Structural Change with Two-Level Nested CES

Functions
Zehra Sena BÜYÜKAVCI

14/01/2019

Abstract

In this paper, we try to analyze the effect of a different production function approach,i.e., two-
level nested CES functions, on the context of structural change for the period of 1987-2016
and make a further explanation behind the reasons of it. For this purpose, we model three
different types of two-level nested CES functions and look at how well the data can be fitted for
manufacturing and services. Then, by comparing these CES functions with traditional one-level
nested CES function,i.e., capital and labor as factor inputs, and between each other, we try to
conclude how much two-level nested CES functions can be helpful to fit the data and also to
provide a further explanation on structural change.

1
1 Introduction
In the literature of structural change, capital and labor are generally taken as the factors of pro-
duction and the role of energy as a factor input is nor taken into consideration because energy is
described as an intermediate product rather than a primary input like capital and labor. In this
paper, we want to add energy as a primary factor of production into the production function. The
data in this paper are taken from the Integrated Industry-Level Production Account (KLEMS) of
the Bureau of Economic Analysis from 1987 to 2016. The structural change from this data can be
seen from the graph below as the GDP share of each sector.:

If we determine the sectors of manufacturing and services based on NAICS, change in the share of
each input can be observed as follows.

All values on the graphs are real values, and it will be explained in the data section. Besides the
explanation in the literature of structural change,e.g., by income effect (Kongsamunt et al., 2001),
by the differences in the capital-labor substitution across sectors(Alvarez-Cuadrado et al.,2013), or
by different capital intensities and capital deepening(Acemoglu and Guerrieri, 2008),whether or not
the elasticity of substitution between energy and the other inputs can be another explanation for
the structural change is the main concern of this paper. For doing this, at first three models for
three different types of two-level nested CES functions will be presented and the parameter values
for each CES functions will be calibrated. after that, the traditional one-level CES function will be
added and four function types will be compared to get a better fit for both sectors.

2
2 Methodology
The traditional, one level nested CES production function used in economics has two-factor inputs
and defined as follows:

−ν/ρ
Y = A aX1−ρ + (1 − a)X2−ρ

In this equation; Y is the output, X1 and X2 are the factor inputs which are generally capital and
labor. α is the distribution parameter, α > 0, ρ is the substitution parameter and lastly ν is the
returns scale parameter, ν ≥ 0. The elasticity of substitution for this one level nested production
function is σ = 1/(1+ρ). If the substitution elasticity σ < 0, then the two-factor inputs are comple-
ments and if σ > 0, they are substitutes. In this paper, we use a two-level nested CES production
function adding energy as another factor of production. Energy can be placed inside a nested CES
function in three ways:

KL(E): Capital and labor in an inner (K-L) nest, and energy in an outer nest:
h i−1/ρ
ρ/α
Y = A a (bK −α + (1 − b) L−α ) + (1 − a) E −ρ

EK(L): Energy and labor in an inner (E-K) nest, and labor in an outer nest:
h i−1/ρ
ρ/α
Y = A a (bE −α + (1 − b) K −α ) + (1 − a) L−ρ

LE(K): Labor and energy in an inner (L-E) nest, and capital in an outer nest:
h i−1/ρ
ρ/α
Y = A a (bL−α + (1 − b) E −α ) + (1 − a) K −ρ

Here, we assume constant returns to scale and Hicks-neutral technological change for reducing the
complexity. Again, a and b are the distribution parameters between parameters. σ1 = 1/(1 + α)
is the elasticity of substitution between the inner nested factor of productions and σ2 = 1/(1 + ρ)
is the elasticity of substitution between the outer nested factor input and composite inner nested
factor inputs.
CES functions are non-linear in parameters and for the estimation of the parameters, there are
three approaches. Firstly, CES functions can be linearised in some form or the other, such as an-
alytically linearised Kmenta approximation (Kmenta, 1967). In these approaches, there are some
deficiencies and elasticity of substitution that can be very low or high as Thusrby and Lovell (1978)
stated. The second approach is the cost minimization approach with a particular production func-
tion then derived a linear system of equations from the corresponding optimal input demand. (e.g.
van der Werf, 2008; Okagawa and Ban, 2008; Dissou, Karnizova, and Sun,2014) But in this ap-
proach, comprehensive price data is needed. Because of that, we don’t have a consistent price data
with our data, we use the third approach of the non-linear estimation method. (e.g.Markandya and
Pedroso-Galinato, 2007; Henningsen and Henningsen, 2011; Koesler and Schymura, 2012; Su, Zhou,
Nakagami, Ren, and Mu, 2012) In this approach, we use non-linear ordinary least squares(OLS)
loss function through the minimization of the following sum of squares, as follows:

3
 Pn  2
min S(A, a, b, α, ρ) = i=1 Yi − Ŷi

Because of that we want to estimate the service and manufacturing sector of the US, n equals to 2
and the data cover the period of 1987-2016 we can explicitly write the function as follows:

 h i−1/ρ 2
P2016 −α −α ρ/α −ρ


min S(A,a,b,α, ρ) = t=1987 Yt − A a bXit + (1 − b) Xit + (1 − a) Xit

Where i is the service sector or the manufacturing sector and for both sector we separately solve
this function with derivatives:
dS(A,a,b,α,ρ)
dA =0
dS(A,a,b,α,ρ)
da =0
dS(A,a,b,α,ρ)
db =0
dS(A,a,b,α,ρ)
dα =0
dS(A,a,b,α,ρ)
dρ =0

Our sample size for each sector is 30, i.e., t = 30 and the number of parameters is 5, i.e., n = 5
then the condition of ’t > n’ is satisfied. Hence, we can find a solution to the equations above.
But because of that, this solution cannot be calculated analytically, numerical methods includ-
ing iterative algorithms which require starting values for the parameters to be estimated should
be used. With these starting values, the iterative algorithm tries to reach an optimum through
some optimization methods. In the literature, there several optimizations such as the Levenberg-
Marquart algorithm (Henningsen and Henningsen, 2011), Newton method (Schnabel, Koontz, and
Weiss, 1985), Broyden-Fletcher-Goldfarb-Shanno algorithm (Broyden, Fletcher, Goldfarb, Shanno,
1970).In this paper, we use the Generalized Reduced Gradient algorithm as the default NLS method
of Excel which is one of the most robust nonlinear programming methods to find a solution for the
above equations for all types of two-level nested CES function and both sectors in the US economy.

3 Data
The data are derived from the Integrated Industry-Level Production Account (KLEMS) of the Bu-
reau of Economic Analysis, as it is stated above. BEA’s industry statistics provide a time series of
nominal and real gross output, intermediate inputs (including decomposition of energy, materials,
and purchased services inputs), and value-added by industry; the statistics are based on the 2007
North American Industry Classification System (NAICS). The data consists of 49 industries, i.e.,
21 industries for manufacturing and 28 industries for services, and covers the period between 1997-
2016. Based on Alvarez-Cuadrado, Van Long, Poschke (2014) the sector classification is as follows:

Manufacturing: Farms, Forestry, Fishing, and related activities, Wood products, Nonmetallic
mineral products, Primary metals, Fabricated metal products, Machinery, Computer and elec-
tronic products, Electrical equipment, appliances, and components, Motor vehicles, bodies and
trailers, and parts, Other transportation equipment, Furniture and related products, Miscellaneous

4
manufacturing, Food and beverage and tobacco products, Textile mills and textile product mills,
Apparel and leather and allied products, Paper products, Printing and related support activities,
Petroleum and coal products, Chemical products, Plastics and rubber products.

Services: Utilities, Wholesale trade, Retail trade, Air transportation, Rail transportation, Wa-
ter transportation, Truck transportation, Transit and ground passenger transportation, Pipeline
transportation, Other transportation and support activities, Warehousing and storage, Publish-
ing industries, except internet (includes software), Motion picture and sound recording industries,
Broadcasting and telecommunications, Data processing, internet publishing, and other information
services, Federal Reserve banks, credit intermediation, and related activities, Securities, commod-
ity contracts, and investments, Insurance carriers and related activities, Funds, trusts, and other
financial vehicles, Real estate, Rental and leasing services and lessors of intangible assets, Legal
services, Computer systems design and related services, Miscellaneous professional, scientific, and
technical services, Performing arts, spectator sports, museums, and related activities, Amusements,
gambling, and recreation industries, Accommodation, Food services and drinking places.

