CAPITAL THEORY AND INVESTMENT BEHAVIOR*

By DALE W. JORGENSON
Universityof California,Berkeley
Introduction
There is no greater gap between economic theory and econometric
practice than that which characterizes the literature on business in-
vestment in fixed capital. According to the neoclassical theory of capi-
tal, as expounded for example by Irving Fisher, a production plan for
the firm is chosen so as to maximize utility over time. Under certain
well-known conditions this leads to maximization of the net worth of
the enterprise as the criterion for optimal capital accumulation. Capi-
tal is accumulated to provide capital services, which are inputs to the
productive process. For convenience the relationship between inputs,
including the input of capital services, and output is summarized in a
production function. Although this theory has been known for at least
fifty years, it is currently undergoing a great revival in interest. The
theory appears to be gaining increasing currency and more widespread
understanding.
By contrast, the econometric literature on business investment con-
sists of ad hoc descriptive generalizations such as the "capacity prin-
ciple," the "profit principle," and the like. Given sufficientimprecision,
one can rationalize any generalization of this type by an appeal to
"theory." However, even with the aid of much ambiguity, it is im-
possible to reconcile the theory of the econometric literature on in-
vestment with the neoclassical theory of optimal capital accumula-
tion. The central feature of the neoclassical theory is the response of
the demand for capital to changes in relative factor prices or the ratio
of factor prices to the price of output. This feature is entirely absent
from the econometricliterature on investment.
It is difficult to reconcile the steady advance in the acceptance of
the neoclassical theory of capital with the steady march of the econo-
metric literature in a direction which appears to be diametrically op-
posite. It is true that there have been attempts to validate the theory.
Both profits and capacity theorists have tried a rate of interest here
or a price of investment goods there. By and large these efforts have
been unsuccessful; the naive positivist can only conclude, so much the
worse for the theory. I believe that a case can be made that previous
attempts to "test" the neoclassical theory of capital have fallen so far
* The researchfor this paper was completedwhile the author was Ford FoundationRe-
searchProfessorof Economicsat the Universityof Chicago.The researchwas supportedby
the NationalScienceFoundation.
247
248 AMERICAN ECONOMIC ASSOCIATION

short of a correct formulation of this theory that the issue of the

validity of the neoclassical theory remains undecided. There is not suf-
ficient space to document this point in detail here; but I will try to
illustrate what I would regard as a correct formulation of the theory
in what follows.
Stated baldly, the purpose of this paper is to present a theory of
investment behavior based on the neoclassical theory of optimal accu-
mulation of capital. Of course, demand for capital is not demand for
investment. The short-run determination of investment behavior de-
pends on the time form of lagged response to changes in the demand
for capital. For simplicity, the time form of lagged response will be
assumed to be fixed. At the same time a more general hypothesis about
the form of the lag is admitted than that customary in the literature.
Finally, it will be assumed that replacement investment is proportional
to capital stock. This assumption, while customary, has a deep justi-
fication which will be presented below. A number of empirical tests
of the theory is presented, along with an analysis of new evidence on
the time form of lagged response and changes in the long-run demand
for capital resulting from changes in underlying market conditions and
in the tax structure.
Summary of the Theory
Demand for capital stock is determined to maximize net worth. Net
worth is defined as the integral of discounted net revenues; all prices,
including the interest rate, are taken as fixed. Net revenue is defined as
current revenue less expenditure on both current and capital account,
including taxes. Let revenue before taxes at time I be R(t), direct taxes,
D(t), and r the rate of interest. Net worth, say W, is

w= e-rt[R(t) - D(t)]dt.

We will deduce necessary conditions for maximization of net worth for

two inputs-one current and one capital-and one output. The ap-
proach is easily generalized to any number of inputs and outputs.
Let p be the price of output, s the wage rate, q the price of capital
goods, Q the quantity of output, L the quantity of variable input, say
labor, and I the rate of investment; net revenue is
R = pQ - sL - ql.

