Int. J.

Production Economics 241 (2021) 108269

Contents lists available at ScienceDirect

International Journal of Production Economics

journal homepage: www.elsevier.com/locate/ijpe

Assessing the business values of e-commerce and information technology

separately and jointly and their impacts upon US firms’ performance as
measured by productive efficiency
Winston T. Lin a, *, Yueh H. Chen b, Chia-Ching Chou c
a
Department of Operations Management and Strategy, School of Management, The State University of New York at Buffalo, Buffalo, NY, 14260-4000, USA
b
Department of Finance, College of Management, National Sun Yat-sen University, Kaohsiung City, 80424, Taiwan
c
Department of Management, College of Business Administration, Central Michigan University, 202C Smith Hall, Mount Pleasant, MI, 48859, USA

A R T I C L E I N F O A B S T R A C T

Keywords: The business values of information technology (IT) and electronic commerce (EC) have been discussed for years,
Information technology yet virtually all previous studies have evaluated the values of IT and EC independently, instead of jointly; and
Electronic commerce ignored their impacts on the firm’s performance and inputs relationships. This pioneering firm-level study
Theory of production
bridges the gap by empirically investigating the complex relationships among IT and EC investments, relation­
Constant elasticity of substitution
Time-varying stochastic production frontier
ships of inputs, and US firms’ performance at the firm, industry, and sector levels, based on a panel dataset. We
approaches undertake a detailed analysis of the empirical results, seek to answer five major research questions, and discuss
Complementarity and substitutability managerial and strategic implications.

1. Introduction separately and jointly at the firm level and analyzing how the different
assessment processes impact the performance of US corporations.
Information technology (IT) has been broadly invested and adopted Moreover, employing the CES production functions makes it possible
by companies since 1990s, and the increasing investment and applica­ to find out whether the complement and substitution relationships
tion of IT have become one of the major reasons that lead to the among ordinary (non-IT) capital, ordinary labor, and IT capital have
increasing competition among companies (McAfee and Brynjolfsson, impacts upon the performance of US corporations. Hence, five important
2008). As the importance of IT is rising, the significant position of research questions arise, as described below:
electronic commerce (EC) follows (Lin et al., 2015, 2016).
Essentially, IT and EC can be recognized as an integral part of Pro­ (i) How do the business values of IT and EC at the firm, industry, and
duction and Operations Management (POM)/Information Systems sector levels differ if they are assessed separately and jointly?
Management (ISM) (Prasad and Babbar, 2000; Lin et al., 2015, 2016); (ii) How the performance of US manufacturing and service firms is
and IT and EC also play an important role in Production Economics (PE) affected differently if the values of EC and IT are evaluated
(Chen and Lin, 2009; Lin and Chiang, 2011). According to Lin et al. separately and simultaneously? As mentioned above, IT and EC
(2015, Fig. 1, Page 816), IT consists of (a) IT investment on IT hardware, play an important role in PE and are an integral part of POM and
IT software, other office equipment and (b) information systems (IS) ISM; but a careful review of the literatures clearly indicates that
staffing expenses such as IT services and internal IT expenses; whereas the values of IT and EC have been assessed independently rather
EC refers to business-to-business (B2B) (Stender and Ritz, 2006) and than simultaneously (Lin et al., 2015, 2016). As a result, the
business-to-customer (B2C) investments. But previous research has values of IT and EC may be seriously over- or under-stated,
assessed the business values of IT and EC separately instead of jointly leading to miscalculating the performance of US corporations.
(Lin et al., 2015, 2016). Consequently, the values of IT or EC could be (iii) Do the complement and substitution relationships of inputs in­
under- or over-estimated, thereby creating a research gap to be bridged. fluence the performance of the manufacturing and service firms
This study intends to fill the gap by evaluating the values of IT and EC in the United States? The complement and substitution

* Corresponding author.
E-mail addresses: [email protected] (W.T. Lin), [email protected] (Y.H. Chen), [email protected] (C.-C. Chou).

https://doi.org/10.1016/j.ijpe.2021.108269
Received 7 November 2020; Received in revised form 8 August 2021; Accepted 13 August 2021
Available online 16 August 2021
0925-5273/© 2021 Elsevier B.V. All rights reserved.
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

