Econometric estimation of the

physical capital stock∗

José A. Hernández†
University of Las Palmas, Spain

Abstract
The stock of physical capital of an economy is one of the basic economic aggregates.
Yet, it is not observable, since its measurement requires the knowledge of its rate of
depreciation. If we assume that that this parameter is not constant, which is consistent
with the persistent technological progress that bring obsolescence on the vintage capital
stock, then, the measurement of the capital stock poses some technical difficulties. In
this contribution, it is suggested a method to estimate a variable rate of depreciation
by means of the econometric estimation of a production function. The formalization of
this method, empirical evidence and simulation exercises are also presented.

Keywords: Depreciation Rate, Production Function, Spanish Economy.

Introduction
The stock of physical capital of an economy - region, industry or country - is one of the
basic economic aggregates: among other uses it lies at the heart of potential GDP and total
employment calculations, it is required to break down total output among contributing fac-
tors, it is necessary to study the evolution of capital’s productivity, and it also enters as an
argument in the labor productivity function. Yet, this variable is hard to measure directly,
part of the problem deriving from the difficulty in assessing capital consumption - this is
one of the main problems pointed out by the OECD in the elaboration of National Accounts
- see OECD (2001). The depreciation of physical capital is also important to distinguish
between net and gross macroeconomic aggregates. In National Accounts the measure of
capital consumption is usually solved by some ad-hoc method based on accounting regula-
tions at firm level. This is a crude measure, not linked with the technology of the economy
∗
This document is based on the paper ”Econometric estimation of a variable rate of depreciation of the
capital stock”, and was the main reference for the course ”Estimating the capital stock” issued at the Primer
Encuentro Internacional de Matemáticas de la Universidad Sergio Arboleda (EI-MUSA)
†
jhernandez@daea.ulpgc.es
José A. Hernández

neither with the productivity associated to capital. It would be interesting then, to devise
alternative methods. Some alternative ways are based on the use of weighted coefficients
for current and lagged investment into the capital stock equation - see Hulten (1991) and
Jorgenson (1991)-. This methodology does not produce satisfactory results when the econo-
metric approach is used to identify a depreciation pattern – see Hulten and Wykoff (1981) –
and poses serious difficulties. Some of these are solved by ad-hoc methods to calculate these
weighted coefficients of lagged investment based on age-efficiency and age-price functions.
In this paper, an alternative method based on the econometric estimation of the stock of
capital is suggested by means of the estimation of the depreciation rate and the initial cap-
ital stock. The method yields the capital stock computed as a weighted sum of current and
lagged investments, the weights being derived in a more systematic way than in previous
works (see also Dadkiah et. al. (1990), Nadiri et. al. (1993), Prucha (1995 and 1997) and
Prucha et. al. (1996) for related work along the same lines).
The econometric analysis is introduced in the problem of measuring the capital stock based
on the fact that capital stock is one of the basic arguments of the production function.
Then, the capital stock can be estimated indirectly by means of the estimation of the
rate of depreciation -unknown parameter - jointly with the remaining parameters of the
production function itself. It could be argued that, since the capital stock enters as a
variable in other economic equations (for example, demand for labor and investment), they
could also be used to estimate the depreciation rate. However, these relationships, being
more behavioral in nature than the production function itself, which is more technically
based, are more subject to specification errors of several kinds, and therefore, would lead
to less robust results.
The econometric estimation of the capital stock using a production function is based on
the following idea: since the depreciation rate is unknown, the capital stock is not known
either. Then, in econometric terms, the problem can be solved as one of estimation with
unobservable variables. This would amount to trying several values for the depreciation
rate, calculating the corresponding capital stock variables, and finally, fitting the production
function under every capital stock so derived: the depreciation rate and the capital stock,
would then be selected on the basis of standard econometric criteria.
As some authors have pointed out - see Prucha (1995)-, the econometric approach imposes
some methodological restrictions, and it is not immediate to apply standard estimation
methods, part of the problem being that the expression for the capital stock, in terms
of the current and lagged investments, has a time dependent number of arguments, and
this prevents an immediate use of standard packages like TSP. If additional assumptions
are made related to the rate of depreciation parameter, i.e. allowing this parameter to
vary across different periods, then the weights of lagged investments have more complex
specifications, and the use of standard packages becomes even more complicated. Yet, as
theory suggests, it seems natural to consider that the rate of depreciation varies due to
several factors: technical shocks, changes of relative prices of inputs, the level of output -

100
 Econometric Estimation of the Physical Capital Stock

see Prucha and Nadiri (1996)-, changes of the interest rates, the maintenance cost of capital
stock or capital utilization level - see Burnside and Eichenbaum (1994)-. Dynamic factor
demand models provide a context in which the methods could be applied, since the optimal
demand for capital depends on the interest rate and on the relative prices of inputs. The
depreciation rate then, has the role of driving the economic depreciation necessary to adjust
the actual capital stock to the optimal level.
The need for more sophisticated methods to measure capital consumption in order to es-
timate the capital stock, when a variable rate of depreciation is considered, is the key
motivation of this paper. The objectives pursued are the following: First, we describe
two econometric estimation methods of a variable rate of depreciation that are easily imple-
mentable with standard econometric packages, assuming that the depreciation rate depends
linearly on a set of explanatory variables. The first method (Method 1) is based on the com-
plete specification of the coefficients of the lagged investment in the capital stock equation,
and it was originally stated by Prucha (1997) in a general framework, while the second,
(Method 2) takes into account a linear approximation of such coefficients and it is an orig-
inal contribution of this paper. Although this second method leads to an approximated
model – and therefore there is a cost in terms of potential biases –, it is shown to have
advantages in some particular situations, where the implementation of Method 1 has some
difficulties – and costs of using it are larger than those of Method 2-. The second goal of the
paper is to show the empirical effectiveness of both methods, and this is shown with real and
simulated data. Real data analysis is applied to the Spanish economy for the period 1970
to 1997, yielding similar results for both methods and rather different from those currently
accepted. As for simulations, the purpose is to bring some light into the comparison of
the proposed methods, and the identification of the cases where one of the methods could
outperform the other. We find that when more that one variable explains the depreciation
rate, Method 2 yields estimates with better properties in terms of bias, and efficiency. The
opposite is found when the depreciation rate explanatory variables are not stationary, since
in this case the omitted part of the approximation of the model is not negligible. This fact
will be explained in detail in Section 6.
The plan of the paper is as follows: In Section 2 we describe the Baseline method of
estimation of a constant rate of depreciation. In Section 3 we describe the two methods to
estimate a variable rate of depreciation. In Section 4 estimation methods are generalized
for m explanatory variables. Empirical and simulation results are presented in Sections 5
and 6 respectively. Finally, the main results are summarized in a concluding section 7, and
two appendices gather some technical points.