Let u be the rate of direct taxation, v the proportion of replacement

chargeableagainst income for tax purposes,w the proportionof interest,
 TOPICS IN ECONOMIC THEORY 249
and x the proportion of capital losses chargeable against income; where
K is capital stock and a the rate of replacement, direct taxes are
D = u[pQ - sL - (v3q + wrq - xq)K]

Maximizing net worth subject to a standard neoclassical production

function and the constraint that the rate of growth of capital stock is
investment less replacement, we obtain the marginal productivity con-
ditions
OQ s
dL p
--UV ufw 1-uX 4
g+
q r- r-~~q
3(Q _ 1- 1-u 1-uqi
OK p

The numerator of the second fraction is the "shadow" price or implicit

rental of one unit of capital service per period of time. We will call this
price the user cost of capital. WVeassume that all capital gains are re-
garded as "transitory," so that the formula for user cost, say c, reduces
to
-uv 1-uw
c = q _1X +1ur.

Second, we assume that output and employment on the one hand and
capital stock on the other are determined by a kind of iterative process.
In each period, production and employment are set at the levels given
by the first marginal productivity condition and the production func-
tion with capital stock fixed at its current level; demand for capital is
set at the level given by the second marginal productivity condition,
given output and employment. With stationary market conditions, such
a process is easily seen to converge to the desired maximum of net
worth. Let K* represent the desired amount of capital stock, if the
production function is Cobb-Douglas with elasticity of output with
respect to capital, y,
pQ
K*-PQ

We suppose that the distribution of times to completion of new in-

vestment projects is fixed. Let the proportion of projects completed in
time r be WT. If investment in new projects is 4 and the level of starts
of new projects is I,, investment is a weighted average of past starts:
250 AMERICAN ECONOMIC ASSOCIATION

E 00N

It = w(L)It
r=O

where w(L) is a power series in the lag operator, L. We assume that in

each period new projects are initiated until the backlog of uncompleted
projects is equal to the difference between desired capital stock, K*,
and actual capital stock, Kt:
N *N
It =Kt - [Kt + (1 -wo)It_1 +
which implies that:

It = w(L) [Kt - Kt-1.

It is easy to incorporateintermediate stages of the investment process
into the theory. For concreteness,we considerthe case of two intermedi-
ate stages, which will turn out to be anticipated investment, two
quarters hence, and anticipated investment, one quarter hence. A sim-
ilar approach can be applied to additional intermediate stages such as
appropriationsor commitments. The distribution of completions of the
first stage, given new project starts, may be described by a sequence,
say {VOT }; similarly, the distribution of completions of a second stage,
given completion of the first stage, may be described by a sequence
IV1T}. Finally, the distribution of investment expenditures, given
completion of a second intermediate stage is described by a sequence
V2T}. Where I,"E represents completions of the first stage, 4s2E com-
pletions of the second stage, and Is actual investment, as before, we
have:

It -E Voit-2 =vo(L)I ,
r.0
0
S2E S E S1E
It = E = vi(L)It
,=O

It EVI2si-
a 2(L)It2
T=0

where vo(L), v1(L), and v2(L) are power series in the lag operator.
Up to this point we have discussed investment generated by an
increase in desired capital stock. Total investment, say It, is the sum of
investment for expansion and investment for replacement, say IR:
E It
It = It + It
We assume that replacement investment is proportional to capital
 TOPICS IN ECONOMIC THEORY 251
stock. The justification for this assumption is that the appropriate
model for replacement is not the distribution of replacements for a
single investment over time but rather the infinite stream of replace-
ments generated by a single investment; in the language of probability
theory, replacement is a recurrent event. It is a fundamental result of
renewal theory that replacements for such an infinite stream approach
a constant proportion of capital stock for (almost) any distribution
of replacementsfor a single investment and for any initial age distribu-
tion of capital stock. This is true for both constant and growing capital
stocks. Representing the replacement proportion by 6, as before,
R
It = Wt;
combining this relationship with the correspondingrelationship for in-
vestment in new projects, we have:

It w(L)[K - Kt-1 + SKt.

Using the assumption that capital stock is continued in use up to the

point at which it is replaced, we obtain the correspondingrelationships
for gross investment at each of the intermediate stages, say 4'1and
Its2

4t = vo(L)[Kt - Kt-i + 8Kt,

82
It vl(L)vo(L)[Kt*-Kt-1] + AK:;
we can also derive the following;

It = vi(L) [Its'-Kt] + SKt,

It = v2(L)[I?2_ -Kt]+ SKt,
It = v2(L)vzd(L) -
[I -Kt] + SKt.