relationships of inputs are a complex and important issue in PE concerning how the complement and substitution relationships of inputs
and POM but are somehow neglected. In the two-factor case, the affect the firm’s performance as measured by productive efficiency, as
relationship between non-IT capital and ordinary labor was well as managerial implications. Section 6 concludes.
recognized by Arrow et al. (1961), Lin and Shao (2006), and Chen
and Lin (2009). In the three-factor case, the importance of the 2. Related literature
relationships among ordinary (non-IT) capital, ordinary labor,
and IT capital in the production and operations processes was Since 1990’s, investments in IT and EC have become more and more
emphasized in the two country-level studies by Chen and Lin prominent in the US companies. For example, the investment in IT
(2009) and Lin et al. (2016) and the firm-level study by Lin and increased at an annual rate of 7.4%, which is more than half of the US
Chuang (2013). economy growth rate (Dewan and Min, 1997; McAfee, 2002). Likewise,
(iv) Since IT is considered as a major driving force of improvements in according to Digital Planet, the amount of the EC transactions in the US
manufacturing and service performance (Triplett and Bosworth, kept increasing significantly from $8 billion in 1996 to $327 billion by
2004; Chou et al., 2012; Lin, 2013), a fourth research question 2002. Therefore, the importance of IT and EC is notable in both business
may be asked: Does the usage of IT along with ordinary capital practices and academic researches.
(non-IT capital) and ordinary labor improve or worsen the per­ A careful literature review indicates that there are two major streams
formance of US firms, industries, and sectors? That is, does the IT in the field of IT/EC studies (Lin et al., 2015), namely, the stream of the
productivity paradox (PP) exist in US firms, industries, and adoption usage of IT and EC and the stream of the assessment of IT and
sectors? EC values. The scholars of the former stream have made contributions in
(v) Does the joint presence of IT and EC in the value-creating pro­ a number o f IT/EC-related issues, including the IT/EC adoption in
duction process enhance or worsen the performance of US firms, supply chain dynamics (Disney et al., 2004), operations management
industries, and sectors? To state alternatively, does the joint (Gunasekaran et al., 2002), mobile contexts (Stender and Ritz, 2006),
presence of IT and EC in the production and operations process purchasing (Cullen and Webster, 2007), supply chain management
create the complementarity (COM) or substitutability (SUB) (Swaminathan and Tayur, 2003; Vakharia, 2002), EC adoption in small
phenomenon? Note that a COM (SUB) phenomenon or relation­ and medium US businesses (Grandon and Person, 2004) and in IS (Gefen
ship means that when IT and EC are invested jointly, the presence and Straub, 2000), relationships between handling of product returns
of IT may enhance (reduce) the value of EC and, hence, COM and customer loyalty in EC (Ramanathan, 2011), and IT adoption in
(SUB), and vice versa (Lin et al., 2015). supporting TQM initiatives (Sánchez-Rodríguez et al., 2006), among
others.
The remainder of the paper is organized as follows. Section 2 is The researchers of the latter stream focus on the assessment of IT and
devoted to a review of related literature. Section 3 describes the theory EC values. They have improved the method used to assess IT and EC by
of production on which this study is based, methodologies, research investigating the issues, such as the impact of EC on transaction cost and
models, built-in performance measures, and estimation method. Section market structure (Lee and Clark, 1996), dynamic benchmarks for op­
4 presents the panel data used and their sources, and the empirical re­ erations evaluation for IT (Warren and Nicholls, 1999), the effect of IT or
sults along with discussion at the firm, industry, and sector levels. Sec­ EC upon business performance (Devaraj et al., 2002; Geoffrion and
tion 5 presents several specification tests on the identification Krishnan, 2003; Gunasekaran et al., 2002), the influence of EC an­
assumptions, a test on the endogeneity of inputs, and a robustness nouncements on the market value (Subramani and Walden, 2001), the
analysis; and summarizes the major findings, including some details impact of IT investment on production efficiency, product quality, and

Fig. 1. The concept of three levels.

