Explaining the Investment Boom of the 1990s

Sta ey Tevlin
Karl Whelan 

Mar h 21, 2000

Abstra t
Real equipment investment in the United States has boomed in re ent years, led by
soaring investment in omputers. We nd that traditional aggregate e onometri models ompletely fail to apture the magnitude of this re ent growth|mainly be ause
these models negle t to address two features that are ru ial (and unique) to the urrent investment boom. First, the pa e at whi h rms repla e depre iated apital has
in reased. Se ond, investment has been more sensitive to the ost of apital. We do ument that these two features stem from the spe ial behavior of investment in omputers
and therefore propose a disaggregated approa h. This produ es an e onometri model
that su essfully explains the 1990s equipment investment boom. 

Division

of Resear h and Statisti s, Federal Reserve Board. Email: kwhelanfrb.gov or stevlinfrb.gov.
We would like to thank Robert Chirinko, David Lebow, Deb Lindner, Dan Si hel, and Peter Tulip for their
valuable suggestions. We would also like to thank seminar parti ipants at the 2000 AEA meetings, the
New York Fed, and the Federal Reserve Board. The views expressed in this paper are our own and do not
ne essarily re e t the views of the Board of Governors or the sta of the Federal Reserve System.

1

Introdu tion

The behavior of equipment investment in the urrent U.S. expansion has been remarkable.
Growth in real equipment investment over the period 1992-98 averaged 11.2 per ent per
year, ex eeding all other seven-year intervals in the post-War era.1 This development
has been of great ma roe onomi importan e: The investment boom has underpinned
the ontinuing strength of U.S. aggregate demand and has probably also had important
supply-side e e ts, perhaps playing a role in the unusual late- y le a eleration in labor
produ tivity.
In this paper, we examine whether existing time series models an explain the astounding behavior of equipment investment in the 1990s. We demonstrate that they annot.
Although we examine the traditional, a elerator-style models that previous investment
\horsera e" studies have found best t the data, we nd that they ompletely fail to apture the magnitude of the 1990s investment boom.2 We show that the models' breakdown
stems from an important element of investment growth in the 1990s|the surge in real
investment in omputing equipment. Our analysis of the behavior of omputer investment
reveals two features that, though ru ial to the investment boom of the 1990s, are ignored
by standard aggregate models. We demonstrate that a disaggregated approa h, whi h models investment in omputing and non- omputing equipment separately, su essfully explains
the behavior of investment in the 1990s.
The rst feature that we do ument is the sharp in rease in the average rate of depre iation in the 1990s. Most e onometri models assume a onstant depre iation rate and thus
a stable relationship between the hange in the apital sto k and the level of investment.
Sin e the optimal apital sto k is a fun tion of the level of output and the ost of apital,
this also implies a stable link between investment and hanges in output and the ost of
apital. However, the in reasing rate of depre iation in the 1990s broke this link: Firms
needed to invest more to sustain a given level of the apital sto k. We show that the in rease
in the depre iation rate was due to a shift in the omposition of apital towards omputers,
1

All gures in this paper refer to 1992-based National In ome statisti s rather than the 1996-based gures
published in O tober 1999. The e onometri models in our paper use apital sto k data, and revised apital
sto ks will not be published until Spring 2000.
2
For earlier evaluations of ompeting investment models, see Clark (1979) and Bernanke, Bohn, and
Reiss (1988). Oliner, Si hel, and Rudebus h (1995) ompare a elerator-style models with models based on
Euler equations or Tobin's Q.

1

whi h depre iate more rapidly than other types of equipment. Aggregate models do not
apture this phenomenon, be ause, by de nition, they ignore ompositional mix-shifts.
The se ond feature that we examine is the role of the ost of apital. The rising average
depre iation rate suggests the need to separately model net investment in omputing and
non- omputing equipment. Doing so reveals an important pattern. Computer investment
is very sensitive to the ost of apital, far more so than investment in non- omputing equipment. As a result, rapid de lines in omputer pri es played a ru ial role in generating the
investment boom of the 1990s. This result ontrasts sharply with most of the empiri al literature on aggregate investment, whi h typi ally nds very little response to ost variables.
We provide a plausible explanation for the di erent estimates of ost-of- apital elasti ities
that we observe: Firms respond more to sho ks per eived as permanent than to those
per eived as transitory, and sho ks to omputer pri es usually result from te hnologi al
innovations that are unlikely to be reversed.
We on lude that the spe ial behavior of equipment investment in the 1990s resulted
from the substantial impa t of rapid omputer pri e de lines on apital a umulation,
and the onsequent need for higher rates of repla ement investment. A simple disaggregated approa h, whi h separately models net and gross investment for omputing and
non- omputing equipment is apable of explaining the re ent behavior of investment.
The paper is organized as follows. Se tion 2 reviews the traditional e onometri models
and do uments their poor empiri al performan e in the 1990s. Se tion 3 examines the
in rease in the average rate of depre iation and its role in the breakdown of the onventional
models. Se tion 4 dis usses why apital a umulation may respond more to the persistent
omponent of the ost of apital than to the less persistent omponent. Se tion 5 presents
our e onometri analysis and do uments the performan e of our approa h in tra king the
behavior of aggregate gross investment in the 1990s. Se tion 6 on ludes.

2

2

Traditional Investment Models and Their Re ent Failure

Traditional models of investment start with a theory relating the optimal fri tionless apital
sto k, Kt, to the produ tion te hnology and fa tor pri es. If rms ould ostlessly adjust
the apital sto k, they would always set Kt = Kt. However, the sluggish behavior of the
apital sto k suggests that there are osts asso iated with adjustment. The traditional neoKeynesian investment models used simple ad ho spe i ations of the e e ts of adjustment
osts, the most ommon being the partial adjustment approa h, whi h assumed that rms
move part of the way towards their optimal fri tionless sto k ea h period. Formulating this
relationship in terms of the logarithm of the apital sto k and using lower ase letters to
denote the log of variables, the partial adjustment equation is 
k = (1
whi h an be re-written as:

)( 

kt

kt

1

)

(1)

(2)
= kt 1 + (1 )kt
Applying repeated substitution to equation (2) gives an equivalent representation for the
apital sto k, this time as an in nite distributed lag fun tion of past kt 's:
kt

kt

X1

= (1
r=0

) 

r kt r

=

X1 r t r

r=0

k

(3)

This has been turned into an empiri al investment equation by taking the following
steps. First, the in nite distributed lag suggested by the partial adjustment theory is
repla ed with a nite approximation, usually about 8 to 12 quarters. Se ond, the equation
is di eren ed to turn it into a net investment equation: 
kt =

XN r t r + t

r=0

k 

(4)

Of ourse, if the apital sto k adjustment equation is orre tly spe i ed, then this di eren ing step is not ne essary. However, the traditional literature largely pre-dated ointegration
methods and used stationarity-indu ing transformations as a pre aution against spurious
regressions. In our empiri al work, we will examine this issue formally.
This equation is operationalized by assuming a form for kt . Spe ifying a CES produ tion
fun tion, Kt is proportional to CYtt , where Yt is output, Ct is the ost of apital, and  is
3

the elasti ity of substitution between apital and labor. Taking logs of Kt we get 
kt =