For empirical implementation of the theory of investment behavior,

it is essential that each of the power series-vo(L), vi(L), v2(L)-have
coefficientsgenerated by a rational function; for example,
s(L)
w(L) = v2(L)v,(L)vo(L) =s L
1(L)
where s(L) and t(L) are polynomials. We will call the distribution
correspondingto the coefficientsof such a power series a rational power
series distribution. The geometric and Pascal distributions are among
the many special instances of the rational power series distribution.
2 52 AMERICAN ECONOMIC ASSOCIATION

Empirical Results
To test the theory of investment behavior summarizedin the preced-
ing section, the correspondingstochastic equations have been fitted to
quarterly data for U. S. manufacturing for the period 1948-60. The
data on investment are taken from the OBE-SEC Survey; first and
second anticipations of investment expenditureas reported in that Sur-
vey are taken as intermediate stages.' With two intermediate stages,
six possible relationships may be fitted. First, for actual investment
and both intermediate stages, the level of investment is determined by
past changes in desired capital stock. Second, investment is determined
by past values at each intermediate stage and the second anticipation
is determined by past values of the first anticipation. The first test of
the theory is the internal consistency of direct and derived estimates
of the coefficients of each of the underlying power series in the lag
operator.
The results of the fitting are given in Table 1. For each of the fitted
relationships coefficients of the polynomials s(L) and t(L) in the ex-
1 Data on capital stock were obtained by interpolating the capital stock series for total
manufacturing given in the U.S. national accounts between 1949 and 1959, using the formula

Kt+1 = It + (1 -)Kt.
Given an investment series, a unique value of a may be determined from the initial and
terminal values of capital stock. Investment data from the OBE-SEC Survey were used for
the interpolation. For desired capital stock, the quantity pQ was taken to be sales plus changes
in inventories, both from the Survey of Current Business. User cost depends on a number of
separate pieces of data. The quantity q is an investment deflator, a is, of course, a fixed pa-
rameter (taken to be equal to .025), r is the U.S. government long-term bond rate. The tax
functions vary with time; as an example, the tax rate u is the ratio between corporate income
tax payments and corporate profits before taxes as reported in the U.S. national accounts.
A detailed description of the data underlying this study will be reported elsewhere.
2 To derive the form of the functions used in the actual fitting, we take v2(L)v1(L) Vo(L)
as an example. First:

I s() [Kt - K*1] + aKt.

t(L)
Secondly,

It = s(L)[K*- K*t-] + [1 - t(L)][It - aKt] + aKt.

The coefficient to may be normalized at unity so that:

1 - t(L) = - t1L -t2L2 - **-

The a priori value 5= .025 was used to compute It-,-Ktr. An estimate of a is given by the
coefficient of Kt. If a is different from its a priori value, the process of estimation can be re-
iterated, using a second approximation to the value of &.
The parameter -y is estimated using the constraint:

w0
T P0
 TOPICS IN ECONOMIC THEORY 2 53
TABLE 1
REGRESSION COEFFICIENTS AND GoODNESS OF FIT STATISTICS, UNRESTRICTEDESTIMATES

Regres- t2 R2 s
'SO y yS2 1 A2/s2

72V1l' O .00102 -1.51911 .63560 .02556 .94265 .10841 2.14039

(.00049) (.09945) (.10098) (.00163)
vivo .00132 -1.25242 .36656 .02618 .89024 .16229 2.00431
(.00073) (.12667) (.12977) (.00240)
7o .00109 -1.26004 .37281 .02549 .87227 .18974 2.37298
(.00085) (.13044) (.13138) (.00278)

V271 .81357 .01962 .92729 .11955 1.16294

(.03492) (.00175)
1. .90024 .02295 .96234 .08604 1.47693
(.02722) (.00127)
.89462 .02337 .95121 .10597 1.70179
(.03145) (.00155)

REGRESS10N COE.F1CI1ENTS AND GOODNESS OF FIT STATISTICS, RESTRICTED ESTIMATES

V21 VO .00106 -1.52387 .63100 .94156 .10830 2.10778

(.00049) (.09925) (.10074)
*'i7 0 -01.25704
.00133 .36769 .88986 .16087 2.00549
(.00073) (.12509) (.12862)
Ic .00109 -1.26395 .37300 .87128 .18848 2.53442
(.00084) (.12942) (.13051)
ra.1 .82764 .89995 .13883 .87127
(.04037)