1 Estimation of a constant rate of depreciation

In this section we present the original method suggested by Prucha (1995) to estimate a
constant rate of depreciation. It is the baseline of the methods we develop in the next

101
José A. Hernández

sections to estimate a variable rate of depreciation.

Consider a standard production function given by

Yt = F (Lt , Kt, θ) (1)

where Yt , Lt and Kt denote respectively output, labor and capital stock at the end of period
t, and θ represents a vector of unknown parameters of the technology. The capital stock
accumulates according to the perpetual inventory method equation:

Kt = It + φKt−1 (2)

where It denotes gross investment at period t, φ = 1 − δ and δ is the unknown rate of

depreciation. Repeated substitution of the lagged capital stock in the original equation of
Kt yields
t−1
X
Kt = φi It−i + φt K0 = Gt (I1, ..., It, K0,δ) (3)
i=0

which is a function of t, since the number of arguments of this function depends on the
period index. Now we can substitute (3) into (1) to obtain:

Yt = Ht (Lt, I1, ..., It, K0, δ, θ) (4)

which is also a function of t, because of K. Note that L1 , ..., LT and I1, ..., IT are observable
variables, and, in principle, K0, δ and θ can be estimated by NLS or ML, for instance.
However, it is not possible to use directly standard econometric packages to estimate (4)
because of the number of arguments changes every period. This difficulty can be easily
overcome if we rewrite Kt with the use of dummy variables. First, taking j = t − i,
t
X
Kt = φt−j Ij + φt K0 (5)
j=1

j
Now, for a given t, we define T new variables It , j = 1, ..., T as follows

Itj = Ij Dtj , (6)


1 j6t
Dtj =
0 j>t

This allows as to rewrite (5) as

T
X
Kt = φt−i Iti + φtK0 = G(It1, ..., ItT , K0,δ) (7)
i=1

102
 Econometric Estimation of the Physical Capital Stock

To illustrate what this transformation imply, vector K = (K1, .., KT )′ can be written as
    
K1 I1 K0 0 . . . 0 0 1
 K2   I2 I1 K0 . . . 0 0  φ 
    
 K3   I3 I2 I1 0 0  φ2 
    
 . = . . . . .  . 
    
 .   . . . . .  . 
    
 .   . . . . .  φT −1 
KT IT IT −1 IT −2 . . . I1 K0 φT

Substitution of (7) into (1) yields

Yt = H(Lt, It1, ..., ItT , K0, δ, θ)

Now the number of variables for each period is constant and the parameters can be estimated
using standard econometric packages like TSP. Note that Kt depends on T variables, but
this does not give rise to multicollinearity, since the number of parameters is fixed and the
model is therefore identified. For linear models, this point was stressed by Greene and Seaks
(1991).

2 Estimation of a variable rate of depreciation

In this section we describe two methods to estimate a variable rate of depreciation, and
hence the capital stock so determined. As in the previous case, this estimation is carried
out jointly with the estimation of a set of parameters θ of the production function. The rate
of depreciation is assumed to depend linearly on a variable x. Initially the two methods are
described for a simple case in which the depreciation rate has two components: a constant
term and a variable term that depends on one variable x. This feature does not impose
any restriction on the validity of the methods, which can be described for more complex
specifications – see Section 4 the description of the multidimensional case –. Variable x is
suggested by the economic theory, and some examples are given in Section 1.
Again, to solve the problem of the estimation of the capital stock we consider a production
function
Yt = F̃ (Lt, Kt, θ)
where Kt depends now on a variable rate of depreciation. By assuming a linear equation
for δ we have δt = δ1 + δ2 xt . Note that since the goal of the method described below is the
econometric estimation of δ1 , δ2 and θ, one of its advantages is the possibility to test if the
rate of depreciation is fixed or variable by testing the null hypothesis H0 : δ2 = 0.
The capital stock equation is given next. Calling f1 = 1 − δ1 , and f2 = −δ2 , we have
φt = 1 − δt = f1 + f2 xt. If we substitute recursively Kt−s for all s = 1, ..., t − 1, into (2),

103
José A. Hernández

and taking φt in the place of φ, we have

Kt = It + φt It−1 + ... + φt φt−1 ...φ2φ1 K0 (8)

Let Ct,s be the coefficient of the gross investment at period t − s for all 0 6 s 6 t − 1 and
Ct,t the coefficient of K0. Then,
 s−1
 Qφ 16s6t
t−j
Ct,s = (9)
 j=0
1 s=0

Now (8) can be written as

t−1
X
Kt = Ct,s It−s + Ct,t K0 (10)
s=0
= G̃t (I1, ..., It, δ0, δ1, K0)

Note that when δt = δ for all t, then Ct,s = φs and G̃t(.) = Gt(.) which is the case described
in Section 2. Again, as shown when δ is constant, equation (10) has a number of arguments
that varies with the period considered. Next we describe two methods to solve this problem.

2.1 Method 1
This method is a particular case of the general method shown in Prucha (1997), when
linear dependence is assumed on the depreciation rate. We present it here to make easier
the understanding of its empirical implementation and also the derivation of the following
alternative method. First taking j = t − s, in the original equation of the capital stock, we
have
Xt
Kt = Ct,t−j Ij + Ct,t K0
j=1

and now using Itj as stated in (6),

T
X
Kt = Ct,t−j Itj + Ct,t K0 (11)
j=1

It is simply the immediate extension to the variable case of the original method presented in
Section 2. Since now we have the same number of arguments for each period in the capital
stock equation, the same applies to the production function. Hence, its parameters θ, δ0
and δ1 can be estimated by ML or NLS using standard packages.

104
 Econometric Estimation of the Physical Capital Stock

In order to highlight the increased complexity given by the variability of the depreciation
rate, let us write down the complete capital stock vector in terms of lagged investment and
K0. By construction, the capital stock is now,
             
K1 I1 φ1 K0 φ1 0 0
 K2   I2   φ2   I1   φ2   φ1   K0 
             
 K3   I3   φ3   I2   φ3   φ2   I1 
             
 . = . + . ∗ . + . ∗ . ∗ .  + ...
             
 .   .   .   .   .   .   . 
             