X y
N

r 

t r

X  + 
N

r

t r

(5)

t

Sin e P1r=0 r = 1, the sum of the oeÆ ients on output should approximately equal one
while the oeÆ ients on the ost of apital should sum to the elasti ity of substitution, 
. These sums have an intuitive interpretation sin e they des ribe the predi ted long-run
response of the apital sto k to permanent unit sho ks to output and the ost of apital.
Models of the form of equation (5) have been estimated by Bernanke, Bohn, and Reiss
(1988). However, this approa h, whi h des ribes the determination of the apital sto k,
only gives us a model of net investment. For ma roe onomists interested in business y le
modelling and fore asting, the variable of interest is gross investment, whi h in ludes both
the hange in apital sto k and the repla ement of depre iated apital. Most empiri al
models assume a onstant average rate of depre iation and estimate an equation for gross
investment. In this ase, approximating the log-di eren e of the apital sto k with the
growth rate, we get 
kt  KKt = KIt Æ
(6)
t 1
t 1
where Æ is the depre iation rate. This gives an equation for gross investment
It
Kt 1

r =0

r =0

X
= Æ + y

X  +  

N

r =0

r

N

t r

r =0

r

t r

t

(7)

We will label this regression the \traditional model". Note that this approa h estimates
the depre iation rate as the inter ept in the KI regression.
Previous empiri al implementations of this model, estimated on data prior to the 1990s,
have found that it provides a fairly good des ription of the y li al behavior of investment.
Indeed, Oliner, Rudebus h, and Si hel (1995) have shown that models of this form provide superior fore asting performan e to popular alternative spe i ations based on Euler
equations or Tobin's Q. This is not to say that these models are without problems. For
instan e, despite mi roe onomi eviden e that the elasti ity of substitution is lose to one,
regressions usually reveal a small and often insigni ant role for the ost of apital. Indeed, omparisons of fore asting power have often favored the pure a elerator formulation
( = 0) over models in luding the ost of apital. Summarizing these results, Chirinko's
omprehensive 1993 survey on luded that \on balan e, the response of investment to pri es
tends to be quite small and unimportant relative to quantity variables."
t
t

4

1

Our estimation of equation (7) on rms these previous results, revealing a small longrun ost-of- apital elasti ity of -0.34. (Our data are des ribed in Appendix A.) However,
Figure 1, whi h shows the in-sample t, reveals a far more serious problem. The model
fails ompletely to apture the 1990s' in rease in investment relative to the apital sto k.
After 1991, the model underpredi ts by larger and larger amounts, with these residuals
prin ipally o set by large negative residuals over the early part of the sample.3 By 1997:4,
the a tual level of investment relative to the apital sto k is 7 per entage points higher
than an be explained by the model; this translates into a 31 per ent error on the level of
investment. A Chow test for parameter stability on rms that the model has gone o tra k
in the 1990s, with the null hypothesis of stable oeÆ ients resoundingly reje ted.4 In the
rest of the paper, we explore the reasons for the traditional model's omplete breakdown
in the 1990s.
3

The Unstable Aggregate Depre iation Rate

The most obvious simplifying assumption made in the derivation of the traditional model
is that the average rate of depre iation is onstant. We an easily he k the validity of this
assumption by solving for the aggregate depre iation rate obtained from re-arranging the
perpetual inventory equation (K = (1 Æ ) K 1 + I ) to get:
I 
K
Æ =
(8)
t

t

t

t

t

t

t

Kt 1

Figure 2 shows this series as the solid line (rather obs urely labeled \Using Chain-Weight
Investment and Capital", for reasons that will be ome apparent in a moment). It shows
that the aggregate depre iation rate has not been onstant, but has in reased substantially
in re ent years, rising from 0.13 in 1989 to 0.16 in 1997.
3

This is also true for popular alternative versions of this equation su h as the Jorgenson\neo lassi al"
model and an augmented version that in ludes ash ow. Models based on Tobin's Q, although predi ting
strong investment over the last few years of our sample, also do not tra k the behavior of KItt 1 in the 1990s
parti ularly well.
4
The gure also reveals a problem reported in previous horsera e papers. Even when residuals are small
during the middle part of our sample, they tend to be positively auto orrelated. We believe this is due to
the nite-lag approximation to the true in nite-lag apital sto k adjustment formula, equation (3). When 
is large (and empiri al estimates suggest it is), this approximation will omit auto orrelated terms. Our
regressions in Se tion 5 are not based on the nite-lag approximation.

5

The main ause of this uptrend is straightforward. Di erent types of equipment depre iate at di erent rates and Oliner (1989, 1994) has shown that omputers depre iate
signi antly faster than other types of equipment. The National In ome and Produ t A ounts (NIPA) apital sto ks used in our analysis are onstru ted under the assumption of
separate, onstant, depre iation rates for ea h of 27 underlying equipment ategories, with
the depre iation rate for omputers taken dire tly from Oliner's resear h.5 Thus, it should
ome as no surprise that the re ent explosion in omputer investment has led to an in rease
in the average rate of depre iation for total equipment. Before we move on to dis uss the
impli ations of a varying pa e of depre iation for e onometri modelling, we need to note
a surprising pattern in our al ulated series for the aggregate depre iation rate.
3.1

A Depre iation Puzzle

While we had expe ted that the high rates of omputer investment would have raised
the aggregate depre iation rate in the 1990s, we were surprised to nd that this was not
just a re ent phenomenon but rather an a eleration of a long-running trend. In fa t, our
al ulated series for the aggregate depre iation rate doubles over 1965-1997. The magnitude
of this apparent mix-shift seems very large, parti ularly as it suggests that variations in the
average rate of depre iation have had an important e e t on aggregate gross investment
throughout the past 30 years, something not found by previous resear hers. The solution
to this puzzle turns out to be a hange in the NIPA methodology for onstru ting real
aggregates.
Sin e 1996, all NIPA real expenditure aggregates, in luding real GDP, have been derived
using a Fisher hain-aggregation methodology.6 Sin e 1997, real apital sto k aggregates
have been onstru ted using the same methodology. Rather than aggregating all quantities
a ording to their base-year pri es, as in the traditional Laspeyres index, the growth rate
of a hained aggregate re e ts a mix of old and new pri es. Given a series of quantities
and pri es for n goods, qi (t) and pi (t), the gross growth rate for the Fisher hain-aggregate
quantity is de ned as:

s P
P p (t 1) q (t)
p
( t ) q (t )
G (t) = P p (t) q (t 1) P p (t 1) q (t 1)
n
i=1

5
6

n
i=1

i

i

n
i=1

i

n
i=1

i

i

i

See Katz and Herman (1997) for a des ription of the NIPA sto ks.
See Landefeld and Parker (1997) for a dis ussion of this methodology.