V2 .91545 .95406 .09409 1.23560

(.02933)
.90271 .94538 .11100 1.53759
(.03276)

pression for each power series as a rational function are given.2 For
example, the power series v2(L)v1(L)vo(L) is expressed as:

.00106L2
V2(L)vj(L)vo(L) = -2
1 - 1.52387L + .63100L2

The value of the replacement proportion a estimated from data on

capital stock is .025. Two sets of regressions were run, one with 5 fitted
from the data (unrestricted), the other with 3=.025 (restricted).
Throughout, the coefficient of multiple determination R2, the standard
error of estimate for the regression s, and the VonNeumann ratio A2/s2
are presented as measures of goodness of fit.
The first set of tests of the theory is the comparison of alternative
estimates of each of the fundamental power series. As an example, one
may take the hypothesis that the direct estimates of the power series
v2(L) and vi(L), when combined, give an estimate of v2(L)v1(L) which
is close to that obtained by direct estimation. Using the unrestricted
estimates, the result of this comparison is:
254 AMERICAN ECONOMIC ASSOCIATION

.91545L0 .90271L = .82639L2;

the derived estimate, which may be comparedwith the direct estimate,
.82764L2.The differencebetween the two estimates is slightly over .03
standard errors. A similar test of the hypothesis that the direct esti-
mates of the power series vi(L) and vo(L),when combined, yield an esti-
mate of vi(L)vo(L)which is close to that obtained by direct estimation
results in
.90271L .00098L
.00109* 2-
1 - 1.26395L + .37300L2 1 - 1.26395L + .37300L2
which may be compared with the direct estimate,
.00133L
1 - 1.25704L + .36769L2
The coefficient of the numerator is within half a standard error of the
derived estimate. The coefficients of the denominator are within .06
and .04 standard errors of the derived estimates. The similarity of
derived and direct estimates for the power series v2(L)vi(L)vo(L) is less
striking. The three possible derived estimates are extremely similar to
each other, but they differ considerably from the direct estimate. Nev-
ertheless, using any of the derived estimates as the null hypothesis for a
test of the direct estimates would probably lead to acceptance of the
null hypothesis. In general, the theory of investment behavior is
strongly confirmedby the set of tests of internal consistency. Of course,
given the internal consistency of the alternative estimates, it is possible
to improve efficiency of estimation for the model as a whole by combin-
ing information from the various sources.
The tests of internal consistency just describedare tests of the theory
of investment in new projects. A test of the theory of replacement in-
vestment is a test of the consistency of the empirical results with the
hypothesis a=.025. This hypothesis is borne out in two ways. First, for
all but one of the regressions, the usual null hypothesis is accepted; a
much stronger result is that for the first three regressions, estimates
of the relationships under the restriction that 6=.025 results in a reduc-
tion in the standard errorof estimate for the regression.Finally, each of
the standard errors of the estimates of a is less than one-tenth the size
of the correspondingregression coefficient. We conclude that the hy-
pothesis that replacement is a constant fraction of capital stock, spe-
cifically, that 6=.025, is strongly validated by the empirical results.
We turn now to comparisons of the fitted regressions with some
simple alternatives. First, as alternatives for the first three regressions,
we take the naive models:
 TOPICS IN ECONOMIC THEORY 255

It = It-1,
S2 S2
I t = It-II
Si S1
I t= It-i.

Simple as these models may be, they are quite stringent standards for
comparison for seasonally adjusted quarterly data, much more strin-
gent, for example, than the correspondingmodels for annual data. The
appropriatestatistics for comparisonare the standard errorsof estimate
and the VonNeumann ratios. Results of this comparison are given
separately for the periods 1948-60 and second quarter 1955 to 1960 in
Table 2.3 For the period as a whole, each of the regressionmodels has a
standard error well below that for the correspondingnaive model. For
the later subperiod the advantage of the regression models is even
greater. Turning to the VonNeumann ratios, there is practically no
evidence of autocorrelated errors for the fitted models and very clear
evidence of autocorrelation for the naive models. Of course, this test is
biased in favor of the fitted regressions. Even with this qualification,
the fitted regressions are clearly superior in every respect to the cor-
respondingnaive models.
As a standard of comparison for the second three regressions, we
take the forecasts actually used by the Department of Commerce in
presenting the results of the OBE-SEC Survey. These alternative
models take the form:

ItI _
It =It-2, 8
2

59 S
1t It-l.

Despite the high level of performance of the OBE-SEC anticipations

data, the fitted regressions constitute a substantial improvement in
both goodness of fit as measured by standard error of estimate and ab-
sence of autocorrelationof residuals. The test for autocorrelationis not
biased in favor of the fitted regressions,so that the evidence is unequiv-
ocal; the fitted relationships are clearly superior to the corresponding
forecasting models for the period as a whole and for the subperiod since
second quarter 1955.
A further comparisonof the fitted regressionswith the corresponding
naive and forecasting models is given in the second half of Table 2,
where an analysis of the conformity of turning points of each of the
3Data for both anticipations and actual expenditures on a revised basis are available
from the Department of Commerce only since the second quarter of 1955. Anticipations data
for the earlier period were revised by multiplying each observation by the ratio of revised to
unrevised actual investment for the period in which the observation was made.
256 AMERICAN ECONOMIC ASSOCIATION

t
if i't 2 Ig
t2I i't2= =
t I,=12It=I't'
NOTE:I =I't'lit-I Iu_i
=1(Ir1)
Ig=f(I')
t2:f(I8l)
) It=f(AKg)Model
It2=f(AKs
=f(AKt)
Total t t 1

R2
.93380 .83673 .95121
.93504 .92729 .83854
.96234 .86193 .87128
.84966 .88986
.94156
percentages

may
S
not .11983 .17391 .10597
.10969 .11955 .20282
.08604 .18058 .16087
.15950 .18848 .10830
add 19481-1960IV
to
GOODNESS
A2/S2
.99342
100%01.45737 .69933 1.70179 1.162941.31901.66900 2.53442
1.47693 2.00549
2.10778 OF

FIT
..04366
because R2
of .92833 .95931
.77138 .96477
.91854 .92169 .81161 .81929 .91921
.81297 .94757
.94298
STATISTICS:

rounding s
821036.93
.12031 .20707 .09045
.12359 .19435
.12996 .19947
.09368 .18410 .14372
.09368
.11378 FITTED,
1955II-1960IV
TABLE
error.
2
NAIVE,
A2/S2
1.25062 .46505 2.26525
.91146 1.65030 .94465
1.09855.83580 .52078 2.25394
1.84699
1.95800 AND

20 14 25 16 14 16 35 24 22 39 29 TP
29%o Error
FORECASTING
49 51 47 45 45 43 25 37 39 33 43
Over-
47%/o estimate
MODELS
1948I-1960IV
31 35 27 39 41 41 39 39 39 27 27
Under-
estimate
24%|

18 18 27 23 18 23 32 32 23 41 41 TP
Error
23%|

45 50 50 41 45 41 27 27 32 27 32
Over-
41% estimate

1955II-1960IV
136 32 23 36 36 36 41 41 45 32 27
36%o estimate
Under-
 TOPICS IN ECONOMIC THEORY 25 7
"forecasts" to the turning points of the actual data is presented. In
general, the first set of fitted regressions is slightly inferior to the naive
models and the second set slightly superior to the forecasting models on
the basis of this criterion. A final comparison is between the fitted re-
gression of investment on changes in desired capital stock and the fore-
cast of investment from its second anticipation. The comparison favors
the fitted regression; however, the anticipations data used in a fitted
relationship between investment and second anticipation provide a
model which is superior to the simple forecasting model and to the
fitted regression of investment on changes in desired capital stock.
Structure of the Investment Process
In the preceding sections, only those aspects of the theory of invest-
ment behavior relevant to testing the theory were presented. In this
section certain further implications of the theory are developed. Specifi-
cally, we will characterize the long-term response of investment to
changes in the underlying market conditions and the tax structure and
the time pattern of response of investment to changes in demand for
capital.
First, using the facts that gross investment is determined by the rela-
tionship:

It = w(L)[Kt- Kt1] + AK:

and that capital stock is determined by past investments, we obtain:

It = [1 - (1 - 8)L]w(L)K*,
= y(L)Kt,
where y(L) is a power series in the lag operator. XVedefine the r-period
response of investment to a change in market conditions or tax struc-
ture as the change in gross investment resulting from a change in the
underlying conditions which persists for r periods. More precisely, sup-
pose that desired capital remains at a fixed level for r periods to the
present; then,