2
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

productivity (Thatcher and Oliver, 2001), an analysis of IT and the distributed according to N(0, σ 2w ), X it = (Kit , Lit ) if IT capital is absent,
creation of output and quality in services (Napoleon and Gaimon, 2004), and Xit = (Kit , Lit , Iit ) if IT capital is present, with Kit , Lit , and Iit repre­
a review on IT value and organizational performance (Melville et al., senting ordinary (non-IT) capital, ordinary labor, and IT capital,
2004), activity-based justification of IT investment (Peacock and Tan­ respectively.
niru, 2005), the IT productivity paradox and the impact of IT on pro­
ductivity (Brynjolfsson, 1993; Brynjolfsson and Hitt, 1996; Dewan and
Min, 1997; Melville et al., 2007), relative sizes of IT investment and 3.3. The theory of errors in variables and the generalized time-varying
productive efficiency in stochastic production frontiers (Lin and Shao, SPF approach: two-equation model
2000), complementarity between IT and EC (Zhu, 2004; Zhu and
Kraemer, 2002), complementarity and substitutability among national Chen and Lin (2009) argue that productive inefficiency uit in the one
characteristics and IT value (Chen and Lin, 2009; Lin and Chiang, 2011) equation model (2) may be affected by various controllable or uncon­
and the impacts of complement and substitution of inputs upon the trollable factors. To account for these situations and, therefore, identify
values of IT and EC and, hence, the decision-making unit’s (DMU’s) the sources of the productive inefficiency, they originate the
performance (Lin et al., 2016), among others. two-equation (generalized) time-varying SPF approach described by
Nonetheless, the studies in both streams took a virtually unanimous Yit = f (Xit ; β) − uit + wit , (3)
stand to consider the adaption and assessment of IT and EC separately
instead of jointly; especially, as far as IT and EC values are concerned, uit = g(Zit ; α) + εit , (i = 1, …, m and t = 1, …, n) (4)
they have been evaluated independently rather than simultaneously,
thereby over- or under-stating the values of IT and EC. The country level where equation (3) is exactly the same as equation (2) and uit ( ≥ 0) is
studies of Lin et al. (2015) and Lin et al. (2016) represent two exceptions composed of two elements, namely, g(Zit ; α) and εit . The former is a
in the sense that they have assessed the values of IT and EC jointly and deterministic element which is subject to (determined by) the influence
separately. However, a careful review of the literature indicates that of Zit , a vector representing a broad set of firm- or industry-specific
assessing the values of IT and EC simultaneously and independently and characteristics and macroeconomic factors common to all firms
their effects on the firm’s performance remain lacking at the firm level. considered, observable and/or unobservable, that cause and explain the
In order to bridge the research gap, this firm-level study attempts to differences in productive (in)efficiency across firms. The latter is an one-
empirically evaluate the values of IT and EC separately and also jointly sided (half-) normally distributed random variable. Finally,α, like β, is a
and compare their impacts on the firm’s performance by pursuing the vector of unknown coefficients to be estimated. In sum, the generalized
answers to the above-stated five research questions with novel (two-equation) time-varying SPF approach constituted by Equations (3)
methodologies. and (4) with a stochastic and dynamic inefficiency is considered as a
significant departure from the traditional (one-equation) time-varying
3. Theory, methodologies, research models, performance SPF of equation (2) or equation (3).
measures, and estimation method Here, we would like to explain equation (4) theoretically, mathe­
matically, and statistically to justify the application of equation (4).
Following the work of Lin et al. (2015) and Lin et al. (2016), we According to equation (3) [or (2)], the productive (technical) in­
apply the well-established research methods and models and built-in efficiency (uit ≥ 0) is simply determined by the firm’s internal control­
performance metrics in these two country-level studies to this research lable elements. In other words, the sources of the productive inefficiency
work that takes the firm, industry, and sector as the levels of research. (uit ) are limited to the functional form of the production function, f (⋅),
We now describe the main points relevant to the present study. and the production factors, Xit , entering f (⋅). Such internal limitations
inherent in equation (3) warrant the existence of observation error in uit .
3.1. The theory of production In order to correct such limitations, it is called for the introduction of the
second equation, equation (4), alongside equation (3), to form a two-
As known in Lin et al. (2015) and Lin et al. (2016), the theory of equation model for the generalized time-varying SPF approach as pro­
production can be described by uit = f(Xit ; β) − Yit ≥ 0, or posed by Chen and Lin (2009).
Equation (4) is introduced based on the econometric (economic or
Yit = f (Xit ; β) − uit , (i = 1, …, m and t = 1, …, n) (1) statistical) theory of errors in variables (Kmenta, 1997; Pindyck and
In equation (1), Yit is the actual (observed) output produced by firm i Rubinfeld 1998; Chen and Lin, 2009; Greene, 2012; and Ord et al.,
at time t, Xit is a vector of inputs, β is a vector of unknown coefficients, 2017). The theory states that an observed variable is equal to the sum of
and f(Xit ; β) is the maximum (expected) output which must be greater its true variable and its observation error, meaning that the observed
than or equal to the actual output, and uit denotes the difference between variable is exactly equal to the true variable iff its observation error is
the maximum and the actual output. Moreover, uit ( ≥ 0) is stochastic zero. Therefore, in equation (4), uit is considered as equivalent to the
and non-negative and, hence, is assumed one-sided (or half-) normally observed variable or the observed productive inefficiency obtained from
⃒ ⃒ equation (3), εit is the true variable or the true productive inefficiency,
distributed according to ⃒N(0, σ2u )⃒ (Aigner et al., 1977;Lovell, 1993; Lin,
2009). Since the production unit can either operate on the frontier (uit = and the observation error of the observed productive inefficiency is
0) or below the frontier (uit > 0), uit is the measure of productive represented or measured by the (linear or nonlinear) g( ⋅) function of a
(technical) inefficiency. variable vector of Zit with an unknown coefficient vector of α. Accord­
ingly, the g( ⋅) function can also be interpreted as a correction (adjust­
ment) factor to the observed productive inefficiency uit ; and Z may
3.2. The traditional time-varying stochastic production frontier (SPF)
contain internal forces (e.g., the firm’s EC investments) and/or external
approach: one-equation model
forces (e.g., federal government monetary policy relative to interest
rates). These forces warrant the existence of the observation error in uit
Introducing the conventional random error (wit ) to the right-hand
(Chen and Lin, 2009; Lin and Chiang, 2011). For example, this is
side of equation (1), we arrive at the traditional one-equation time-
particularly true and possible in our modern high-tech and digital era, in
varying SPF model (Aigner et al., 1997, and others) given by
the light of the fact that the firm’s production process can’t be immune
Yit = f (Xit ; β) − uit + wit , (i = 1, …, m and t = 1, …, n) (2) from its own internal activities of EC investments, thereby EC in­
⃒ ⃒ vestments have indirect influence on the firm’s observed productive
where uit ( ≥ 0) is assumed distributed as ⃒N(0, σ 2u ⃒, wit is assumed inefficiency and, hence, the occurrence of the observation error of the