6

i

i

(9)

In the base-year (1992 in our data), all pri e indexes are set equal to one and the level of
ea h aggregate is set equal to its nominal value. For all subsequent and previous years, the
real level series are simply \ hained" forward and ba kwards using the Fisher-aggregation
growth rates. For NIPA aggregate real equipment investment and sto ks, the Fisher hain
pro edure aggregates 27 omponent series.
This hain aggregation pro edure helps to redu e biases due to valuing goods at pri es
that be ome irrelevant on e we move away from the base year. However, a omplexity
it introdu es is that the level of the onstru ted real aggregate is no longer the additive
sum of its real omponents, with this la k of additivity being most evident when there
are large relative pri e shifts within a bundle of goods (as is the ase with the equipment
bundle be ause of the substantial de lines in the pri e of omputing equipment). This
non-additivity invalidates the al ulation of the aggregate depre iation rate. To illustrate,
onsider the following simple example.
There are two types of apital, A and B . Suppose now that the aggregate apital sto k
is onstru ted a ording to the traditional Laspeyres xed-weight formula. In this ase,
the real aggregates for investment and the apital sto k are the simple sum of their real
omponents: I FW = I A + I B and K FW = K A + K B . It is easy to use this fa t to show that
the aggregate depre iation rate is a weighted average of the two underlying depre iation
rates, with the weights given by the real quantities for the two sto ks:
!
!
KtA 1
KtB 1
ItFW 

KtFW
FW
A
B
Æt
= K FW
= Æ KA + KB + Æ KB + KA
(10)
t 1
t 1
t 1
t 1
t 1
Note, however, that the stri t additivity of the xed-weight formula was ne essary to
obtain this weighted-average expression for the aggregate depre iation rate. On e this additivity breaks down, only in the base year an we interpret the depre iation rate al ulated
from equation (8) as a weighted average; this is be ause in the base year all real series
are equal to their nominal ounterparts and so for this year additivity does hold. In fa t,
as we show in Appendix B, moving away from the base year, the aggregate depre iation
rate al ulated from equation (8) with hain-weighted data di ers systemati ally from a
weighted-average depre iation rate, displaying a long-run upward trend even in the absen e
of mix shifts towards faster depre iating equipment.
The explanation for this result is fairly subtle; a full derivation is available in Appendix
B. However, the intuition is as follows. The growth rate of a hain-weighted aggregate
7

e e tively equals a weighted-average of the growth rates of its omponents, where the
weights are given by the omponents' nominal shares. It turns out that when real investment
in one type of apital grows faster than others be ause its relative pri e is de lining, then
the nominal share of this type of apital in investment will be higher than its nominal share
in the apital sto k, implying that the aggregate for real investment grows faster than the
aggregate for the real apital sto k. As a result, the ratio of the level of real aggregate
investment to the level of the real aggregate apital sto k, will trend upwards. Hen e, the
series al ulated from equation (8) will also trend upwards.
To demonstrate the e e t on the aggregate depre iation rate of the hange in aggregation
methodology, we onstru ted xed-weight aggregates for equipment investment and the
equipment apital sto k by adding up the underlying real series for the 27 equipment
investment ategories. We then al ulated a depre iation rate for these xed weight series,
exa tly as in equation (10). This xed-weight depre iation rate series is shown as the
thi k dashed line on Figure 2 (labeled \Using Fixed-Weight Investment and Capital"). It
rises steeply over the past few years but in reases only very slightly prior to the 1990s,
remaining for most of the sample in the range of 0.13, the value most ommonly used in
studies that onstru t equipment sto ks from a onstant aggregate depre iation rate. In
ontrast, the orresponding series for hain-weighted data limbs steadily from the mid1960s on. Thus, although the in reasing aggregate depre iation rate in the 1990s mainly
re e ts a omposition shift, the uptrend evident prior to the 1990s is mainly an artifa t of
hain aggregation.7
Sin e mix-shifts towards equipment-types that are faster depre iating and de lining in
relative pri e an explain the substantial rise over time in our perpetual-inventory estimate
of the aggregate depre iation rate, an obvious question is whether removing omputing
equipment (whi h depre iates rapidly and has the largest pri e de lines) will result in a
stable depre iation rate. The nal series on Figure 2, the thin dashed line (labeled \NonComputing Equipment, Chain-Weight Investment and Capital"), tells us that the answer
is: Almost. This series was al ulated by applying equation (8) to newly- al ulated hainaggregates for investment and apital sto k for all equipment ex ept omputers, and it
shows a very slow and modest up reep over time.8
7

Be ause previous resear h in this eld used the old xed-weight data, this explains why other resear hers
did not note this urious pattern.
8
Appendix A des ribes how we al ulated these aggregates for non- omputing equipment.

8

3.2

Impli ations for Aggregate Investment Modelling

Returning to the re ent failure of the traditional model, we have seen that the a tual
depre iation rate required to onvert aggregate net investment (kt = KI Æt ) into KI
is not a onstant that an be proxied by the inter ept, as is assumed when we dire tly
estimate equation (7), but in fa t has been rising rapidly in re ent years. If our aim is
a stable e onometri model, then a solution to this problem is to instead dire tly model
the behavior of net investment by estimating equation (5). Figure 3 illustrates how this
step radi ally improves in-sample t. It ompares the residuals from our estimation of
the gross investment ( KI ) model, equation (7), with the residuals from estimation of the
net investment ( KI Æt ) model, equation (5). On e we do not have to a ount for the
variations in the aggregate depre iation rate, we no longer have residuals that trend up over
time and the re ent net investment residuals, though still positive and relatively large, are
not histori ally unpre edented.
This is something of a hollow vi tory, however, if our ultimate goal is a model of gross
investment expenditures. Worse still, these aggregate models annot explain the sour e of
the in reasing aggregate depre iation rate{the explosion in net investment in omputing
equipment. A omplete model of gross investment expenditures in the 1990s must a ount
for the di erent behavior of investment in omputing and non- omputing equipment. In
the next se tion, we present an alternative to the partial adjustment model that provides
intuition for su h an approa h by illustrating why omputer pri es may have a di erent
e e t on investment than other elements of the ost of apital.
t
t

t

t

t

t

1

1

9

1

t

t

1

4

Cost of Capital Sho ks and Capital A umulation

Empiri al tests of the traditional models that we have fo used on thus far nd only a small
role for pri e variables. Therefore, they imply that rapidly de lining omputer pri es have
had little impa t on investment in omputers. The sheer magnitude of the in rease in
omputer investment in re ent years suggests that this may be in orre t. Consider now an
alternative theoreti al approa h, previously presented by Ni kell (1979) and Kiyotaki and
West (1996), that explains why omputer pri e de lines may a e t apital a umulation
more than other sho ks.
The models we have looked at thus far rely on very simple modelling of the e e ts of
adjustment osts. An alternative is to expli itly model the impli ations of adjustment osts
for an optimizing rm with rational expe tations. To apture only the essential features of
the investment problem, we use a quadrati approximation to the underlying pro t fun tion:
Changes in the apital sto k and deviations from the fri tionless optimal sto k both lead to
osts whi h in rease a ording to a simple quadrati fun tion. For a given expe ted path
of k , rms hoose the urrent apital sto k to solve
"
#
1
n
o
X 