 .   .   .   .   .   .   . 
KT IT φT IT −1 φT φT −1 IT −2

and by defining the Tx1 vectors Fj and I¯j , as follows,

F1 = (φ1, φ2, ..., φT )′ I¯1 = (I1, I2, ..., IT )′

F2 = (0, φ1, ..., φT −1)′ I¯2 = (K0, I1, ..., IT −1)′
. .
. .
FT = (0, 0, ..., 0, φ1)′ I¯T +1 = (0, ..., 0, K0)′

the vector K can be written now, using a compact notation, as,

K = I¯1 + F1 ∗ I¯2 + F1 ∗ F2 ∗ I¯3 + ... + F1 ∗ · · · ∗ FT ∗ I¯T +1 (12)

where ∗ is the element-by-element product of the vectors.

2.2 Method 2
Note that the number of factors in Ct,s is s and that for each t such elements are different.
This fact together with the highly non linearity in the original parameters explain some
difficulties in the convergence of the estimation procedure, possibly due to the intractability
of the criteria used to optimize - maximum likelihood or squared residuals, for instance -.
In that case it would be interesting to investigate an expression for Kt where weights of
the current and lagged investment values had a simpler form and where the cost in terms
of bias and efficiency of the estimates were low. In this context a new method is proposed.
The empirical properties of this method and its comparison with Method 1 are shown with
real data and in several Monte Carlo experiments on sections 5 and 6 respectively. This
method is based on the substitution of the set of coefficients given by (9) by their linear
approximation around the sample mean of x, x̄, in the capital stock equation. The idea
is to simplify the non linearity associated to the cross products of the coefficients φt . It is
shown in the Appendix 1 that the linear approximation of Ct,s yields

105
José A. Hernández

s−1
X
C̃t,s = (f1 + f2 x̄)s−1 ((f1 + f2 x̄) + f2 (xt−j − x̄))
j=0

Although it is not required, we consider now x̄ = 0 into the above equation, to make the
notation and discussion more succinct. By substituting Ct,s by C̃t,s into the capital stock
equation (11), we obtain,

Kt = It + (f1 + f2 xt )It−1 + f1 (f1 + f2 (xt + xt−1 ))It−2 + ... (13)

t−2
X t−1
X
... + f1t−2 (f1 + f2 xt−j )I1 + f1t−1 (f1 + f2 xt−j )K0
j=0 j=0

Now, taking j = t − s, we can rewrite the above equation as

t
X
Kt = C̃t,t−j Ij + C̃t,t K0
j=1

Again, using Itj as defined in (6),

T
X
Kt = C̃t,t−j Itj + C̃t,t K0 = G̃(C̃t,t, ...C̃t,t−T , It1, ..., ItT , K0)
j=1

where C̃t,t−j and xt−j are arbitrary numbers if j > t since for this case we have that Itj
equals zero. Finally, the above equation is plugged into the production function to estimate
the parameters of the model.
It is interesting to write down more explicitly the expression for the capital stock under this
approximation. From (13), the capital stock complete vector can now be written in matrix
form as

       
K1 I1 K0 . . 0 1 I1 x1 K0 . . 0 0
 K2   I2 I1 . . 0  f1   I2 x2I1 . . 0  f2 
       
 . = . . .  . + . . .  . 
       
 .   . . .  .   . .  . 
P .
KT IT IT −1 . . K0 f1T IT xT IT −1 . . ( xt)K0 f1T −1 f2

and using a more compact notation, as

K = AΦ1 + BΦ2 (14)

where A and B are the Tx(T+1) matrixes and Φ1 and Φ2 are the parameters vectors.

106
 Econometric Estimation of the Physical Capital Stock

2.3 Comparing both methods

Here we focus on the differences between Methods 1 and 2 in order to bring some light in the
understanding of our contribution and the situations where there will be possibly success
in its application. By means of our approach, we propose going one step ahead with the
first order approximation of the coefficients of the lagged investment and the initial capital
stock in its original equation. To illustrate what this means, let us consider, for instance,
C2,2 , the true coefficient of K0 in the equation of K2, given by equation (11). Assuming
that x̄ = 0, which is not relevant for comparison purposes but simplifies it, by construction,
following (9), C2,2 = φ2 φ1 , where

φ2φ1 = (f1 + f2 x2) (f1 + f2 x1)

= f12 + f1 f2 (x2 + x1 ) + f22 (x2 x1 ) (15)

which ought to be estimated if we use Method 1. The simple approximation leads to C̃2,2
that is given by
C̃2,2 = f12 + f1 f2 (x2 + x1 ) (16)
which will be a good approximation of C2,2 as long as the variable xt does not deviate too
much from its sample mean and therefore the product x1x2 is of second order of magnitude
- i.e., negligible for practical purposes -. The meaning of ‘too much’ in the preceding
sentence cannot be specified in greater detail beforehand, being itself an empirical matter.
This particular comparison presented above highlights one important difference between
both methods: the capital stock in Method 2, is linear in variables, the non linearity being
left just to the parameters (see (16)), whereas Method 1 is exact, at the cost of retaining
the full non linearity of the problem (see (15)). Although the approximated equation C̃2,2
is still non linear in parameters, it does not retain the original non linearity in parameters.
In terms of the considered example, the element f22 (x2 x1 ) of C2,2 is not included into
C̃2,2 . The trade off is then, exact full non linearity, versus an empirically simpler linear
approximation. We note further, that apart from the log, customarily applied to variables in
econometric specifications, our approach is ‘almost’ linear. This simplification should reduce
the computational burden, and facilitate the convergency properties of the appropriate
algorithms. Whether this holds in practice is an entirely empirical matter, that may also
depend on the specific case considered. We finally note that Method 2 can also be combined
with Method 1, or other numerical procedure for that matter, since it can also be understood
as a first step in the optimization procedure, providing initial values for more involved non
numerical software routines.
From the computational perspective, our simplification gives rise to another significant
difference. With standard econometric software packages such as for example, TSP , which
is the software discussed by Prucha (1995 and 1997), the computation of Method 1 as given
above requires the programming of (T+1) FRML statements for the Fj columns in (12);
in our case, it is enough to supply the program with the matrices A and B of (14) in

107
José A. Hernández

the data set. This, again, reflects the fact that the non linearity involved in Method 1 is
quite considerable, and points to another possible advantage of our method, i.e., its relative
programming simplicity. Also, related to this, the lower probability of falling in errors when
programming, which is an important issue for empirical analysis.