In the literature, the choice for output, Y, in energy embedded two-level nested function is made
in four ways: gross domestic product(GDP), gross output, gross value added(GVA) or gross value
added plus energy cost. (Brockway, Heun, Santos, and Barrett, 2017) Based on the data structure
we have, and the statement of a logical measurement somewhere between GVA and gross output
in the capital-labor–energy case, lacking materials, by Kander and Stern(2013) and Van der Werf
(2008)of, we use GVA+energy cost as Y in the OLS loss function. Furthermore, the data consists
of the real values for variables based on a quantity index taking the year 2009 as the base year and
all values for variables appear as 100 in 2009. Hence, we rescale the data in two steps. At first,
the price of value-added of the goods in the farm industry is taken as numeraire and the values for
energy, capital, and labor in millions of dollars for the farm industry are rescaled for the year of
2009. In the second step, all values for all industries are rescaled according to this rescaled farm
industry quantity index. In this way, all variables are taken in comparable forms of real units. The
real-valued data for both sectors with all variables is graphed as follows:

5
4 Estimation Results
In the paper, the estimations are made with the use of the GRG algorithm, as stated above. Specific
starting input values for the parameters are required because of the nature of nonlinear least-squares
estimation and these values were chosen according to the existing literature on a similar estimation
and also grid search was performed for these parameters and the same output was taken. The
results are presented for three types of two-level nested functions:

4.1 KL(E)
Firstly, capital and labor are inner nested, and then energy is added in the function. The specific
non-linear least square estimation for this two-level nested CES function is as follows:
 h i−1/ρ 2
P2016 −α −α ρ/α −ρ


min S(A,a,b,α, ρ) = t=1987 Yt − A a bKit + (1 − b) Lit + (1 − a) Eit

When we solve this non-linear least square estimation function for different grid values, we take the
parameters in the table below:

Manufacturing Services
A 2,524 2,552
a 0,951 0,915
b 0,772 0,990
α 3,277 2,000
ρ 0,354 0,084
σKL,E 0,738 0,922
σK,E 0,234 0,333
SSE 0,035 0,013

Table 1: Parameter Values for KL(E)

When we look at the table, it can be observed that TFP value and the distribution parameter for
energy doesn’t change prominently. The most salient observation in the table is that the service
sector is higher for both values of elasticity of substitution. It means that the substitutability of
energy with composite capital and labor and the substitutability between capital and labor is much
higher in the service sector when we use two-level nested KL(E) CES function.

4.2 EK(L)
In the second type of two-level nested CES function, energy and capital are inner nested, and then
labor is added in the function. The specific non-linear least square estimation as follows:
 h i−1/ρ 2
P2016 −α −α ρ/α −ρ


min S(A,a,b,α, ρ) = t=1987 Yt − A a bEit + (1 − b) Kit + (1 − a) Lit

6
The parameter solution for this OLS loss function is following:

Manufacturing Services
A 2,297 2,586
a 0,910 0,999
b 0,999 0,918
α 2,548 0,099
ρ 0,241 0,608
σEK,L 0,806 0,622
σE,K 0,282 0,910
SSE 0,032 0,013

Table 2: Parameter Values for EK(L)

In this table, TFP values and the distribution parameters for energy change more than the table
of KL(E). This time the elasticity of substitution between labor and composite energy and capital
is higher in the manufacturing sector, but the elasticity of substitution between energy and capital
is much higher in the service sector.

4.3 LE(K)
At last, labor and energy are inner nested, and then capital is added in the function. The specific
non-linear least square estimation for the third type of two-level nested CES function is as follows:
 h i−1/ρ 2
P2016 −α −α ρ/α −ρ


min S(A,a,b,α, ρ) = t=1987 Yt − A a bLit + (1 − b) Eit + (1 − a) Kit

Manufacturing Services
A 2,226 2,535
a 0,187 0,083
b 0,999 0,918
α 0,990 0,001
ρ 1,147 0,095
σLE,K 0,466 0,913
σL,E 0,222 0,586
SSE 0,032 0,013

Table 3: Parameter Values for LE(K)

When we look at the table, the value for the distribution parameter of capital is very high. Again,
the values for elasticity of substitution between labor and energy and between capital and labor-
energy composite is higher in the service sector.

7
5 Comparison and Results
At first, the parameter values for one-level nested function CES must be founded. As can be
seen from the table above, if we use the traditional CES function it gives us the Cobb-Douglas
production function. When we compare the sum of squared errors of one-level nested CES produc-
tion function with two-level nested production functions, it can be observed that it is slightly higher.

Manufacturing Services
A 1,982 2,081
a 0,001 0,985
ρ 39,022 33,624
σ 1,000 1,000
SSE 0,035 0,014

Table 4: Parameter Values for KL

As can be seen in the graph in the data section, the output data may not be stationary. To be
able to see this, we look at the Augmented Dickey-Fuller test at Table 5 and it can be seen, even
though all series, i.e., the data, and the estimated outputs for all function types, for raw data are
nor stationary the first differenced version of all series are stationary. Then it means there is a
trend in series, hence we also look at the differenced series as well.