2
2 
m kt+m kt+m + (kt+m kt+m 1 )
(11)
Min Et
m=0
where  is the rm's dis ount rate.
The model's rst-order onditions are:  

1

1 

1

1 

kt+1 + 1 + +
k
k = 0
k
(12) 
 t  t 1  t
Letting L be the lag operator,
F bethe lead operator, and using the fa t that the har
1
1 x + 1 = 0 has two roots su h that one root () is
2
a teristi equation x
1 +  +  

Et

1 , this an be re-expressed as
between zero and one while the other equals 

Et
Implying a solution 

(F  

) F

1  

Lkt 

1 
kt = 0 

#
"
1
X 

n 

() kt+n
kt = kt 1 + Et
n=0

(13)

An intuitive re-formulation of this equation that illustrates the model's fundamental
property, omes from using the fa t that  = (1 ) (1 ) (whi h omes from re10

arranging the hara teristi equation). Making this substitution we get 
kt = (1
where

k = (1 

)(kt kt 1 ) 

) Et

t

"1
X

(14)

() kt+n
n

n=0

#

(15)

Thus, ea h period, the log of the apital sto k adjusts towards the moving target, kt , whi h
is a weighted average of expe ted future kt's. It an be shown that  depends positively
on , implying that higher adjustment osts lead to a slower speed of adjustment towards
kt .9
The model is ompleted by a spe i ation of the pro ess for kt . Pro t maximization
(using a generalized CES produ tion fun tion) will give us a rst order ondition: kt = 
t + yt  t , where y and are as before and t summarizes the e e ts of apital-biased
te hnologi al hange. To give a on rete example of what t means, the sto k of omputing
apital may tend to rise independently of output and the ost of apital if the stru ture of
produ tion hanges in ways that fa ilitate in reased usage of omputers.
To implement this model empiri ally, we need to spe ify time series pro esses for output
and the ost of apital. Let yt =  (L) yt 1 + t and t =  (L) t 1 + t where  and  are
m-th order distributed lag polynomials and t and t are white noise. Given these pro esses,
we an solve for the e e ts of output and the ost of apital on kt in terms of empiri ally
observable variables by using the following formula of Hansen and Sargent (1980):
Et

where 

(L) =

Similarly, letting 
(L) =
9

1
1

"1
X

n=0

1 

()

1 

()

()

n

2

41 +
2

41 +

#

yt+n

m
X1
k=1

m
X1
k=1

0

=  (L) yt
m
X 

r=k+1

0 

m
X

r=k+1

(16)
1

()

r k

()r 

r A L
1

k

3
k5

3 

r A Lk 5

(17)
(18)

This type of apital sto k pro ess an also be derived from more general assumptions about te hnology

and adjustment osts: See Auerba h (1989).

11

the pro ess for the apital sto k is now

kt = kt 1 + (1 ) (1 ) ( (L) yt  (L) t ) + ^t

(19)

where ^t depends on a weighted average of urrent and future values of the apital-biased
te hnologi al hange term.
Suppose now we estimate equation (19). The te hnology-bias variable, ^, annot be
observed, so this ends up in the error term. Thus, our estimating equation is

kt = + kt 1 +

XN iyt i + XN i t i + ut
i=0

i=0

(20)

where the model predi ts that

(L) = (1 ) (1 )  (L)
(L) =  (1 ) (1 )  (L)
+ ut = ^t
Equation (20) bears a lose resemblan e to the apital sto k equation under partial
adjustment. However, the oeÆ ients on y and now depend on the variables' own timeseries pro esses and the dis ount rate, , as well as on the underlying produ tion te hnology
and the adjustment speed, . Spe i ally, onsider the long-run elasti ities with respe t to
y and , de ned as the sum of oeÆ ients on these variables divided by (1 ). These
values depend positively on the persisten e of the explanatory variables. In Appendix C,
we show that if is an I (1) series, then (1 )  (1) = 1. But, if is an I (0) series, then
this term is less than 1, and will be approximately zero if is white noise. The reason for
this result is intuitive: Firms are less likely to rea t to sho ks to the \fri tionless optimal"
sto k that they per eive as being temporary than to sho ks per eived to be permanent.10
10

Note that these are onditional elasti ities, not long-run impulse responses of a multiple equation system:
They des ribe the behavior of the apital sto k onditional on the paths of output and the ost of apital.
This ontrasts with the work of Kiyotaki and West (1996). They have also noted that this model an allow
the apital sto k to have di erent elasti ities with respe t to output and the ost of apital. However, their
empiri al implementation imposed the assumption that the ost of apital was an I (1) series, thus ruling
out this possibility. Their implementation of this model instead fo used on long-run impulse responses
of the (k; y; ) system. Their nding of smaller long-run impulse responses to sho ks to omes from
their estimated pro ess for being a less persistent I (1) pro ess than the I (1) pro ess for output (for
instan e although both are I (1) pro esses, yt = 1:5yt 1 0:5yt 2 + t implies larger impulse responses than
yt = 0:5yt 1 + 0:5yt 2 + t ). It does not ome from smaller onditional elasti ities for k with respe t to
than with respe t to y .

12

In light of these results, it is informative to examine the persisten e properties of the
ost of apital for omputing and non- omputing equipment. We de ne the ost of apital
a ording to the standard Hall-Jorgenson rental rate formula:
Ct

= Pt

Rt

P_t

!

Pt

1

IT C

1   

DEP 

(21)

where Pt is the pri e of apital relative to the pri e of output, Rt is the real interest rate,
IT C is the investment tax redit, DEP is the present value of depre iation allowan es per
dollar invested, and  is the marginal orporate in ome tax rate.
Expressed in logs, the ost of apital is the sum of two series{the relative pri e of apital,
and the non-relative-pri e omponent, whi h measures the tax-adjusted gross required rate
of return on investment. As Figure 4 shows, these two omponents a e t the omputer
and non- omputer ost of apital series in very di erent ways. The upper panels show that
the omputer ost of apital is highly non-stationary, exhibiting ontinuous rapid de lines
as a result of the remarkable pattern of falling pur hase pri es. The lower panels show
that the relative stability of the non- omputer ost of apital omes from a ombination
of an uneven de line in the relative pri e of this equipment and a hoppy pattern for the
non-pri e omponent.
Even looking within spe i ategories, the ost of apital ombines omponents that
appear to have very di erent persisten e properties. For instan e, the relative pri e of omputers appears to be a very persistent series; the relative pri e of non- omputing equipment
seems to have a downward trend, although one that is less dominant than for omputers;
the non-pri e omponents for both variables seem to be relatively stable, mean-reverting
series. More formal e onometri hara terizations of the persisten e of these series, using
simple autoregressions and unit root tests, on rm the intuition implied by these graphs.
These tests suggest that the relative pri e series for both omputing and non- omputing
equipment almost ertainly have unit roots, while the non-relative pri e omponents appear
more likely to be stationary series.
There are also good e onomi reasons to believe that the pri e and non-pri e omponents
of the ost of apital have di erent persisten e properties. The pattern of de lining relative
pri es for equipment omes from te hnologi al innovations in the equipment-produ ing
industries, and it seems likely that on e pri es have fallen as a result of innovations, these
pri e redu tions will be permanent. In ontrast, real interest rates will, in the long-run,
13

be related to the marginal produ tivity of apital, whi h will be a stationary variable in
any general equilibrium model. Similarly the Hall-Jorgenson tax term is bounded and has
tended to be mean-reverting.
To summarize, expli itly modelling the e e ts of adjustment osts tells us that the e e t
on investment of sho ks to the ost of apital depends on the per eived persisten e of the
sho ks. We have also shown that the persisten e of the ost of apital varies substantially
a ross equipment type, with the ost of apital for omputers being dominated by the
persistent de line in pur hase pri es. These results suggest using a disaggregated approa h
that allows di erent types of equipment to have di erent elasti ities with respe t to the
ost of apital.

