)
(0.08)
(0.70)
(0.10)
(0.16)
(0.07)

1.00
0.87
0.99
0.05
0.98
0.87

(0.00)
(0.04)
(0.00)
(0.07)
(0.01)
(0.03)

The U.S. time series used are real GNP. total consumption
expenditures.
and gross private domestic
investment (all in 1972 dollars). The capital stock series includes nonresidential
equipment and structures.
The hours series includes total hours for persons 81 work in non-agricultural
industries as derived from the
CWWW Popctluriott Sun~,r. Productivity
is output divided by hours. All series arc seasonally adjusted.
logged and dctrcndcd.
hThc standard deviations and correlations with output arc sample mcnns of statistics computed lor each
of 100 simulations.
Each simulation consists of 115 periods. which is the same number of periods as the
U.S. sample. The numbers in parcnthcscs
arc sample standard deviations
of these statistics.
Before
computing
any statistics each simulated time series was logged and dctrcndcd
using the same procedure
used lor the U.S. time series.

trend. The detrending procedure used is the method employed by Hodrick

and Prescott (1980).*
Since much of the discussion in this section centers on the variability of
hours worked and productivity (output divided by hours worked), some
discussion of the hours series is appropriate. The time series for hours worked
used in constructing these statistics is derived from the Current Population
Survey, which is a survey of households. This series was chosen in preference to
the other available hours series which is derived from the establishment stiey.
The hours series based on the household survey is more comprehensive than
This
following

method

involves

choosing

smoothed

values

( s, $. 1 for the series ( x,}T-

1 which

solve the

-_problem:

min (l/r~~(r.-.,)+(h/r)~~~[(2*I-~,)-(+-4-~)12),
t-1
I
where X > 0 is the penalty on variation,
where variation
is measured
by the average squared second
difference.
A larger value of X implies that the resulting
(3,) series is smoother.
FollowGig
Prescott
(1983).
I choose
A= 1600. Deviations
from the smooth
series are form&l
by taking
d, - x, -s,.

This method is used in order to filter out low frequency

fluctuations.
Although
other methods
(spectral
techniques,
for example)
are available.
this method was chosen because of its simplicity
and the fact that other methods lead to basically
the same results [see Prescott
(1983)].

322

G. D. Hanserl,

Indivisible

lubar

and the business

cycle

the establishment series since self-employed workers and unpaid workers in

family-operated
enterprises are included. Another advantage is that the
household series takes into account only hours actually worked rather than all
hours paid for. That is, it doesnt include items such as paid sick leave. A
disadvantage is that the household series begins in the third quarter of 1955,
which prevented me from using data over the entire post-war period.
Sample distributions of the summary statistics describing the behavior of the
artificial economies were derived using Monte Carlo methods. The model was
simulated repeatedly to obtain many samples of artificially generated time
series. Each sample generated had the same number of periods (115) as the
U.S. time series used in the study. Before any statistics were computed, the
data were logged and the same filtering procedure applied to the U.S. data was
applied to these time series. One hundred simulations were performed and
sample statistics were calculated for each data set generated. The sample
means and standard deviations of these summary statistics are reported in the
last four columns of table 1.
When comparing the statistics describing the two artificial economies, one
discovers that the economy with indivisible labor displays significantly larger
fluctuations than the economy with divisible labor. This shows that indivisible
labor increases the volatility of the stochastic growth model for a given
stochastic process for the technology shock. In fact, it is necessary to increase
uc by 30 percent (from 0.00712 to 0.00929) in order to increase the standard
deviation of output for the divisible labor economy so that it is equal to the
standard deviation of GNP for the actual economy, which is 1.76. It is still the
case that 0.00929 is in the interval suggested by the data (see paragraph on
measuring crcin the previous section). However, since it is likely that there is
significant measurement error in our empirical estimate of the production
function residual, one should prefer the lower value of Us.
Another conclusion drawn from studying this table is that the fluctuations in
most variables are larger for the actual economy than for the indivisible labor
economy. It is my view that most of this additional fluctuation (except in the
case of the consumption series) is due to measurement error. Work in progress
by the author attempts to correct for measurement error in the hours series
(and hence some of the measurement error in the productivity series).
Preliminary findings seem to suggest that the above hypothesis is correct. In
addition, the fact that the consumption series fluctuates much more in the
actual economy than in the artificial economy can probably be explained by
the fact that nothing corresponding to consumer durables is modeled in the
economies studied here.
The work referred to is a chapter of my dissertation. Copies will soon be available upon
request.