3
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

observed productive inefficiency, as evidenced by the empirical results Model 3 (Case 1): The two-equation two-factor model with Eit
′

of Models 3, 4, and 6 in Section 4. entering the second equation,

To sum up, in equation (4), the observed productive inefficiency, uit ,
Yit = f (Kit , Lit , ; β) − uit + wit
in the one-equation model (3) or (2) must be reformulated as the sum of
a deterministic component specified as a function of vector Zit , linear or
uit = g(Eit ; α) + εit . (i = 1, …, m and t = 1, …, n)
nonlinear, which represents the observation error of the observed pro­
ductive inefficiency, and the true productive inefficiency, εit , where εit is Model 4 (Case 1): The two-equation two-factor model with Iit and Eit
′

a right-sided normally distributed random error. Furthermore, the appearing in the second equation,
choice of the elements of Xit in f( ⋅) is limited by the concept of stock,
Yit = f (Kit , Lit , ; β) − uit + wit
whereas the selection of the components of Zit in g( ⋅) is more flexible,
having a wider variety: (i) ranging from flow to stock, (ii) from observed uit = g(Eit , Iit ; α) + εit . (i = 1, …, m and t = 1, …, n)
to unobserved, (iii) from microeconomic to macroeconomic, (iv) from
Model 6 (Case 2): The two-equation three-factor model with Eit
′
quality to quantity, etc.; and depending on research objectives. There­
fore, consider an example. IT capital may be treated as a determinant of entering the second equation,
inefficiency to be delineated in the error in variable (inefficiency) model Yit = f (Kit , Lit , Iit ; β) − uit + wit
represented by equation (4), e.g., in the second equation of Models 2 and
4. uit = g(Eit ; α) + εit . (i = 1, …, m and t = 1, …, n)
In estimation, equations (3) and (4) in the generalized two-equation
model can be combined into a single equation, that is, uit in equation (3)
is replaced by g(Zit ; α) + εit to get 3.6. The built-in performance metrics of the firm
( )
Yit =f X it; β − g(Zit ;α)− εit +wit (3’) The one-equation and two-equation time-varying SPF approaches
are derived respectively from the theory of production and the theory of
where. εit = f( ⋅) − Yit − g( ⋅) ≥ 0. errors in variables, the six research models are a companion of the time-
If, mathematically, the g( ⋅) function or the error component doesn’t varying SPF approaches, and the performance metrics are built in the
exist; or, statistically, the data produce insignificant estimates of α, then research approaches or models. First, based on equation (2) or equations
εit = uit , equation (4) disappears, and equation (3’) collapses to the (3) and (4), we are able to transform productive (technical) inefficiency
traditional one-equation model (2) or (3). Otherwise, g( ⋅) exists; and uit or εit into productive (technical) efficiency PEit given by
equation (4) and, hence, equation (3’) are justified, meaning that uit in
equation (3) must be replaced by εit as shown in equation (4) and (3′ ). PEit = exp( − uit ), if g( ⋅ ) is insignificant (i = 1, ⋯, m and t = 1, ⋯, n)
Note that our research Models 1 and 5 follow equation (3) or (2), while (5)
Models 2, 3, 4, and 6 follow equation (3’) or equations (3) and (4).
The specification error tests on the identification assumptions asso­ = exp( − εit ), if g( ⋅ ) is significant (i = 1, ⋯, m and t = 1, ⋯, n) (5′ )
ciated with the SPF models conducted in Section 5.1 below strongly for firm i at time t. Since uit or εit lies in the interval [0,∞], PEit lies in
suggest that the two-equation Models 4 and 6 are superior to the one- the range of [0, 1]. Thus, uit and PEit have one-to-one correspondence.
equation Model 1 and that Model 6 > Model 4 > Model 5> Model 2 Second, in order to compare and assess the values of IT and EC and the
> Model 3> Model 1 in terms of the comparisons of the observed values performance of individual firms, we can use the average productive
of the test statistics to the critical values in the Wald and LM tests. efficiency of firm i (APEi ) defined as
Now, we consider two cases. In Case 1, Iit is absent from the CES /
production function; and, in Case 2, Iit is present as a third input of X it in ∑
APEi = PEit n, (i = 1, …, m) (6)
Equation (3). In other words, Iit is not treated as an input variable along t
with Kit and Lit in Case 1, whereas it is used as a production factor along
with Kit and Lit in Case 2. and the overall performance can readily be calculated via
/
∑∑
3.4. Two one-equation research models APE = PEit mn. (7)
i t
Equation (2) yields two one-equation research models designated as
Model 1 and Model 5 :
′ ′

3.7. The CES frontier specifications and the CES-based research models
Model 1 (base of Case 1): The classical or traditional one-equation
′

two-factor model (without investments in EC and IT) given by

As mentioned earlier, the one-equation and two-equation time
Yit = f (Kit , Lit ; β) − uit + wit . (i = 1, …, m and t = 1, …, n) varying SPF models require the deployment of the CES production
functions (two-factor and three-factor). The two-factor and the three-
Model 5 (base of Case 2): The one-equation three-factor model (with
′

factor CES-based SPF models are presented in Appendix A and Appen­

IT Investment but without EC investments) described by
dix B, respectively.
Yit = f (Kit , Lit , Iit ; β) − uit + wit . (i = 1, …, m and t = 1, …, n) It is true that, like other types of production functions available in the
literature, a CES production function is rather dated and traditional and
3.5. Four two-equation research models is known not to be flexible. However, we must develop our research
models based on the CES production functions for the following three
The discussions conducted in the preceding Subsection 3.3 lead to compelling reasons.
four two-equation research models described as follows. First of all, a comprehensive analysis of the impacts of the complex
Model 2 (Case 1): The two-equation two-factor model with Iit substitution and complement relationships of production factors upon
′

appearing in the second equation, the performance of the manufacturing and service firms in the United
States is one of the five research questions as listed at the very outset of
Yit = f (Kit , Lit , ; β) − uit + wit
this work. Only the CES-based research models allow us to answer this
important question by calculating and comparing the substitution and
uit = g(Iit ; α) + εit . (i = 1, …, m and t = 1, …, n)
complement parameters from the CES production functions of two