5

E onometri Modelling

5.1

Regressions

We estimated the apital sto k adjustment formula, equation (20), for aggregate equipment
as well as for omputing and non- omputing equipment. Be ause the proposed regressions
ontain nonstationary variables, we rst addressed whether there is a ointegrating relationship. We ran the potential ointegrating regressions and applied Phillips-Ouliaris-Hansen
tests for a unit root in the residuals. We ould not reje t the hypothesis that the error
term has a unit root for any of the three ategories. (This may be be ause our error term
ontains the biased te hnologi al hange term ^t , and it is possible that this term has a unit
root.) These results indi ate that the onventional approa h in the \horsera e" literature
of di eren ing to avoid a spurious regression was probably well-founded.11 We will follow
this approa h in estimating a di eren ed version of equation (20).12
11

For ompleteness, we also estimated our regressions in levels; the important results of this se tion were
un hanged.
12
Note, though, that our approa h of dire tly estimating the apital sto k adjustment equation di ers
from the approa h of the traditional models. These models applied repeated substitution of the lagged
kt term to transform the theoreti al ARMA equation into an MA ( ) equation, and then approximated
this equation using an an MA (n) regression. However, if the adjustment ost parameter, , is high (and
empiri al estimates suggest that it is), then terms omitted in this MA (n) approximation will still have large
oeÆ ients. Sin e these terms are probably positively auto orrelated, we believe that this a ounts for the
poor auto orrelation properties of the traditional models.

1

14

The results are shown in Table 1. The aggregate results ( olumn 1) are familiar from
previous empiri al investment papers. The estimated  of 0.93 implies relatively slow
adjustment. The sum of the oeÆ ients on output is signi antly positive and the sum of
the oeÆ ients on the ost of apital, though negative as expe ted, is quite small. The
long-run elasti ities are shown in the bottom part of the table. For the ost of apital, this
elasti ity is only -0.18.
The se ond olumn of Table 1 shows this regression for omputing equipment. Limited
data availability requires us to estimate over a smaller sample for omputing equipment
(1980-97), whi h leads to less tightly estimated oeÆ ients.13 Nonetheless, this olumn
ontains an important result: The estimated long-run elasti ity of the omputer apital sto k
with respe t to the ost of apital is -1.6, nearly 9 times the estimate from the aggregate

Column 3 reports the results for non- omputing equipment; these are similar to the
aggregate regression.

model.

A ording to the model in the previous se tion, regressors with more persistent time
series pro esses should have higher elasti ities. Thus, part of the explanation for the larger
ost-of- apital elasti ity for omputing equipment ould be that the varian e for the omputer ost of apital is dominated by persistent sho ks (falling omputer pri es). Columns
4-6 examine this hypothesis and provide on rmation. For both omputing and non omputing equipment, the elasti ities with respe t to the more persistent omponents of
the ost of apital (the relative pri e terms) are larger|in the ase of omputers, signifi antly so. Moreover, the long-run investment elasti ity with respe t to omputer pri es
is also statisti ally signi antly larger than the non- omputer elasti ity with respe t to
non- omputer pri es.14
In fa t, by estimating the persisten e properties of the various regressors we an al ulate exa tly how mu h higher the elasti ities on persistent regressors should be. We estimated pro esses for pri e and non-pri e variables for both omputing and non- omputing
equipment, using a stationary representation for the non-pri e variables, and imposing the
13

We hose this starting data be ause the sto k of omputing equipment was very small before 1980. None
of the results reported here are sensitive to the hoi e of sample.
14
The results we have shown in this se tion are robust. Durbin's h statisti s are low indi ating that the
regressions are free of residual auto orrelation. Spe i ation hanges (su h as in luding a trend and adding
extra lags) did not signi antly alter any of our results. Furthermore, the regressions show no eviden e of
parameter instability in the 1990s.

15

assumption that the pro esses for the pri e variables are I (1). Using these pro esses along
with equations (17), (18), and (20), we nd that the ross-equation restri tions implied by
the model tell us that, for both omputing and non- omputing equipment, the onditional
elasti ity of the apital sto k with respe t to the non-pri e variables should equal about half
the elasti ity with respe t to the pri e variables. A Wald test of these ross-equation restri tions reveals that they annot be reje ted. However, be ause of the relative impre ision
of the estimates we are relu tant to pla e too mu h emphasis on these tests.
Our assumption that the relative pri e series are I (1) also implies that the estimated
long-run elasti ities with respe t to these variables should equal the elasti ities of substitution for ea h type of apital. The implied elasti ity of substitution for non- omputing
equipment is -0.33, in line with standard estimates from previous investment studies, although still perhaps surprisingly low. For omputing equipment, the implied elasti ity of
substitution of -1.83 is extremely large. A possible interpretation of this result is that
omputer te hnologies are more easily substitutable for other fa tors.
5.2

Impli ations of Computer Pri e Measurement Error

One question about our large estimate of the elasti ity of omputer net investment with
respe t to its relative pri e is whether it ould be a e ted by errors in the measurement
of omputer pri es. The reasons to suspe t that measurement error may be a e ting this
oeÆ ient are twofold. First, the NIPA omputer pri e index is a onstant-quality series.
This pri e is onstru ted from so- alled \hedoni " pri e regressions, and there is ertainly
room for mis-spe i ation and mis-measurement in these regressions. Se ond, like almost
all NIPA expenditure ategories, real investment in omputing equipment is onstru ted
by de ating the nominal expenditure series by the pri e index. Thus, any measurement
error in the pri e index will a e t both the right- and left-hand sides of our net investment
regression.
While su h measurement error may a e t our regressions, we believe that onsideration
of this fa tor points to a pri e elasti ity for omputing equipment that is larger in magnitude
than our estimate. This is be ause this type of measurement error biases the estimated
long-run elasti ity with respe t to pri es towards minus one and our estimate is -1.83. To
illustrate this result, onsider a simpli ed version of our theoreti al investment equation,