G.D.

Hansen,

Indivisible

labor and the business

cycle

323

Perhaps the most significant discovery made by examining table 1 is that the
amount of variability in hours worked relative to variability in productivity
is
very different for the two model economies. This relative variability can be
measured by the ratio of the standard deviation in hours worked to the
standard deviation in productivity. For the economy with indivisible labor, this
ratio is 2.7, and for the economy without this feature the ratio is not
significantly above 1. 2o For the U.S. economy the ratio is equal to 1.4, which is
between these two values.
As explained in the introduction, accounting for the large variability in
hours worked relative to productivity has been an open problem in equilibrium
business cycle theory. Kydland and Prescott (1982) study a version of the
stochastic growth model where labor is divisible and the utility function of
individuals is non-time-separable with respect to leisure. This non-time-separability property is introduced to make leisure in different periods better
substitutes. However, this feature enables these authors to report a value for
this ratio of only 1.17, which is still much too low to account for the
fluctuations found in U.S. data.
On the other hand, the economy with indivisible labor studied here has
exactly the opposite problem Kyland and Prescotts model has. The ratio
implied by this model is much larger than the ratio implied by the data.
However, this should not be surprising. In fact, it would be bothersome if this
were nor the case. After all, we do observe some adjustment along the intensive
margin in the real world. Examples include workers who work overtime in
some periods and not in others or salesmen who work a different number of
hours each day. Since indivisible labor implies that all fluctuations are along
the extensive margin, one would expect - even without looking at statistics
calculated from the data - that the ratio discussed above should be somewhere
between the one implied by an indivisible labor economy and a divisible labor
economy.
6. Conclusion
A dynamic competitive equilibrium economy with indivisible labor has been
constructed with the aim of accounting for standard deviations and correlations with output found in aggregate economic time series. Individuals in this
economy are forced to enter and exit the labor force in response to technology
shocks rather than simply adjusting the number of hours worked while
remaining continuously employed. Therefore, this is an equilibrium model
which exhibits unemployment (or employment) fluctuations in response to
aggregate shocks. Fluctuations in employment seem important for fluctuations
*This ratio is still not significantly different from one even when uCis increased to 0.00929.

324

G. D. Hmsen,

Indioisible

labor

and the business

qwzle

in hours worked over the business cycle since most of the variability in total
hours is unambiguously due to variation in the number employed rather than
hours per employed worker.
An important aspect of this economy is that the elasticity of substitution
between leisure in different periods for the aggregate economy is infinite and
independent of the elasticity of substitution implied by the individuals utility
function. This distinguishes this model, or any Rogerson (1984) style economy,
from one without indivisible labor. These include the model presented in
section 3.1 and the economy studied by Kydland and Prescott (1982). In these
divisible labor models, the elasticity of substitution for the aggregate economy
is the same as that for individuals.
This feature enables the indivisible labor economy to exhibit large fluctuations in hours worked relative to fluctuations in productivity. Previous equilibrium models of the business cycle, which have all assumed divisible labor,
have been unsuccessful in accounting for this feature of U.S. time series. This
is illustrated in this paper by showing that a model with divisible labor fails to
exhibit large fluctuations in hours worked relative to productivity while the
model with indivisible labor displays fluctuations in hours relative to productivity which are much larger than what is observed. This seems to indicate that
a model which allowed for adjustment along both the extensive margin as well
as the intensive margin would have a good chance for successfully confronting
the data.
In conclusion, this study demonstrates that non-convexities such as indivisible labor may be important for explaining the volatility of hours relative to
productivity even when individuals are relatively unwilling to substitute leisure
across time. They are also useful for increasing the size of the standard
deviations of all variables relative to the standard deviation of the technology
shock. Therefore, a smaller size shock is sufficient for explaining business cycle
fluctuations than was true for previous models such as Kydland and Prescotts
(1982). In addition, these non-convexities make it possible for an equilibrium
model of the business cycle to exhibit fluctuations in employment. Therefore,
non-convexities will inevitably play an important role in future equilibrium
models of the cycle.
Appendix:

A market for unemployment insurance

The purpose of this appendix is to show that the equilibrium of the economy
presented in section 3.2 is equivalent to the equilibrium of an economy where
labor is still indivisible but households are able to purchase any amount of
unemployment insurance they choose. In the original economy, agents are
assumed to buy and sell contracts which specify a probability of working in a
given period as opposed to buying and selling hours of work. A lottery
determines which households must work and which do not. A household is

G.D. Hansen,

Indivisible

labor arld the business

cycle

325

paid according to the probability that it works, not according to the work it
actually does. In other words, the firm is automatically providing full unemployment insurance to the households.
In this appendix, households choose a probability of working each period
and a lottery is held to determine which households must work, just as
in the original economy. Also, preferences, technology, and the stochastic
structure are exactly the same as for the original model. However, this
economy is different in that households only get paid for the work they
actually do - unemployed individuals get paid nothing by the firm. But, the
household does have access to an insurance market which preserves the
complete markets aspect of the original model. It is shown below that the
equilibrium of this economy is equivalent to that of the original economy since
individuals will choose to be fully insured in equilibrium. This is shown by
proving that the problem solved by households is the same as the problem
solved by households (12) in the original model.
The problem solved by the households can be described as follows: Each
period, households choose a probability of working, (Y,, a level of unemployment compensation, y,, and consumption and investment contingent on
whether the household works or not, c,, and i,, (s = 1,2). These are chosen to
solve the following dynamic programming problem (primes denote next period
values) :

(A-1)
subject to
cl + i, I w(h, K)h,+
c2 + i, ly

r(h, K)k -p(a)y,

+ r(X, K)k -p(a)y,

ki = (1 - S)k + i,,

s= 1,2.

64.2)
(A-3)

(A.4

The function V(A, K, k) is the value function which depends on the households state. The state vector includes the capital owned by the household, plus
the economy wide state variables A and K, where K is the per capita capital
stock.21 The functions w(h, K) and r(h, K) are the wage rate and rental rate
Since we are allowing households to choose any level of unemployment insurance they wish,
we have to allow for the heterogeneity that may come about because different households will have
different income streams. This is why the distinction is made between the per capita capital stock,
K, and the households accumulated capital stock, k. However, this heterogeneity will disappear in
equilibrium since all households will choose full insurance, so K = k in equilibrium.

of capital respectively, and p(a) is the price of insurance, which is a function

of the probability that the household works. Also, since individuals preferences are the same as for the original model, u(c) = log c and v(I) = A log I.
The insurance company in this economy maximizes expected profits which
are given by p(a)v - (1 - a)~. That is, the firm collects revenue p(cu)v and
pays y with probability 1 - (r. To guarantee that profits are bounded, p(a) =
(1 - a). Therefore, the price the household must pay for insurance equals the
probability that the household will collect on the insurance.
One can now substitute this expression for p into constraints (A.2) and
(A.3). After eliminating the constraints by substituting out i,v and c,~(s = 1,2),
one can write the following first-order necessary conditions for k; and .y:
~(c,)=BEV,.(A,K,k:),

s= 1,2,

u( c,) = u( cz).