Kt = Kt-, (v = 1, 2 ),

and
00

It = E yvKt-p)
co
+
=zKt E yvKt-',
P-7+ 1
258 AMERICAN ECONOMIC ASSOCIATION

where {Z4} is the sequence of cumulative sums of the coefficients of

y(L). As an example, the response of gross investment to a change in the
rate of interest is:
AI dK*
- ZT .
ar ar
The coefficients {Z-r} characterize the time pattern of response. Obvi-
ously,
lim z7 - lim yv =6,
7T* X 7t+ P0

so that the long-term response of gross investment to changes in, say,

the rate of interest, is
ai dK*
clr ar
Clearly, the short-term responses approach the long-term response as a
limit; the approach is not necessarily monotone, since the coefficients
of the power series y(L) are not necessarily non-negative.
Long-term response and elasticities of gross investment with respect
to the price of output, price of capital goods, and the rate of interest are
given in the top half of Table 3. The corresponding responses and
elasticities for the income tax rate, the proportion of replacement and
the proportion of interest chargeable against income for tax purposes
are given in the bottom half of Table 3. It should be noted that the rate
of interest and the tax rate are measured as proportions, not percent-
ages. For example, a decrease in the rate of interest by 1 per cent in-
creases manufacturinggross investment by $.15178 billions per quarter
in the long run, at least to a first approximation.
The time pattern of response is presented in Table 4, where the func-
tions w(L), y(L), and z(L) are derived from the fitted regressions. The
TABLE 3
RESPONSES AND ELASTICITIESOF INVESTMENTWITH RESPECT TO CHANGES IN MARKET
ANDTAX STRUCTURE
CONDITIONS

RESPONSE ELASTICITY
Average End of Period Average End of Period
Market Conditions
Price of output .......... .35830 .35299 1.00000 1.00000
Price of capital goods .... -.35273 -.32106 -1.00000 -1.00000
Rate of interest ......... -14.23653 -15.17789 -.29143 -.37866

Tax Structure
Income tax rate ......... -.37487 -.33016 -.50959 -.42064
Proportion of replacement .18729 .20502 .39181 .48565
Proportion of interest.... .55656 .79840 .19428 .32659
 TOPICS IN ECONOMIC THEORY 2 59
TABLE 4
TImE FORM OF LAGGEDRESPONSE

Lag w(L) y(L) z(L)

0 0 0 0
1 0 0 0
2 .11277 .11277 .11277
3 .14209 .03214 .14491
4 .13700 -.00154 .14337
5 .11965 - .01393 .12944
6 .09969 - .01697 .11247
7 .08101 - .01619 .09628
8 .06491 -.01407 .08221
9 .05159 - .01170 .07051
10 .04081 - .00949 .06102
11 .03219 - .00760 .05342
12 .02535 - .00604 .04738
13 .01994 - .00478 .04260
14 .01567 - .00377 .03883
15 .01231 - .00297 .03586
16 .00967 - .00233 .03353
17 .00760 - .00183 .03170
18 .00597 - .00144 .03026
19 .00469 - .00113 .02913
20 .00368 - .00089 .02824

Remaining ..01346 -.00325

Rate of decline ..78531 .78531

average lag between change in demand for capital stock and the cor-
responding net investment is, roughly, 6.5 quarters or about a year and
a half. Of course, this estimate is affected by the essentially arbitrary
decision to set the proportion of the change invested in the same period
and period immediately following the change at zero. The coefficients
of the power series z(L) are of interest for computation of short-period
responses of investment to changes in the demand for capital stock.
For example, the 2-period response of manufacturing gross investment
to a change in the rate of interest of 1 per cent is:

AK* .11277
Z7 Or = . .15178 = .68465 billions/quarter.
ar .02500

By comparison, the corresponding 10-period response is .37046 billions

per quarter. The response dies out, almost to its long-term level of
.15178 billions/quarter, by twenty periods from the initial change in
demand for capital stock. Similar calculations of the response of gross
investment to changes in market conditions or the tax structure may be
made for any of the six determinants of demand for capital by combin-
ing the responses given in Table 3 with the time pattern presented in
Table 4.