Manufacturing Services
Raw Data -0.9593 -0.4596
Data
First Differenced -4.5407 -3.4281
Raw Data -0.8601 -0.5825
KL(E)
First Differenced -3.6281 -3.3206
Raw Data -1.4045 -0.5839
KE(L)
First Differenced -3.9358 -3.3232
Raw Data -1.1484 -0.5863
LE(K)
First Differenced -3.7598 -3.3300
Raw Data -1.4641 -0.4935
KL
First Differenced -4.6348 -2.9533

Table 5: Augmented Dickey-Fuller Test Statistics Values

To be able to look at the relationship between the data and the estimated outputs, we calculate the
correlation values for bot raw data and first differenced series after estimating the output with all
of these parameters above tables for all types of CES functions. At the table below, this correlation
coefficient values can be seen
.

8
 Raw Data First Differenced
Manufacturing 0,952 0,441
KL(E)
Services 0,996 0,745
Manufacturing 0,957 0,453
EK(L)
Services 0,996 0,745
Manufacturing 0,956 0,454
LE(K)
Services 0,996 0,746
Manufacturing 0,963 0,580
KL
Services 0,996 0,716

Table 6: Correlation between the Data and the Estimated Values

Together with the tables, to give a better understanding visualization of the data and the model
can be seen as follows for the manufacturing sector:

The same graphs for the service sector are as following:

As can be seen from the tables for the values of the sum of squared errors and correlation, and also
from the graphs, one level-nested KL CES function model is much better for manufacturing. When
we look for services, it can be said that all of the two-level nested function models can be used.
But among them, LE(K) CES function shows a slightly better representation. This conclusion can
be interpreted in the context of structural change as follows: because of that manufacturing sector
can be ideally represented by the Cobb-Douglas production function, energy cannot be used as
a substituted input factor in the production process. On the other hand, the service sector with
two-level nested LE(K) CES production function has a relatively high elasticity of substitution

9
between capital and labor-energy composite. Therefore, this extra elasticity can lead to efficiency
in the service sector and the GDP share for services increases while the one for manufacturing is
decreasing.

10
References
[1] D. Acemoglu and V. Guerrieri. Capital deepening and nonbalanced economic growth.
Journal of political Economy, 116(3):467–498, 2008.

[2] F. Alvarez-Cuadrado, N. Van Long, and M. Poschke. Capital-labor substitution, structural

change and the labor income share. Journal of Economic Dynamics and Control, 87:206–231,
2018.
[3] P. Brockway, M. Heun, J. Santos, and J. Barrett. Energy-extended ces aggregate production:
Current aspects of their specification and econometric estimation. Energies, 10(2):202, 2017.

[4] Y. Dissou, L. Karnizova, and Q. Sun. Industry-level econometric estimates of

energy-capital-labor substitution with a nested ces production function. Atlantic Economic
Journal, 43(1):107–121, 2015.
[5] A. Henningsen, G. Henningsen, et al. Econometric estimation of the “constant elasticity of
substitution” function in r: Package miceconces. Institute of Food and Resource Economics,
University of Copenhagen—FOI Working Paper, 9, 2011.
[6] A. Kander and D. I. Stern. Economic growth and the transition from traditional to modern
energy in sweden. Energy Economics, 46:56–65, 2014.
[7] J. Kmenta. On estimation of the ces production function. International Economic Review, 8
(2):180–189, 1967.
[8] S. Koesler and M. Schymura. Substitution elasticities in a ces production framework-an
empirical analysis on the basis of non-linear least squares estimations. 2012.
[9] P. Kongsamut, S. Rebelo, and D. Xie. Beyond balanced growth. The Review of Economic
Studies, 68(4):869–882, 2001.

[10] A. Markandya and S. Pedroso-Galinato. How substitutable is natural capital? Environmental

and Resource Economics, 37(1):297–312, 2007.
[11] A. Okagawa and K. Ban. Estimation of substitution elasticities for cge models. Discussion
Papers in Economics and Business, 16, 2008.

[12] R. B. Schnabel, J. E. Koonatz, and B. E. Weiss. A modular system of algorithms for

unconstrained minimization. ACM Transactions on Mathematical Software (TOMS), 11(4):
419–440, 1985.
[13] X. Su, W. Zhou, K. Nakagami, H. Ren, and H. Mu. Capital stock-labor-energy substitution
and production efficiency study for china. Energy Economics, 34(4):1208–1213, 2012.

[14] J. G. Thursby and C. K. Lovell. An investigation of the kmenta approximation to the ces
function. International Economic Review, pages 363–377, 1978.
[15] E. Van der Werf. Production functions for climate policy modeling: An empirical analysis.
Energy economics, 30(6):2964–2979, 2008.

11