3 Multidimensional case
If variable x has dimension m > 2, Methods 1 and 2 can also be extended to estimate
the parameters involved in the production function together with those behind the rate of
depreciation. Method 1’s extension is immediate, since the modification to take into account
is simply to write down the full coefficients in terms of the lagged investment and then apply
the dummy variable approach to fix the number of arguments in the equation of the capital
stock. Method 2 is also easy to extend, simply by computing the linear approximation of
Ct,s when δ has dimension m.
When x ∈ Rm the rate of depreciation is given by

δt = δ1 + δ2 x1t + ... + δm xmt

and the term φt = 1 − δt = f1 + f2 x2t + ... + fm xmt , where f1 = 1 − δ1 and fj = −δj ,

j = 2, ..., m. The original coefficients of lagged investment is given by

s−1
Y s−1
Y
Ct,s = φt−j = (f1 + f2 x2,t−j + ... + fm xm,t−j )
j=0 j=0

for all s = 1, ..., t and Ct,0 = 1. The first order linear approximation of Ct,s around x = 0
yields -see Appendix 1 for the approximation around x = x̄ –

s−1
X s−1
X
C̃t,s = f1s−1 (f1 + f2 x2,t−j + ... + fm xm,t−j )
j=0 j=0

Replacing Ct,s by C̃t,s in the equation of Kt and at the same time calling j = t − s, we
have,
X t
Kt = C̃t,t−j Ij + C̃t,t K0
j=1

At this point, again, it is applied the dummy variable transformation described in Section 2
to have the same number of argument for Kt for all t. Then it is included in the production
function to finally estimate the parameters under a standard estimation method.

108
 Econometric Estimation of the Physical Capital Stock

4 Empirical Results
In this section empirical results are presented fitting a Cobb-Douglas production function
to the Spanish economy data using Methods 1 and 2. Results are given in Table 1. Returns
to scale are assumed to be constant, since they yield the most coherent estimates values.
Taking logs, the production function becomes,

yt = c + dD1t + αlt + (1 − α)kt + ut (17)

where yt is GDP, c, the constant term, lt , employment, kt, the capital stock, and ut, a
random shock. A dummy variable D1t = 0, t 6 1983, and D1t = 1, t > 1983, is detected to
be significant. The breaking point is picked selecting the best fit among all possible dates.
A time trend turns out non significant, hence technological progress is not included in the
model, but the structural change in the constant term could capture the modernization
process initiated in the Spanish economy in 1982. Hence, the variable D1 could explain the
technological progress that has occurred, not as a constant rate across the sample period,
but notably in the early eighties.
The agricultural sector is excluded, since the resulting sample yields more admissible esti-
mates and it only accounts for a small fraction of GDP (less than 5%), and investment is
negligible (according to input-output tables). Non residential investment is considered only,
and data are measured at 1986 prices. The sample data spans from 1970 to 1997, and is
taken from the Spanish National Statistical Institute (the Tempus data base, available in
the internet). The data require the random error to follow an AR(1). This fact could be
understood as a result of total factor productivity shock1 . The notation is as follows,

ut = ρut−1 + εt (18)
εt ∼ N (0, σε2)

A variable depreciation rate has been considered in order to test the validity of the methods
described. Both Methods 1 and 2 are implemented in order to cross check their results.
The notation used is the following:

Kt = It + (1 − δt )Kt−1 (19)
δt = δ1 + δ2 x2t

Four cases are considered depending on the assumptions made about the variability of δ. In
Case I, δ is assumed to be constant. Although this is not the main focus of the paper, the
results provide baseline model estimates for the more complex specifications. In this case,
there is no distinction between Methods 1 and 2, since there are no explanatory variables
1
Following real business cycle models, a standard assumption for technology consists of a time trend and
an innovation which follows an AR(1) process. In this case the error in the production function follows an
AR(1), what supports our results.

109
José A. Hernández

for δ. Results are shown in the first column of the Table 1. In Case II, x2t = D2t, where
D2t is a dummy variable, defined as D2t = 0, t 6 1989, and D2t = 1, t > 1989. The breaking
point of D2t is picked selecting the best fit among all possible dates. Results are shown
in the 2nd and 3rd columns of the Table 1. In Case III, the GDP growth rate is assumed
to explain δ, in order to check the role of the intensiveness of the use of capital in the
depreciation pattern. Accordingly, a positive demand side shock would increase the use of
the capital stock and thereby increase its rate of depreciation. In this case we would expect
δ2 estimates to be positive. An interesting economic argument to be checked is the effect of
the changes in the rental prices of the capital due to the obsolescence in the depreciation2.
To check this argument one would check the significance of the growth rate of investment
as an explanatory variable of δ, but no satisfactory estimates were found. In case IV the
depreciation rate is assumed to follow the equation δt = δ1 + δ2 x2t + δ3 x3t, where x2t = D2t
and x3t is the GDP growth rate.
The estimation method used in all models is full maximum likelihood, assuming a Gaussian
distribution. In particular, in models involving an AR(1), the first observation is never
dropped (see, for example, Hamilton (1994)). The initial capital stock could be understood
as a further parameter (see section 2), so that beyond entering the model in a non linear
way, it does not pose any special technical problem. Nevertheless, it is not identified (see
Appendix 2), so that it has been judged a more sensible solution to estimate it from alter-
native accounting sources. Accordingly, it has been set equal to 1.8x107 millions of ptas.
at 1986 prices.
We turn now to the discussion of the results themselves, given in Table 1. The average
depreciation rate is always the coefficient δ, since both the dummy and the GDP growth
rate are measured as deviations from their sample means. Column 1 gives the results
obtained for the model with a constant rate of depreciation, – i.e., Case I – using the basic
method of Prucha (1995) which yields an estimate of 3.7%. Since this value must be greater
than zero, a one sided significance test is in order, and it turns out to be significant at the 5%
level (this also applies to the remaining results). This value is rather low when compared to
conventional values (see Corrales et.al.(1989)). Therefore, more complex specifications seem
in order. The remaining results pursue this point, making the depreciation rate dependent
on the mentioned variables (see columns 2 to 7).
In Case II, columns 2 and 3 give the results obtained allowing for a dummy variable, D2t
using Methods 1 and 2 respectively. The dummy is statistically significant when estimation
is carried out using both methods, and it points to an increased depreciation rate in the
second subsample, yielding an average value for the whole sample of 4.27% and 4.23%
(Method 1 and 2 respectively). These values are a priori more admissible, as discussed
above. A significant economic argument to explain the selected breaking year of D2 is
the fact that the interest rate changes its trend in 1990, year in which it reached 14.7%.
From that year on, the interest rate decreased to 10 per cent in 1994. This changing
2
This argument could be found in OECD (2001) as sources of changes in the age-price profile.