16

without dynami s or non-pri e ost-of- apital terms: 
k = + y
t

pt + t

t

Suppose now that the NIPA pri e, p , is measured with error so that 
p = p + u
t

t

t

The measured real net investment series is the nominal series divided by the measured
pri e: 
k = k + p p
= + y p +  + p p
= + y p + (1 ) (p p) + 
= + y p (1 ) u + 
t

Note now that

t

t

t

t

t

t

t

t

t

t

t

t

t

t

t

t 

p; (1 ) u ) = (1 ) 2
Thus, the sign of the bias in the estimate of depends on the value of itself. If < 1,
then the bias is positive, while if > 1 the bias is negative. Sin e our estimate of the
oeÆ ient on the relative pri e of omputing equipment is greater than one in magnitude,
this suggests that, if measurement error is a fa tor, then the true oeÆ ient is greater in
magnitude than our estimate.15
5.3

Cov (

t

t

t

u

Out-of-Sample Fore asting

Our interpretation of the results in Table 1 is that they are broadly onsistent with the
theoreti al approa h outlined in the previous se tion. However, what of the fa t that
prompted this exploration, the investment boom of the 1990s? To test whether our twoequation pro edure for predi ting net investment helps to explain the re ent behavior of
the apital sto k, we estimated our preferred equations for omputing and non- omputing
equipment (Columns 5 and 6 of Table 1) through 1989:4. We then simulated them out
15

In any ase, we believe the eviden e on NIPA pri e de ators suggests a sanguine interpretation of the
measurement error problem. Re ent resear h by Doms (1999) has shown that pri e de lines measured from
mat hed models (following the pri e of the same ma hine over time) are similar to the NIPA measures based
on hedoni regressions.

17

of sample, taking the realized paths of output and the ost of apital as given, to obtain
simulated apital sto k series for omputing and non- omputing equipment.
Applying hain aggregation to our two simulated apital sto k series, we obtained a
simulated series for the aggregate apital sto k. As shown in Figure 5, the two-equation
system produ es a series (the dotted line) that tra ks the a tual behavior of the equipment
apital sto k (the solid line) in the 1990s mu h better than the out-of-sample simulated series for the aggregate version of the same regression (the dashed line). The series generated
by the aggregate regression, like the in-sample residuals from the aggregate net investment
model in Figure 2, fall further and further behind observed apital sto k growth as the
1990s pro eed. In ontrast, while the disaggregated system underpredi ts a tual apital
sto k growth somewhat for a number of periods from 1993 on, it moves ba k in line by the
end of our sample (1997:4). The reason for the superior tra king performan e of the disaggregated system is intuitive: This approa h allows the massive de line in omputing pri es
to feed through to apital a umulation far more than aggregate e onometri regressions.
More important than the system's ability to tra k the aggregate apital sto k, however,
is its ability to explain the behavior of gross equipment investment. As the perpetual
inventory depre iation rates for omputing and non- omputing are relatively stable over our
sample, we an use a simple out-of-sample fore asting pro edure for gross investment: We
onvert the disaggregated out-of-sample fore asts for apital sto ks into fore asts for gross
investment using the most re ently observed depre iation rates. Applying this pro edure
to our system estimated through 1989:4 produ es gross investment series for omputing
and non- omputing investment. Aggregating these series, we obtain a good des ription of
the re ent behavior of aggregate equipment investment: Our simulated out-of-sample series
for aggregate gross investment grows 6.9 per ent per year over 1990-97, pretty lose to the
observed value of 7.5 per ent. Moreover, as shown in Figure 6, our simulated series (the
dotted line) aptures the move to rapid investment growth in 1992 and the sustained high
rate of growth thereafter. In ontrast, an aggregate model|using the same spe i ation
and the 1989 aggregate depre iation rate|would have averaged about 3.1 per entage points
too low over the period 1990-97 (the dashed line).

18

6

Con lusions

Boosted by exploding investment in omputing equipment, the behavior of equipment investment in the U.S. in the 1990s has been unpre edented. Thus, it should not be too
surprising that the traditional e onometri models of investment, based as they are on histori al orrelations, have ompletely failed to explain the boom. We on lude that these
developments provide three important lessons for ma roe onomists:

Pri es Matter : 

Many previous studies have found limited roles for pri e variables,
stressing the ability of an a elerator model to explain the y li al behavior of investment. In ontrast, we nd an important role for equipment pri es. Spe i ally,
falling omputer pri es played a ru ial role in the investment boom of the 1990s. 

Depre iation Matters :

Most empiri al studies have tended to ignore the role played
by the repla ement of depre iated apital. We have shown that an in reasing depre iation rate was of rst-order importan e in the extraordinary behavior of equipment
investment in the 1990s. Moreover, we have pointed to an important issue in the measurement of depre iation rates: Methodologi al hanges to the NIPAs have made the
standard measure of the average depre iation rate based on aggregate data invalid. 

Aggregation Matters :

Depre iation rates vary widely a ross di erent types of equipment. Also, a model with rational expe tations and adjustment osts tells us that
the e e ts of ost of apital sho ks will not be uniform a ross all types of equipment.
We show that a two-equation system for net and gross investment in omputing
and non- omputing equipment, estimated through 1989, is apable of explaining the
magnitude and pattern of the U.S. equipment investment boom of the 1990s, while
aggregate models ompletely fail.

Put simply, our explanation of equipment investment in the 1990s is that de lining
omputer pri es had a very large e e t in boosting the a umulation of omputer apital.
Consequently, this led to even greater rates of repla ement investment. Ultimately, of
ourse, the true test of any model is its ability to fore ast future developments. We hope
that the future does not turn out to be as unkind to our empiri al approa h as the 1990s
proved to be to the traditional e onometri models.
19

Referen es
[1℄ Auerba h, Alan (1989). \Tax Reform and Adjustment Costs: The Impa t on Investment and Market Value", International E onomi Review, 30, 939-962.
[2℄ Bernanke, Ben, Henning Bohn, and Peter Reiss (1988). \Alternative Non-Nested Spe i ation Tests of Time Series Investment Models", Journal of E onometri s, 37, 293326.
[3℄ Chirinko, Robert (1993). \Business Fixed Investment Spending: A Criti al Survey of
Modelling Strategies, Empiri al Results, and Poli y Impli ations", Journal of E onomi Literature, 1875-1911.
[4℄ Clark, Peter K. (1979). \Investment in the 1970s: Theory, Performan e, and Predi tion", Brookings Papers on E onomi A tivity.
[5℄ Doms, Mark (1999). Hedoni and Mat hed Model Indexes of PC Pri es, mimeo, Federal
Reserve Board.
[6℄ Feldstein, Martin and Mi hael Roths hild (1974). \Towards an E onomi Theory of
Repla ement Investment", E onometri a, 42, 393-423.
[7℄ Gravelle, Jane (1994). The E onomi E e ts of Taxing Capital In ome, Cambridge:
MIT Press.
[8℄ Hansen, Lars Peter and Thomas Sargent (1980). \Formulating and Estimating Dynami Linear Rational Expe tations Models", Journal of E onomi Dynami s and
Control, 2, 7-46.
[9℄ Katz, Arnold and Shelby Herman (1997). \Improved Estimates of Fixed Reprodu ible
Tangible Wealth, 1929-95", Survey of Current Business, May, 69-92.
[10℄ Kiyotaki, Nobuhiro and Kenneth West (1996). \Business Fixed Investment and the
Re ent Business Cy le in Japan", in NBER Ma roe onomi s Annual, 1996, eds., Ben
Bernake and Julio Rotemberg.
[11℄ Landefeld, Steven and Robert Parker (1998). \BEA's Chain Indexes, Time Series, and
Measures of Long-Term E onomi Growth ", Survey of Current Business, May, 58-68.