4
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

inputs and three inputs. Models 1 to 4 are related to Case 1, where a two-factor (Kit − Lit ) CES
Second, reliability and popularity are our second reason. Although function is used, whereas Models 5 and 6 are relative to Case 2, where a
the CES production function is rather dated and is known not to be three-factor (Kit -Lit -Iit ) CES function is involved.
flexible, it is popularly and widely applied in management, productivity
analysis, applied economics, etc. and, because of its solid theoretical 3.8. Method of estimation
foundations, the empirical results from the CES-based models are
considered reliable and useful for managers and policy makers. The six CES-based research models are estimated by the two-step
Third, furthermore, our six (6) CES-based models are linear in co­ nonlinear maximum-likelihood (2SNML) procedure as specified in the
efficients and regressors (data), enabling us to perform the Hausman various versions of LIMDEP statistical package software, including the
(1978) (or Wu, 1973) test on the endogenity of inputs (see Subsection 2016 LIMDEP Version 11 Software, the latest version. The software is
5.4 below). capable of carrying out the OLS, the NLML, the Waldman test, and
The true fixed-effects model as specified in the statistical software productive (in)efficiency (Lin and Shao, 2000). It is noted that the
LIMDEP package is used to test our models. This method comes with a 2SNML method is a panel-data procedure and that we must enter our six
reasonable assumption that allows the inefficiency to change over time, research models specified as input since they are not available in the
separating firm-specific heterogeneity, and provides reliable research LIMDEP Software.
results (Chou et al., 2012). The six unrestricted (non-structural or esti­ More specifically, the stepwise procedure is developed to deal with
mable) CES-based SPF research models (referring to Model 1 to 6
′ ′
panel data. In the first step, the OLS estimates are obtained and used as
above) are given as follows: the starting values for the nonlinear maximum likelihood (NML) method
in the second step (Schmidt, 1985). If the exception condition, i.e.,
Model 1 : ln Yit = β0i + β1 lnKit + β2 lnLit + β4 (lnKit − lnLit )2 − uit + wit , wrong (positive) skewness exists, or if the OLS residuals are positively
(8) skewed, the estimation process would stop and no estimates can be
obtainable (Waldman, 1982). The two-step NML estimates of the un­
1
where β0i = lnγ, β1 = αδ, β2 = α(1 − δ), and β4 = − 2 αδρ(1 − δ) known coefficients are biased and consistent. As such, the two-step NML
method of estimation is both appropriate and valid, especially under the
Model2:lnYit =(β0i − α0i )+β1 lnKit +β2 lnLit +β4 (lnKit − lnLit )2 − α1 lnIit − εit +wit situation where the endogeneity of inputs is absent (see Subsection 5.4
(9) below).
Indeed, the endogeneity of inputs is at the heart of modern discus­
sions on inefficiency. Several researchers have paid attention to the issue

(10)
Model3:lnYit =(β0i − α0i )+β1 lnKit +β2 lnLit +β4 (lnKit − lnLit )2 − α2 lnEit − εit +wit

(11)
Model 4 : ln Yit = (β0i − α0i ) + β1 lnKit + β2 lnLit + β4 (lnKit − lnLit )2 − α1 lnIit − α2 lnEit − εit + wit

by providing its solutions via the route of developing new estimation

(Refer to Appendix A for Models 1 to 4.) methods. For instance, Kutlu (2010) and Kutlu et al. (2019) address the
problem in the maximum likelihood estimation context. Later, following

(12)
Model 5 : ln Yit = β0i + β1 lnKit + β2 lnLit + β3 lnIit + β4 (lnKit − lnLit )2 + β5 (lnLit − lnIit )2 + β6 (lnKit − lnIit )2 − uit + wit

(Refer to Appendix B for the unique correspondence between the un­ Kutlu (2010), Tran and Tsionas (2013) propose a GMM variant of the
restricted or non-structural and restricted or structural coefficients.) Kutlu’s work; and Shee and Stefanou (2015) develop a methodology to
overcome the problem. In short, these methods are good solutions if the

(13)
Model 6 : ln Yit = (β0i − α0i ) + β1 lnKit + β2 lnLit + β3 lnIit + β4 (lnKit − lnLit )2 + β5 (lnLit − lnIit )2 + β6 (lnKit − lnIit )2 − α2 lnEit − εit + wit

where Yit , Kit , Lit , Iit , uit , wit , and εit were defined above; β0i denotes the issue regarding the endogeneity of inputs is properly tested to confirm
parameter to account for cross-firm heterogeneity of firm i; and α0i is an its existence.
unknown coefficient. (Refer to Appendix B for Models 5 to 6.)