110
 Econometric Estimation of the Physical Capital Stock

trend explains variations in the user cost of capital, which start to decrease in the early
nineties and hence, resulted in an intensification of the amortization process, due to the
obsolescence of the vintage capital stock. As shown in columns 2 and 3, the sign of δ2
verifies this conjecture for both methods.
It must be admitted that a dummy, although required by the data, precludes a full economic
explanation. That is why it was deemed appropriate to replace it by the GDP growth rate,
following the argument about the demand side shocks and the intensiveness in the use
of capital explaining δ. The interest rate and the investment growth rate turn out non
significant, while the GDP growth rate is.
Columns 4 and 5 give the results for Case III, and the coefficient δ2 is significant and negative
for both methods. This would mean that a decrease in the GDP growth rate increases the
depreciation rate (and conversely). Since the GDP growth rate decreased in the second
subsample, implying an increased depreciation rate, this is coherent with the previous result
of Case II. Nevertheless, this result is not consistent with the economic argument pointed
out previously, i.e., an increase in demand explains an intensive use of capital and then an
increase in its depreciation rate. Such argument leads us to expect a positive sign for δ2
in opposition with our findings in the estimation. A remark is in order to explain that,
based on particular feature of the Spanish economy. One of the characteristic features of
the traditional Spanish industry is its almost null attention paid to the R&D activities
and to invest in new technologies, what in fact could determine a paradoxical response to
face their decision of replacing capital assets and adding new technology. Such behavior
allocates to the capital stock the feature of being an element for saving cost, mainly focused
on the maintenance cost of capital. Under this viewpoint, in periods where economy is
increasing its activity at a high rate, traditional firms tend to increase its maintenance cost
of capital and then to compensate its depreciation. In this way, firms avoid incorporating
new technology and the expected increase depreciation rate of the vintage capital stock.
This behavior, in turn, explains a decrease in the depreciation rate when output increases.
On the other hand, in periods where the economy faces a declining in its growth rate, firms
decide to scrape old capital, which bring an increase in the depreciation rate, in order to
save maintenance cost of the underutilized capital stock. Finally the Case IV results are
shown in columns 6th and 7th . Here we see that D2 and GDP growth rate are significant –
in fact this last variable is almost significant – what is a strong empirical evidence about
the increasing trend of depreciation rate in time.
In Figure 1 we present a plot of the estimated depreciation rate over time for Cases I, II and
IV (Method 2). Case I shows the results of the estimation of a constant depreciation rate,
while Cases II and IV show the depreciation path when δ is allowed to vary across years. The
average resulting value of the depreciation rate is close to 5.5%. Also, an increasing trend
over time is apparent for these estimates. Note that in Case IV, the estimated depreciation
rate shows to be very cyclical across years, what is due mainly to the variables chosen to
explain δ. Since the growth rate of output of the Spanish economy is very cyclical, the

111
José A. Hernández

resulting estimated depreciation rate also is. Moreover, the dummy variable D2t chosen
to explain δ produces a shift in the estimated δ from 3% to 5% for year 1990, which is a
very strong increase for this parameter in one year. Apart from the explanatory variables,
a second reason to explain the cyclical estimated δ is that estimates values of δ2 and δ3 ,
are very different from zero -relative to magnitudes of variables involved- what pass on the
variability of the GDP growth rate to the depreciation path. Below, we will come back to
this argument, which is closely related with the approach we use to estimate the model.
The resulting path of the estimated δ in Case IV can be objected on the bases of what
should be expected from other sources of estimation of δ for developed countries. Given
the nature of most of the fixed assets that compose the capital stock, whose depreciation
pattern is considered to be very stable across years – see for instance, buildings, we would
expect a less cyclical estimated path for δ.
Nevertheless, an important argument strengthens our findings, based on a relevant property
of the approach we use to estimate δ. The method we use to estimate the production func-
tion assumes that the economy is efficient, i.e., that output is produced with the minimum
required capital level. In other words, the approach excludes the possibility that the econ-
omy could locate in an interior point of the production set. It means that given the output
level of each period, the estimated capital stock is fitted to its efficient level. Hence, the
depreciation rate has to adjust as necessary to fit the estimated capital stock to that efficient
level. Since inefficiency or adjustment costs of capital are not allowed, depreciation rate has
to vary enough in order to reach the efficient capital stock level. The previous argument is
independent of the variables that explain the depreciation rate, and is the counterpart of
the high value of the estimates δ2 and δ3 . As a conclusion, if adjustment costs of the capital
stock or inefficiency were included into the model -which is a very natural assumption for
the Spanish economy- the capital stock will not need to vary rapidly from year to year to
fit the efficient level. In this case, δ2 and δ3 would be closer to zero and hence the depreci-
ation rate would show a more gently increasing trend. The mentioned characteristic of the
used approach can also explain the differences in paths that could be found from different
methods to estimate δ.
The final and complementary aspect of the results concerns the statistical checking of the
estimated equations. Several tests have been conducted to that end: a) functional form, b)
dynamic specification, c) cointegration, d) stability, e) heteroscedasticity, and, f) omitted
variables. They are discussed next in more detail.
The Cobb-Douglas specification has been tested against a more general CES production
function, given by
 −ν/ρ
Y = A γK −ρ + (1 − γ)L−ρ
Since ρ = 0 yields a Cobb-Douglas function, a maximum likelihood ratio test is easily
conducted estimating both forms of the function. Under the null of ρ = 0 the test follows a
χ2 (1). The estimates of ρ in all cases are such that the probability of the parameter lying
below the estimates ρ̂ is smaller than 0.05. Then H0 is not rejected at the confidence level

112
 Econometric Estimation of the Physical Capital Stock

0.95 and the Cobb-Douglas specification is accepted.

Several tests have been conducted to check the integration order of all variables. Overall,
and after accounting for possible breaks and stochastic and deterministic trends, all variables
involved seem to be I(1) (see Perron, (1989)). The errors of the estimated models do not
have a unit root, so that the estimated equations are cointegrated, as required (in all cases).
Finally, and as for the number of cointegrated vectors, the usual analysis yields only one
such a vector (besides, it would not make any economic sense if there were two or more,
since there is no economic link, behavioral, technical, or otherwise, among the variables
considered).
The estimated equations are stable, once the constant dummy is added. This has been
tested by means of two types of tests: a) estimating the model after splitting the sample
in two equal subsamples, and, b) testing the forecasting ability of the equations for the
last two observations. With the a) type test, the null of constancy is accepted, although
the parameters apparently change somewhat. This can be due to a lack of power, because
of the small size of both subsamples, and as a result of a changing depreciation rate (the
parameters are more stable, once this feature is incorporated into the model).
Heteroscedasticity has been tested and rejected in several ways (ARCH tests, and making
the variance depend on a set of assorted explanatory variables). As for omitted variables,
there are more complex theoretical specifications that would grant their introduction (for
example, demand shocks). However, they are of second order explanatory power, and could
make the estimation results less robust.
The main empirical results given in this section can be summarized as follows: 1) the
depreciation rate of the capital stock varies significantly over time, with a likely average
value close to 5.5%; both a dummy and the GDP growth rate are significant explanatory
variables for the depreciation rate, although this last variables loses some significance in a
joint estimation; finally, the depreciation rate increases in the second part of the sample,
possibly pointing to a decrease in the maintenance cost of capital. 2) both Methods, 1 and
2, yield virtually identical results for all practical purposes, so that we can safely conclude
that the bias in the approximation of method 2 is empirically negligible; this reinforces the
computational usefulness of this method, since it is far less non linear than method 1, what
makes it easier to program and converge.