(A-5)
64.6)

Eq. (A.6) implies, given the strict concavity of U, that c, = cz. This plus eq.
(A.5) imply that k{ = k$. This, in turn, implies that i, = i,. Therefore, the
left-hand sides of eqs. (A.2) and (A.3) are identical. Since these constraints will
be binding in equilibrium, y will be chosen so that the right-hand sides are
equal as well. This means that y = wh, in equilibrium. That is, households will
choose to insure themselves fully. This has the implication that all households
will choose the same sequence of capital stocks, so K = k.
Substituting these results into the households optimization problem (A.l)
yields the following problem: Households choose c, i, k, and a! to
maxV(h,k)=u(c)+av(l-h,)+(l-ol)v(l)+pEV(X,k),
(A.7)
subject to

k = (1 - 6)k + i.

This problem is identical to problem (12). Therefore, the equilibrium allocation for the original economy, where the firm provides full unemployment
insurance to workers by assumption, is equivalent to the equilibrium allocation
for an economy where households get paid by the firm only for work done but
have access to a risk-neutral insurance market. This result, of course, depends
crucially on tire probability OLbeing publicly observable and the contract being
enforceable. That is, it must be the case that the agent announces the same (Y
to both the firm and the insurance company, and if the agent loses the lottery
(that is, has to work) this is known by all parties. For example, this result

G. D. Honsen,

Indivisible

lubor and rite busiuess

cycle

327

would not hold if cr depended on some underlying choice variable like effort
that was not directly observed by the insurance company. In this case a
difficult moral hazard problem would arise.
References
Altonji, J.G.. 1984, Intertemporal substitution in labor supply: Evidence from micro data,
Unpublished manuscript (Columbia University, New York). .
Asbenfelter. 0.. 1984, Macroeconomic analyses and microeconomic analyses of labor supply,
Carnegie-Rochester Conference Series on Public Policy 21, 117-156.
Bcrtsekas, D.P.. 1976, Dynamic programming and stochastic control (Academic Press, New York).
Coleman, T.S., 1984, Essays on aggregate labor market business cycle fluctuations, Unpublished
manuscript (University of Chicago, Chicago, IL).
Debreu. G., 1954, Valuation equilibrium and Pareto optimum, Proceedings of the National
Academy of Sciences 40, 588-592.
Heckman. J.J., 1984, Comments on the Ashenfelter and Kydland papers, Carnegie-Rochester
Conference Series on Public Policy 21,209-224.
Heckman, J.J. and T.E. MaCurdy, 1980, A life cycle model of female labor supply, Review of
Economic Studies 47,47-14.
Hodrick, R.J. and EC. Prescott, 1980, Post-war U.S. business cycles: An empirical investigation.
Working paper (Carnegie-Mellon University, Pittsburgh, PA).
Kydland, F.E. and E.C. Prescott, 1982, Time to build and aggregate fluctuations. Econometrica 50,
1345-1370.
Kydland, F.E. and E.C. Prescott, 1984, The workweek of capital and labor, Unpublished
manuscript (Federal Reserve Bank of Minneapolis, Minneapolis. MN).
Lucas. R.E., Jr.. 1977, Understanding business cycles, Carnegie-Rochester Conference Series on
Public Policy 5, 7-29.
Lucas, R.E.. Jr. and EC. Prescott, 1971, Investment under uncertainty, Econometrica 39,659-681.
Lucas, R.E., Jr.. EC. Prescott and N.L. Stokey. 1984, Recursive methods for economic dynamics,
Unpublished manuscript (University of Minnesota, Minneapolis, MN).
MaCurdy. T.E.. An empirical model of labor supply in a life-cycle setting, Journal of Political
Economy 89. 1059-1085.
Rogerson, R., Indivisible labour, lotteries and equilibrium, Unpublished manuscript (University of
Rochester, Rochester, NY).
Prescott, E.C.. Can the cycle be reconciled with a consistent theory of expectations? or a progress
report on business cycle theory, Unpublished manuscript (Federal Reserve Bank of Minneapolis, Minneapolis, MN).
Solow, R.M.. Technical change and the aggregate production function, The Review of Economics
and Statistics 39, 312-320.