5
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 1
Industry groupings (based on the 2007NAICS classifications).
Sector NAICS Industry Firm No. n

Manufacturing 311 Food Manufacturing 18, 30, 34, 49 4

313 Textile Mills 74 1

314 Textile Product Mills 52 1
315 Apparel Manufacturing 12, 44, 77 3
322 Paper Manufacturing 40, 68 2
323 Printing and Related Support Activities 11, 21, 70 3
324 Petroleum and Coal Products Manufacturing 47 1
325 Chemical Manufacturing 2, 7, 8, 14, 22, 23, 24, 29, 32, 38, 43, 50, 64, 65, 67, 76, 80 17
327 Nonmetallic Mineral Product Manufacturing 58 1
331 Primary Metal Manufacturing 75, 79 2
332 Fabricated Metal Product Manufacturing 60, 66 2
333 Machinery Manufacturing 6, 9, 15, 17, 28, 54, 61 7
334 Computer and Electronic Product Manufacturing 3, 5, 13, 19, 25, 27, 33, 35, 37, 41, 46, 48, 51, 53, 55, 56, 57, 62, 63, 71, 72, 78, 23
81
335 Electrical Equipment, Appliance, and Component 1, 26, 69 3
Manufacturing
336 Transportation Equipment Manufacturing 4, 10, 16, 20, 31, 36, 39, 45, 59, 73 10
337 Furniture and Related Product Manufacturing 42 1
Service 423 Merchant Wholesalers, Durable Goods 85, 88, 90, 92, 105, 106, 117, 121 8
424 Merchant Wholesalers, Nondurable Goods 87, 91 2
441 Motor Vehicle and Parts Dealers 84 1
444 Building Material and Garden Equipment and Supplies Dealers 111 1
448 Clothing and Clothing Accessories Stores 93, 103, 125 3
453 Miscellaneous Store Retailers 124 1
454 Nonstore Retailers 118 1
481 Air Transportation 86, 96, 99, 108, 122 5
482 Rail Transportation 97 1
483 Water Transportation 119 1
484 Truck Transportation 110 1
492 Couriers and Messengers 101 1
511 Publishing Industries 94, 102, 113, 115, 116 5
515 Broadcasting 95 1
517 Telecommunications 123, 128 2
518 Data Processing, Hosting, and Related Services 83 1
532 Rental and Leasing Services 104, 120 2
541 Professional, Scientific, and Technical Services 100, 126 2
561 Administrative and Support Services 82 1
562 Waste Management and Remediation Services 129 1
622 Hospitals 127 1
623 Nursing and Residential Care Facilities 109 1
721 Accommodation 107, 112, 114 3
722 Food Services and Drinking Places 89, 98 2

Note: n = the number of firms.

4. Data, data sources, results, and discussion tween the present period’s and last period’s inventories. IT capital is
defined as computer capital +3 * IS labor which has been applied by
4.1. Data and data sources Brynjolfsson and Hitt (1996) and others.
All the data on the relevant variables are expressed in terms of mil­
The dataset consists of 129 US firms from 1999 to 2009, collected lions of 2005 US dollars.
from multiple different databases. The data on actual output (Yit ), or­
dinary capital (Kit ) and ordinary labor (Lit ) were obtained from the
4.2. Results and discussion
Standard and Poor Corporation’s COMPUSTAT database using the
Wharton Research Data Services (WRDS); and the data on EC in­
As mentioned above, this study evaluates the business values of IT
vestments (Eit ) and IT (Iit ) capital were collected from COMPUSTAT and
and EC separately and jointly by comparing the contributions of IT and
Information Week 500 (IW500), respectively. The time period considered
EC to the firm’s productive efficiency in two scenarios - Case 1 and Case
and the number of firms contained in the sample are constrained by the
2. For each case, the results are presented at the firm, sector, and in­
availability of the data on IT. Nevertheless, the panel dataset consists of
dustry levels as follows.
129 × 11 = 1,419 observations.
The 129 firms are classified into manufacturing sector and service
4.2.1. Case 1: two factors (traditional capital and traditional labor)
sector in order to compare how the performance of the manufacturing
sector differs from that of the service sector. Moreover, according to the
first three digits of the 2007 North America Industry Classification 4.2.1.1. Model estimates. Table 2 shows the estimation results of the
System (NAICS) codes, we further classify the 81 manufacturing firms stochastic production frontier models in the settings of Case 1. First, the
into 16 industries and the 48 service firms into 24 industries (see coefficients of both IT and EC are significant at the 1% level, with
Table 1) to compare how the performance differs among those in­ negative signs, implying that their effects upon productive efficiency are
dustries. Fig. 1 shows the concept of the firm, sector, and industry level. positive since, originally, they bear negative signs in Models 2, 3, and 4
Actual output (Yit ) refers to the value added by production factors (i.e., equations (9)–(11)). Moreover, according to the positive substitu­
(inputs) and is measured by the sum of gross sales and difference be­ tion parameter estimates (see Table A), the substitution relationships
between traditional capital and traditional labor exist in all four models

6
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 2
Estimated results of the CES-based SPF models: Case 1 (2-factor case).
Model Parameters APE R2 λ ρ
β1 β2 β3 β4 β5 β6 α1 α2
1 0.72*** 0.14*** – − 0.05*** – – – – 0.69 0.78 2.93*** ρKL 0.82
2 0.22*** 0.21*** – − 0.01*** – – − 0.49*** – 0.76 0.88 6.60*** ρKL 0.26
3 0.62*** 0.15*** – − 0.04*** – – – − 0.20*** 0.73 0.86 1.38*** ρKL 0.67
4 0.22*** 0.22*** – − 0.01*** – – − 0.40*** − 0.14*** 0.79 0.92 5.55*** ρKL 0.22

Notes: *** demotes significant at the 1% level; ρKL denotes the substitution parameter for ordinary capital and labor.