5 Monte Carlo exercise

In this section we describe and implement a simulation exercise in order to validate the
effectiveness of the methods proposed and also to compare them in average terms across a
high number of replications of the same model. Finally we reach conclusions about situations
where Method 2 could be suggested as a more adequate solution in the estimation of a
variable rate of depreciation.
In this simulation exercise we considered the standard model given by a production function,

113
José A. Hernández

the capital stock equation and a linear equation for δ. We took the Spanish economy
observed labor (L) and investment (I) data from 1970 to 1997, used in the previous section.
From this data we generate S = 1000 simulation paths in each Monte Carlo exercise, one
for each case to be described below. We considered three cases. In Case I and Case II,
δ depends on one variable, a time trend or the real interest rate respectively as suggested
by theory. In the Case III, δ depends on two variables, the growth rate of output and the
growth rate of investment. Simulation paths are generated from observed L and I, δ values,
and a vector of parameters, as follows.
i) First, we generate the variable rate of depreciation by taking values δ1 and δ2 in the Cases
I and II, (and δ3 in Case III). Then, the rate of depreciation is given by δt = δ1 +δ2 x2t +δ3 x3t
– the third term is only taken into account in Case III –.
ii) Taking It and δt , the capital stock series Kt is generated following the perpetual inventory
method equation.
iii) We took the parameters values c = 6, α = 0.25, and σε2 = 2.25x10−4 – Following the
same notation used –. The standard deviation of the white noise processes of the production
function equation, σε2 , was chosen to bring a R2 close to 0.97, the value found in the real
data estimation –. Values of the coefficients were taken to be close to the obtained estimates
shown in Table 1.
iv) Finally, we generate S = 1000 simulated output vectors from the technology specification
and from a simulated disturbance εst ∼ N (0, σε2) following the equation

yts = c + αlt + (1 − α)kt + εst

where lt and kt are labor and capital stock in logarithms, and yts is the simulated output at
the period t, for simulation s. Note that in each simulation path, the disturbance is included
only in the production function error term. It means that labor and capital stock remain
constant across simulations and hence across each estimation of the model.
The results of the simulation exercise are S replications of the model, that is, 1000 vectors
of size T of y. This data, together with labor (l), investment (I) and x, (note that K is
not observable) will be taken into account in the estimation of the model to be carried out
by both Methods 1 and 2 described in Section 3. The parameters vector of the model is
(c, α, δ1, δ2, σε2). Also, δ3 is included in Case III.
Estimation are carried out by Methods 1 and 2 and results are shown in Table 2. In Case I,
the true parameter vector is θ0 = (c, α, δ1, δ2, σε2) = (6, 0.25, 0.06, 0.02, 2.25x10−4). Results
show that average estimates of c, α, and σε are close to the true parameters for both
methods, and only for Method 1 estimates of δ1 , and δ2 , are unbiased. A biased estimate
value of δ2 is found for Method 2. The reason for this finding is that the variable explaining
δ is not constant in time. Note that in this case, the nonstationarity of x determines that the
population mean does not exist and hence that the linear approximation of Ct,s around the
sample mean of x was not a good approximation, since the sample mean is not representative
and hence the omitted part in the approximation is not negligible. In Case II, δ depends on

114
 Econometric Estimation of the Physical Capital Stock

the real observed interest rate, r, and the true parameters vector is θ0 = (c, α, δ1, δ2, σε2) =
(6, 0.25, 0.06, −0.02, 0.015). Parameter δ2 is chosen to be negative since r increases the cost
of capital and then decreases the replacement process described by δ. Real interest r varies
notably in the sample period considered, but it could be said that it is stationary in mean.
In this Case, the results in average estimates and standard deviations are equally acceptable
for both methods. This result confirms that the nonstationarity of x in Case I could be
the reason for the observed biased in the estimates of δ2 when using Method 2. In Case
III, δ depends on two variables, x2t = Y∆Y t
t−1
and x3t = I∆I t
t−1
. The true parameters vector
2 −4
is θ0 = (c, α, δ1, δ2, δ3 , σε ) = (6, 0.25, 0.06, 0.02, 0.12, 2.25x10 ). Results show that average
estimates of c, α, and σε2 are close to the true parameters for both methods. In contrast,
results are not so good for parameters δi in any one of the methods. Nevertheless, some
remarks are in order. First, although both methods show biased estimates values, such
biases are not equally significative across methods and parameters. It can be said that
Method 2 bring less biased estimates than Method 1, for all of the parameters, specially
for δ2 and δ3 . Second, we have found a large standard deviation of the estimates of δi ,
compared to Cases I and II, and significant differences in the results between methods.
As shown in columns 5 and 6, the standard deviations of the estimates of δi are notably
smaller for Method 2 than for Method 1. Possibly, this is due to the complexity of the
model considered, transferred to the optimization process, what in the practice means
points estimates very different to the true parameters for Method 1. In short, in Case III, it
can be said that, although Method 2 estimates results are not as satisfactory as they were
for Case II, it performs better than Method 1, specifically in terms of the bias and standard
deviations of the estimates of the parameters involved in δi .
The implemented Monte Carlo exercises yield the following conclusions. 1) In general,
estimates of the simulated model are close to the real value of the parameters, which shows
the feasibility and effectiveness of the described methods. 2) When a nonstationary variable
explains δ, Method 2 yields biased estimates of δ1 and δ2 , and efficiency losses; this clearly
points out to the requirement that the explanatory variables of the depreciation rate do
not deviate too much from their sample means, for the approximation of method 2 to work
adequately. Nevertheless, note that since δ must lie in the interval (0,1), only stationary
explanatory variables are allowed in the long run. 3) When two variables explain δ, results
show in both methods the complexity of the model and the estimation process. Nevertheless,
Method 2 results are more adequate in terms of bias, possibly due to the simplification of
the computational burden that it brings.