20

[12℄ Ni kell, Steve (1979). Expe tational Investment Models, London S hool of E onomi s,
manus ript.
[13℄ Oliner, Stephen (1989). \Constant-Quality Pri e Change, Depre iation, and Retirement of Mainframe Computers", in

Pri e Measurements and Their Uses

, ed. Allan

Young, Murray Foss, and Marilyn Manser, Chi ago: University of Chi ago Press.
[14℄ Oliner, Stephen (1994). Measuring Sto ks of Computer Peripheral Equipment: Theory
and Appli ation, mimeo, Federal Reserve Board.
[15℄ Oliner, Stephen, Glenn Rudebus h, and Daniel Si hel (1995). \New and Old Models of
Business Investment: A Comparison of Fore asting Performan e",

Journal of Money,

, 27, 806-826.

Credit, and Banking

[16℄ Shapiro, Matthew (1986). \The Dynami Demand for Capital and Labor",
Journal of E onomi s

Quarterly

, 101, 513-542.

[17℄ U.S. Dept. of Commer e, Bureau of E onomi Analysis (1998). \Fixed Reprodu ible
Tangible Wealth in the United States: Revised Estimates for 1995-97 and Summary
Estimates for 1925-97",

, September, 36-46.

Survey of Current Business

21

Appendi es
A The Data
Our dataset onsists of quarterly series over 1950:1-1997:4 for real output, as well as real
investment, real apital sto k, and the ost of apital for total equipment, omputing equipment, and non- omputing equipment. Our output series is real 1992 dollar output of the
private business se tor, whi h is de ned as GDP minus output from government and nonpro t institutions and the imputed in ome from owner-o upied housing.
Our series on real investment for total equipment is private nonresidential produ ers'
durable equipment expenditures from National In ome and Produ t A ounts (NIPA) Table
5.5. The data for real omputer expenditures is the Computers and Peripherals series from
the same sour e. For real apital sto k series for total equipment and omputing equipment,
we started with the annual NIPA apital sto k data, whi h are available through 1997 and
published in Department of Commer e (1998). These annual data, whi h represent yearend sto ks, were then onverted to quarterly series using an interpolation routine that sets
the growth rate for ea h quarter a ording to its share in the annual total for investment
expenditures.
Our series on real equipment investment and real apital sto k ex luding omputing
equipment were not reated by subtra ting the real series for omputing equipment from
the real aggregates. The la k of additivity of the hain-aggregation formula means that this
is an in orre t al ulation. Rather, in theory, we need to onstru t a new aggregate from the
26 disaggregated non- omputing equipment ategories. In pra ti e, a \ hain-subtra tion"
pro edure whi h applies equation (9) to aggregate equipment and the negative for omputer
investment works just as well and does not require data on 26 investment series.
The ost of apital is measured using the Hall-Jorgenson rental rate formula:
Ct

= Pt

Rt

P_t

!

Pt

1

IT C

1   

DEP 

where Pt is the pri e of apital relative to the pri e of output, Rt is the real interest rate,
IT C is the investment tax redit, DEP is the present value of depre iation allowan es, and 
is the marginal orporate in ome tax rate. The relative pri e series, Pt , are de ned relative
22

to the de ator for private business output. The \ apital gains" term is implemented as a
three-year moving average of the per entage hange in Pt . To onstru t the real interest
rate, Rt , we subtra ted expe ted in ation { proxied by the average in ation rate of the
private business output de ator over the previous ve years - from the nominal rate on Baa
orporate bonds. We then added a onstant \risk premium" that normalized this required
rate of return so that its average equalled the average rate of return on physi al apital
in our sample (6.8 per ent), where this is measured as the ratio of nominal apital in ome
to the nominal apital sto k. The tax term was onstru ted using data on investment tax
redits and servi e lives (used in the al ulation of depre iation allowan es) from Gravelle
(1994). We used Æ = 0:31 for omputing equipment, Æ = 0:13 for non- omputing equipment,
and a nominal- apital-sto k weighted average of these two rates for aggregate PDE. We use
a nominal apital sto k weighted average be ause of the problems with aggregate perpetual
inventory depre iation rates dis ussed in Se tion 3.

B Depre iation Rates with Chained Aggregates
What will the al ulated aggregate depre iation rate from equation (8) look like with hainaggregated data? To keep the analysis transparent, we will look at a simple ase. The Fisher
hain formula is somewhat umbersome, so instead we will use the Tornqvist aggregation
formula. This pro edure weights the growth rate of ea h ategory a ording to its share
in the nominal aggregate, and produ es aggregates with almost identi al properties to the
Fisher pro edure. We will also make the following assumptions. Both types of apital
depre iate at the same rate Æ; the pri e of type-A apital falls at rate relative to the
pri e of type-B apital and output, whi h are both normalized to equal
rms 
 one. 
1 Finally,

A
B
produ e with a Cobb-Douglas produ tion fun tion (Qt = At Kt
Kt
) and there
are no adjustment osts. Now, assuming no taxes, the ost of apital for type A simpli es
to P A (r + Æ + ). The ost of apital for type B is (r + Æ). Given our assumptions, rms
a umulate apital a ording to the rst-order onditions:

KtA =

Qt
A
Pt (r + Æ + )

KtB =

(1 )Qt
(r + Æ)

23

These onditions imply apital sto k growth rates gA = gQ + and gB = gQ . Using the
Tornqvist formula, the growth rate for the hain-aggregated apital sto k is

gCW = (gQ + ) + (1 )gQ = gQ +
where  is the share of apital of type A in the aggregate nominal apital sto k. Note
that these nominal sto ks are de ned as the \repla ement value" of the apital sto k and
are obtained by re ating the real apital sto k for ea h ategory by the urrent-period
pri e of new apital. Now onsider the behavior of a hain-aggregate for real investment.
Re-arranging the expressions for the growth rate of the apital sto k we get:

ItA
= gQ + + Æ
KtA 1
ItB
= gQ + Æ
KtB 1

Thus, for ea h type of apital, the ratio of real investment to the real apital sto k is a
onstant. So, real investment for apital of types A and B also grow at rate gA and gB .
To al ulate the growth rate of the hain aggregate, we need nominal shares of investment:
! B !
! A ! A A!
Kt 1
Kt 1
PtA ItA
ItA
KtB
Pt Kt
=
B
B
A
B
B
A
Pt It
Kt 1
It
Kt 1
Kt
PtB KtB
!
!
!
gQ + Æ +
gQ + 1
PtA KtA
=
gQ + Æ
gQ + 1 + PtB KtB
!
PtA KtA
>
PtB KtB
The share of apital of type A in nominal investment is larger than its share in the
nominal apital sto k. The reason for this is intuitive. The real apital sto k of type A is
growing faster than the real sto k of type B . This means that, measured in today's dollars
at repla ement ost, there is more investment relative to the apital sto k for type A than
there is for type B ; as a result the nominal share of investment for type A is higher. Sin e
real investment of type A grows at rate gQ + while real investment of type B grows at
rate gQ , the growth rate of the Tornqvist hain aggregate for real investment pla es more
weight on the faster growing ategory than does the orresponding growth rate for the
aggregate apital sto k. Hen e, the hain aggregate for investment will always grow faster
24

than the hain aggregate for the apital sto k. This example, in whi h relative pri e shifts
ause the fast growing ategory to have a larger share in nominal investment than in the
nominal apital sto k, lines up pre isely with reality: Computers urrently have a mu h
larger share in nominal equipment investment (14 per ent in 1997) than in the nominal
equipment apital sto k (5 per ent in 1997).
Now, suppose we solve for the aggregate depre iation rate from the hain-aggregates for
investment and the apital sto k:
ÆtCW =

ItCW

KtCW
1

gtCW

Then this value will equal Æ only in the base year. Sin e I CW grows faster than K CW in
ea h period, this \depre iation rate" gets larger ea h period. More generally, if we allowed
the two types of apital to have varying depre iation rates, the depre iation rate estimated
from this equation would only equal a weighted average of the underlying depre iation rates
in the base year, as we move forward from the base year this measure would eventually be
higher than ea h of the underlying depre iation rates.

25

C Omitted Proof
Proof that 

(1) (1 

) = 1 when  (1) = 1:

Inserting the expression for

(1 

(1) we need

2 m 10 m
X X 
) 41 +
k=1

r

r=k+1

()

13
k A5
=1 
r 

()

Re-arranging the left-hand-side of this equation we get

2 m 10 m
X X 
) 41 +

(1

k=1

Now use

(1

r

r=k+1

() 

(1) = 1:

2 m 10 m
X X 
) 41 +
k=1

r

r=k+1

()

13
k A5
= (1 
r

13
k A5
= (1 
r

2 m 10 m
X X 
) 41 +
k=1

" 
) 1 +

X1

r=k+1

!
k
X

m

k=1

1
3 
r A ()k 5

1

r=1 

r

k

#

()

Expanding this expression we get
1 + (1  

1 ) () + (1
(1

=

1 

1 ()

=

1 

() 

1 ) ()2 

2 ()2 

1
(1

::: 

2 ) ()2 + :::::: + (1 
1 

2 ) ()3 

n ()n

as required.

26

::::::::::: 

1 

2
: (1

:: 
1 

n 
2

m

1 ) ()
:: 

n

1
m

1 ) ()

Table 1
Capital Sto k Growth Regressions
(Standard errors in parentheses)

Total

Computers

Ex luding

Total

Computers

Computers

Ex luding
Computers

(50-97)

(80-97)

(50-97)

(50-97)

(80-97)

(50-97)

0.93

0.90

0.92

0.94

0.86

0.93

(.02)

(.04)

(.02)

(.02)

(.05)

(.02)

0.10

0.18

0.11

0.10

0.21

0.10

(.01)

(.13)

(.02)

(.01)

(.10)

(.02)

-0.02

-0.17

-0.01

(.01)

(.09)

(.005)

-0.03

-0.26

-0.03

(.01)

(.10)

(.02)

Cost of apital

-0.01

-0.08

-0.01

without pri es

(.005)

(.08)

(.006) 

Sum of the
CoeÆ ients on:
Output

Cost of apital

Relative pri es

Long-run Elasti ities:
Output

Cost of apital

1.49

1.75

1.33

1.46

1.44

1.33

(.36)

(1.41)

(.31)

(0.37)

(0.84)

(.32)

-0.18

-1.59

-0.13

(.09)

(.75)

(.083)
-0.45

-1.83

-0.33

(.26)

(.45)

(.31)

-0.18

-0.53

-0.14

(.11)

(.50)

(.10)

Relative pri e of apital

Cost of apital
without relative pri e

27

Figure 1
The Demise of the Traditional Investment Regression
Investment relative to the Capital Stock
Data
Fitted Values

0.24

0.22

0.20

28

0.18

0.16

0.14

0.12

0.10

1950

1955

1960

1965

1970

1975

1980

1985

1990

1995

0.08

Figure 2
Perpetual Inventory Depreciation Rates for Aggregate Equipment
Using Chain-Weight Investment and Capital
Using Fixed-Weight Investment and Capital
Non-Computing Equipment, Chain-Weight Investment and Capital

0.17

0.16

0.15

0.14

29

0.13

0.12

0.11

0.10

0.09

0.08

1966

1968

1970

1972

1974

1976

1978

1980

1982

1984

1986

1988

1990

1992

1994

1996

0.07

Figure 3
Residuals from Gross and Net Aggregate Equipment Investment Regressions
Errors Expressed As Percentages of the Capital Stock
Gross Investment Residuals
Net Investment Residuals

0.08

0.06

0.04

30
0.02

0.00

-0.02

1950

1955

1960

1965

1970

1975

1980

1985

1990

1995

-0.04

Figure 4
The Cost of Capital
In Logs
Computer Cost of Capital

31

1970

1980

1990

Relative Price of Computers

Non-Price Component - Computers

5

5

4

4

3

3

2

2

-0.5

1

1

-0.6

0

0

-1

-1

-2

Cost of Capital Excluding Computers

1970

1980

1990

-2

Relative Price of Non-Computing Equipment

-0.3
-0.4

-0.7

1970

1980

1990

Non-Price Component - Excluding Computers

-1.2

0.15

-1.2
-1.3

-1.3

0.10

-1.4

-1.4
-1.5

0.05

-1.5

-1.6

0.00

-1.6
-1.7

-1.7
-0.05

-1.8
1950 1960 1970 1980 1990

-0.8

-1.9

1950 1960 1970 1980 1990

-0.10

-1.8
1950 1960 1970 1980 1990

-1.9

Figure 5
Out-of-Sample (Post-1989) Forecasts of Growth in the Stock of Equipment
Annualized Growth Rates
Data
Aggregate Model (Accelerator with Cost of Capital)
Disaggregated Model (Same Specification)

0.07

0.06

0.05

32

0.04

0.03

0.02

0.01

1988

1989

1990

1991

1992

1993

1994

1995

1996

1997

0.00

Figure 6
Out-of-Sample (Post-1989) Forecasts of Growth in Real Equipment Investment
Year-over-Year Percentage Changes
Data
Aggregate Model (Accelerator with Cost of Capital)
Disaggregated Model (Same Specification)

20

15

10

33
5

0

-5

1988

1989

1990

1991

1992

1993

1994

1995

1996

1997

-10