Table 3
Estimated results of the CES-based SPF models: Case 2 (3-factor case).
Model Parameters APE R2 λ ρ
β1 β2 β3 β4 β5 β6 α1 α2
1 0.72*** 0.14*** – − 0.05*** – – – – 0.69 0.78 2.93*** ρKL 0.82
5 − 0.19*** 0.10*** 1.01*** − 0.01*** − 0.06*** 0.05*** – – 0.77 0.89 10.25*** ρKL − 0.69
ρLI 1.16
ρKI 0.53
6 − 0.15*** 0.11*** 0.89*** − 0.001*** − 0.05*** 0.06*** – − 0.14*** 0.80 0.92 7.23*** ρKL − 0.10
ρLI 0.88
ρKI 0.73

Notes: *** demotes significant at the 1% level; ρKL denotes the substitution parameter for ordinary capital and labor; ρLI denotes the substitution parameter for labor
and IT capital; ρKI denotes the substitution parameter for ordinary capital and IT capital.; and *** denoted significant at the 1% level.

(0.82, 0.26, 0.67, and 0.22 for Models 1, 2, 3 and 4, respectively). and EC investments (Model 4) exert the greatest impact on the firm’s
Second, both IT and EC tend to increase the overall average pro­ performance as measured by productive efficiency, followed by IT in­
ductive efficiency (APE), regardless of whether they are separately or vestment only (Model 2) that is followed by EC investments only (Model
jointly present. For example, in the absence of both IT and EC in­ 3); and deploying no IT and EC investments (Model 1) is the worst
vestments, the APE is 0.69 (Model 1). However, when IT and EC are strategy.
considered as the determinants (corrections or adjustments) of in­ However, the results of some companies on the list of those non-first-
efficiency separately, the APEs increase to 0.76 (Model 2) and 0.73 place winners tell the different stories. Take firm 34 in the group of
(Model 3), respectively; but when IT and EC are present jointly, the APE manufacturing firms as an instance. In the absence of IT and EC in­
is 0.79 (Model 4). It is obvious that, overall, there doesn’t exit the IT PP. vestments, its APE is 0.70 (Model 1); when IT investment is made
And the results indicate that as IT and EC are assessed alone, the APEs separately, its APE is 0.57 (Model 2, the smallest); while if EC in­
increase by 10% (from 0.69 to 0.76) and 5.8% (from 0.69 to 0.73), vestments are engaged independently of IT, its APE is 0.73 (Model 3, the
respectively, in comparison with the situation in which the APE in­ largest); and when IT and EC are invested jointly, its APE is 0.60 (Model
creases by 14.5% (from 0.69 to 0.79) as IT and EC are assessed simul­ 4). Consider firm 37 in the group of manufacturing firms as a second
taneously. The empirical evidence clearly suggests that the contribution example. The results of this firm also imply varying effects of IT and EC
of IT to APE is greater than that of EC, meaning that IT is more powerful adoptions: in the absence of IT and EC, its APE is 0.58 (Model 1); yet,
in increasing APE than EC, and that the complementarity (COM) phe­ when IT is treated separately from EC, its APE is only 0.52 (Model 2, the
nomenon exists since the APE from the joint presence of IT and EC smallest); but when EC investments are made separately from IT, its APE
(Model 4) is larger than its counterparts as IT and EC are present sepa­ is 0.70 (Model 3, the largest); and as IT and EC are made jointly, its APE
rately (Model 2 and Model 3), implying that the value of IT and EC are is 0.57 (Model 4). Consequently, both firms 34 and 37 appear to suffer
under-stated if they are assessed independently rather than jointly. from the productivity paradox (PP) of IT, but face the COM phenomenon
created by the joint presence of IT and EC.
4.2.1.2. Firm level analysis. At the firm level, both IT and EC tend to Likewise, some firms in the group of service firms also demonstrate
show positive effects on APEs no matter whether they are assessed the different effects of IT and EC investments upon the performance of
separately or jointly in both manufacturing firms (Table 4) and service firms. Take firm 83 as an example. In the absence of IT and EC, the firm’s
firms (Table 5). The results are consistent with the empirical evidence of APE is 0.71 (Model 1); as IT is treated independently of EC, its APE is
model estimates presented above. Here, we undertake an analysis of the 0.59 (Model 2, the smallest); when EC is engaged separately from IT, its
individual firms with great details. APE becomes 0.73 (Model 3, the largest); and, as IT and EC are treated as
Analyzing the empirical results at the firm level, we are able to tell inefficiency corrections (adjustments) jointly, its APE decreases to 0.61
which firm benefits most from the IT and/or EC adoption. Among the (Model 4). Thus, firm 83 is featured by (i) the IT PP, (ii) IT and EC are
manufacturing firms, firm 20 wins the first place in the absence of both complementary, and (iii) EC is more powerful to increase its APE than
IT and EC (Model 1: APE20 = 0.71); firm 65 benefits most from its IT IT. As a second example, consider firm 90. The absence of IT and EC
adoption (Model 2: 0.84); firm 49 ranks the first because of its EC leads to its APE of 0.69 (Model 1); however, as IT is made alone, its APE
adoption (Model 3: 0.73); and firm 80 is the clear winner in the joint is 0.63 (Model 2, the smallest); as EC is valued alone, the firm’s APE is
presence of IT and EC (Model 4: 0.85). In contrast, among the service 0.72 (Model 3, the largest); and as IT and EC are made jointly, its APE is
firms, firm 84 ranks the first place in the absence of both IT and EC 0.66 (Model 4). Therefore, firm 90 is under the same conditions as firm
(Model 1: APE84 = 0.71); firm 113 benefits most from its IT adoption 83.
(Model 2: 0.84); firm 105 benefits most from its EC adoption (Model 3: Thus, based on the empirical results at the firm level, apart from
0.73); while firm 91 is the biggest winner in the joint presence of IT and some exceptions (such as 34, 37, 83, and 90) mentioned above, the
EC (Model 4: 0.84). The results of the first-place winners in Models 1 to 4 positive contributions of IT and EC to APEs are confirmed. Moreover, the
strongly imply that in both manufacturing and service firms, the joint IT contributions of IT and EC to APEs vary when IT and EC are assessed