6 Conclusions
In this paper we present and discuss two methods to estimate a variable rate of deprecia-
tion of the capital stock with standard econometric packages. The first method is just a
particular case of the one proposed in Prucha (1997), and leads to a fully fledged non linear

115
José A. Hernández

capital stock expression, in variables as well as in parameters. The second and alterna-
tive method that we introduce here, is based on a linear approximation to the capital stock
equation, the idea being to linearize the multiple product of the depreciation rates measured
in different moments in time, around the mean sample values of the explanatory variables
(see section 3.2). This procedure leads to a capital stock formulation linear in variables,
the non linearity being left just to the parameters, as in the constant rate of depreciation
case. This linearization has, in principle, several implications: first, it introduces a bias in
the expression for the capital stock, and thereby in all parameter estimators; how serious
this bias may be, however, is purely an empirical matter. Second, it greatly simplifies the
non linearity of the formulation, making convergency and programming easier, in principle;
again, how useful these properties are, will depend on the specific empirical case considered.
At the very least, however, we can say that both methods can be taken as complementary
in two ways: the estimates of the linearized method could be taken as the starting point
for the fully non linear optimization algorithm of the first method, and the cross check
of the estimates offered independently by both of them, allow a kind of check for global
convergency -which is generally a problem with non linear optimization algorithms -.
In order to check empirically the properties discussed, two empirical exercises have been
conducted: first, both methods are applied to real data from the Spanish economy for
the period 1970-1997, and second, Monte Carlo simulation exercises for several cases are
conducted. Overall, we can say that the results confirm broadly the theoretical guesses
about the relative merits of both methodologies. We summarize first the main results for
the Spanish economy: 1) the depreciation rate of the capital stock varies significantly over
time, with a likely average value close to 5.5%; both a dummy and the GDP growth rate are
significant explanatory variables for the depreciation rate, although this last variables loses
some significance in a joint estimation; finally, the depreciation rate increases in the second
part of the sample, possibly pointing to a decrease in the maintenance cost of capital; 2)
both Methods, 1 and 2, yield virtually identical results for all practical purposes, so that we
can safely conclude that the bias in the approximation of method 2 is empirically negligible;
this reinforces the computational usefulness of this method, since it is far less non linear
than method 1, what makes it easier to program and converge.
As for the results of the Monte Carlo exercise, the main results are the following: 1) In
general, estimates of the simulated model are close to the real value of the parameters
in all cases, which shows the feasibility and effectiveness of both methods. 2) When a
nonstationary variable is added to explain the depreciation rate, Method 2 yields significant
biased estimates of δ1 and δ2 , and efficiency losses; this clearly points out to the requirement
that the explanatory variables of the depreciation rate do not deviate too much from their
sample means, for the approximation of method 2 to work adequately. Nevertheless, note
that since the depreciation rate must lie in the interval (0,1), only stationary explanatory
variables are allowed in the long run. 3) Method 2 is proved to be useful, and even more
adequate, when two variables explain the depreciation rate, where large standard deviation

116
 Econometric Estimation of the Physical Capital Stock

are obtained in Method 1 for estimates of δ2 and δ3. This is likely to be due to the high non
linearity present, specially when several explanatory variables for the depreciation rate are
included, making convergency more difficult to attain (because of possible local optima, or
a flatter likelihood, or both).
As a final comment, perhaps we can say that both methods have their own merits, in
many practical situations being possible to look at them as complementary rather than
alternative. The linearization of method 2 introduced in this paper, should enlarge the
scope for the applicability of econometric estimation of variable depreciation rates, and
the stock of capital, which is an important research topic in National Accounting (see, for
example, OECD related publications)

117
José A. Hernández

Appendix 1: Linear approximation of Ct,s

The first order linear approximation of Ct,s = c(xt, ..xt−s+1) around the sample mean x = x̄
is  ′
Ct,s ≃ c(x̄) + c′(x)|x=x̄ (x − x̄)
When δ ∈ R2 , we want to approximate the coefficient of the lagged investment:
s−1
Y
Ct,s (xt , ..., xt−s+1) = (f1 + f2 xt−j )
j=0

First, valued at x = x̄, we have,

Ct,s (x̄) = (f1 + f2 x̄)s

and for all j = 0, ..., s − 1,


 s−1 
∂Ct,s  Y 
= f2 (f1 + f2 xt−i ) = f2 (f1 + f2 x̄)s−1
∂x 
t−j x=x̄ 
i=0,i6=j
x=x̄

Then,
∂Ct,s
= (f2 (f1 + f2 x̄)s−1 , ..., f2(f1 + f2 x̄)s−1 )′
∂x
and,
s−1
X
C̃t,s = (f1 + f2 x̄)s−1 ((f1 + f2 x̄) + f2 (xt−j − x̄))
j=0

The above equation is the suggested approximation that is considered in the definition of
Method 2. In Sections 3 and 4, we consider the particular case given by x̄ = 0, which may
not be a valid approximation in general, but simplifies the analysis.
The multidimensional case is given by the assumption δt = δ1 + δ2 x2t + ... + δm xmt . Now
Ct,s is
s−1
Y
Ct,s = (f1 + f2 x2,t−j + ... + fm xm,t−j )
j=0

where notation follows that used in the simple case. Now, Ct,s (x̄1, ..., x̄m) = (f1 + f2 x̄2 +
... + fm x̄m )s and the derivative of Ct,s with respect to each variable in x at period t − j, is

∂Ct,s 
= fh (f1 + fh x̄h )s−1
∂xh,t−j xh =x̄h

where h = 2, ..., m. Also,

118
 Econometric Estimation of the Physical Capital Stock

 s−1
X
∂Ct,s  s−1
(xh − x̄ h ) = f h (f + f h x̄h ) (xh,t−j − x̄h )
∂xh xh =x̄h
1
j=0

Then,
 
s−1
X s−1
X
C̃t,s ≃ (f1 +...+fm x̄m )s−1 (f1 + ... + fm x̄m ) + f2 (x2,t−j − x̄2 ) + ... + fm (xm,t−j − x̄m )
j=0 j=0

as shown in Section 4 for the particular case in which x̄ = 0.

119
José A. Hernández

Appendix 2. The lack of identification of the initial capital stock

Consider the simple production function,

log(yt) = A + α log(Kt) + ut

where the capital stock, Kt, is given by,

t−1
X
Kt = K0 φt + φs It−s
s=0

and φ = (1 − δ). If we take K0 as a parameter in the previous equation, the variance of the
OLS estimator of K0, following asymptotic theory, and after straightforward manipulations,
is given by:
   −1
T
P
2
u2t " #−1
∂  T

X
c 2 t=1 
V ar(K0OLS ) ≈ σu  2  = σu2 2α2
(φt /Kt)2
 ∂ (K0)  t=1

c0OLS ) approaches to a positive constant larger

Since φ < 1, as sample size increases, V ar(K
than zero, and the parameter cannot be consistently estimated, or loosely speaking, is
not identified. In other words, as the sample size gets larger, the information about the
parameter does not, just because the associated regressor, φt , goes to zero (see Schmidt
(1976)). Moreover, since Kt is a trending variable, the term φt /Kt goes to zero at a faster
rate than φt and V ar(K c0OLS ) approaches even faster to the constant term. Hence, we
conclude that K0 is not identified, in the sense that, although it can be estimated, its
variance does not approach zero as the sample size increases. In practice, this result means
that we can expect poor estimates for this value.

120
 Econometric Estimation of the Physical Capital Stock

Table 1∗
Production function and depreciation estimation results
with δt = δ1 + δ2 x2t(+δ3 x3t)
Case I Case II Case III Case IV
x2t = 0 x2t = D2t x2t = Y∆Y
t−1
t
x 2t = D2t ; x3t = Y∆Y t
t−1
Basic meth. Meth. 1 Meth. 2 Meth. 1 Meth. 2 Meth. 1 Meth. 2
c0 5.1 5.04 5.05 4.7 4.71 4.64 4.66
(12.9) (12.0) (12.1) (9.37) (9.39) (7.45) (7.54)
c1 .021 .0299 .0297 .0209 .0209 0.0261 0.026
(1.75) (2.36) (2.35) (1.96) (1.94) (2.27) (2.26)
α .344 .349 .349 .392 .391 0.399 0.398
(12.9) (11.9) (12.0) (9.37) (9.4) (7.43) (7.53)
δ1 .0374 .0427 .0423 .049 .0488 0.0541 0.0538
(1.78) (1.77) (1.77) (2.16) (2.16) (2.0) (2.01)
δ2 .0156 .0157 -.577 -.538 0.0147 0.0144
(2.97) (2.91) (-2.24) (-2.15) (2.47) (2.43)
δ3 -.395 -.376
(-1.54) (-1.48)
ρ .63 .378 .382 .766 .756 .592 .586
(3.6) (1.62) (1.64) (5.54) (5.36) (2.59) (2.58)
σε2 1.1x10−4 9.3x10−5 9.3x10−5 9.6x10−5 9.6x10−5 8.5x10−5 8.5x10−5
δm 3.74% 4.72% 4.97% 4.9% 4.8% 5.4% 5.3%

∗ t-values of the estimates in brackets

121
José A. Hernández

Table 2∗
Monte Carlo mean and standard deviation for methods 1 & 2
Case I 1 Case II 2 Case III 3
δt = δ1 + δ2 log(t) δt = δ1 + δ2 r δt = δ1 + δ2 I∆It
t−1
+ δ3 Y∆Y t
t−1
Method 1 Method 2 Method 1 Method 2 Method 1 Method 2
c 6.001 5.996 5.9981 5.9975 6.0268 6.0427
(0.034) (0.038) (0.036) (0.038) (0.3771) (0.4103)
α 0.249 0.251 0.2503 0.2504 0.2468 0.2463
(0.004) (0.005) (0.004) (0.005) (0.0513) (0.0552)
δ1 0.059 0.080 0.0602 0.0603 0.0558 0.0624
(0.003) (0.002) (0.003) (0.003) (0.0384) (0.0391)
δ2 0.020 0.011 -0.0205 -0.0207 0.1741 0.0612
(0.0007) (0.0003) (0.004) (0.004) (0.2754) (0.1629)
δ3 - - - - -0.0955 0.1681
- - - - (0.6503) (0.1937)
σε2 2.2x10−4 1.9x10−4 1.9x10−4 1.9x10−4 1.6x10−4 1.6x10−4
(0.0003) (0.0003) (0.0002) (0.0002) (0.0011) (0.0013)
δm 0.108 0.109 0.060 0.061 0.055 0.062

∗
Typical deviation of estimates in brackets
1
In this case θ0 = (c, α, δ1, δ2, σε2) = (6, 0.25, 0.06, 0.02, 2.25x10−4).
2 In this case θ0 = (c, α, δ1, δ2, σε2) = (6, 0.25, 0.06, −0.02, 2.25x10−4).
3
In this case θ0 = (c, α, δ1, δ2, δ3, σε2) = (6, 0.25, 0.06, 0.02, 0.12, 2.25x10−4)

122
 Econometric Estimation of the Physical Capital Stock

References
[1] Burnside, C. and Eichenbaum, M. (1996) Factor hoarding and the propagation of
business cycles shocks, American Economic Review, 86, 1154-1174.
[2] Corrales, A., y Taguas, D. (1989) Series macroeconómicas para el perı́odo 1954-1988:
Un intento de homogeneización, Ministerio de Economı́a y Hacienda, Madrid.
[3] Dadkiah, K and Zahedi, F. (1990) Estimating a cross country comparison of the capital
stock, Empirical Economics, 25, 383-408.
[4] Greene, W. and Seaks, T (1991) The Restricted Least Squares Estimator: A Pedagog-
ical Note, Review of Economics and Statistics, 73, 563-567.
[5] Hamilton, J. (1994) Time Series Analysis, Princeton: Princeton University Press,
Princeton.
[6] Hulten, C.R. (1991) The Measurement of Capital, in, Fifty years of economic measure-
ment, (Ed) Berndt-Triplett, Chicago.
[7] Hulten, C. R. and F. Wykoff (1981) The estimation of economic depreciation using
vintage asset prices, Journal of Econometrics, 367-396.
[8] Jorgenson, D.W. (1991) Productivity and Economic Growth, in, Fifty years of economic
measurement, Berndt-Triplett (Ed). Chicago.
[9] Nadiri, M.I. and Prucha, I.R. (1993) Estimation of the depreciation rate of physical
and R&D capital in the US total manufacturing sector, National Bureau of Economic
Research, WP 4591.
[10] OECD (2001) Measuring Capital - Measurement of Capital Stocks, Consumptions of
Fixed Capital and Capital Services. OECD Publications Service, Paris
[11] Perron, P. (1989) The Great Crash, the oil price shock, and the unit root hypothesis,
Econometrica, 57, pp. 1361-1401.
[12] Prucha, I.R. (1995) On the econometric estimation of a constant rate of depreciation,
Empirical Economics 20, pp. 299-302.
[13] Prucha, I.R. (1997) Estimation of a variable rate of depreciation: a dummy variable
approach. Structural Change and Economic Dynamics 3, pp. 319–325
[14] Prucha, I. R. and Nadiri, M.I. (1996) Endogenous capital utilization and productivity
measurement in dynamic factor demand models. Theory and application to the U.S.
electrical machinery industry, Journal of Econometrics, 71, 343-379.
[15] Schmidt, P. (1976) Econometrics, Marcel Dekker, New York.

123