7
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 4
Comparisons of the APEi’s of firms at the firm and sector levels: manufacturing (Case 1).
Firm No. Model 1 Model 2 Model 3 Model 4

APEi Rank APEi Rank APEi Rank APEi Rank

1 0.70 28 0.82 12 0.73 35 0.84 13

2 0.69 48 0.75 57 0.73 15 0.82 26
3 0.69 42 0.82 15 0.73 58 0.82 24
4 0.70 12 0.70 71 0.73 50 0.77 66
5 0.68 62 0.82 13 0.73 59 0.83 14
6 0.67 70 0.73 65 0.72 72 0.76 69
7 0.68 60 0.73 61 0.73 29 0.80 47
8 0.68 55 0.77 42 0.73 32 0.79 52
9 0.70 24 0.80 22 0.73 13 0.84 9
10 0.70 33 0.77 39 0.73 39 0.80 45
11 0.70 13 0.82 14 0.73 21 0.84 8
12 0.70 9 0.83 3 0.72 61 0.81 30
13 0.71 5 0.81 16 0.73 17 0.83 19
14 0.69 43 0.75 55 0.73 8 0.82 25
15 0.69 50 0.78 34 0.73 31 0.81 29
16 0.69 46 0.74 59 0.73 46 0.78 62
17 0.68 52 0.76 50 0.73 48 0.80 42
18 0.67 64 0.81 17 0.73 56 0.82 22
19 0.70 15 0.79 28 0.73 49 0.81 32
20 0.71 1 0.83 7 0.73 12 0.84 6
21 0.70 14 0.84 2 0.73 3 0.84 11
22 0.64 76 0.69 74 0.70 78 0.57 79
23 0.67 71 0.73 64 0.73 52 0.78 58
24 0.68 57 0.78 36 0.73 11 0.85 5
25 0.70 8 0.79 29 0.73 28 0.82 20
26 0.70 17 0.83 5 0.73 22 0.84 10
27 0.69 38 0.82 10 0.73 55 0.83 18
28 0.67 66 0.82 9 0.72 63 0.84 7
29 0.68 63 0.68 76 0.73 30 0.77 67
30 0.69 37 0.77 41 0.73 51 0.77 64
31 0.70 16 0.76 46 0.73 23 0.79 48
32 0.68 58 0.78 35 0.73 53 0.82 27
33 0.67 73 0.76 47 0.72 70 0.79 55
34 0.70 10 0.57 79 0.73 6 0.60 78
35 0.70 23 0.76 48 0.73 33 0.80 43
36 0.70 26 0.81 18 0.73 34 0.82 23
37 0.58 81 0.52 80 0.70 79 0.57 80
38 0.69 41 0.78 37 0.73 18 0.78 61
39 0.70 25 0.73 62 0.73 40 0.76 68
40 0.69 39 0.80 21 0.73 9 0.81 38
41 0.66 74 0.72 66 0.72 76 0.73 74
42 0.70 7 0.77 43 0.73 19 0.79 54
43 0.68 61 0.71 68 0.73 36 0.78 59
44 0.70 20 0.75 56 0.73 24 0.80 46
45 0.69 47 0.79 33 0.73 57 0.80 44
46 0.67 72 0.77 38 0.72 66 0.80 39
47 0.66 75 0.70 72 0.73 26 0.79 53
48 0.64 77 0.73 63 0.70 81 0.74 73
49 0.70 31 0.76 52 0.73 1 0.77 65
50 0.64 78 0.71 69 0.70 80 0.67 77
51 0.67 69 0.50 81 0.72 74 0.54 81
52 0.70 30 0.80 26 0.73 42 0.81 28
53 0.63 80 0.63 78 0.72 77 0.67 76
54 0.67 67 0.76 51 0.72 73 0.77 63
55 0.70 19 0.80 24 0.73 38 0.83 16
56 0.69 36 0.76 49 0.72 67 0.80 41
57 0.68 51 0.80 25 0.72 68 0.81 31
58 0.70 21 0.79 32 0.73 25 0.83 17
59 0.68 56 0.70 73 0.72 69 0.72 75
60 0.69 45 0.83 4 0.73 20 0.85 3
61 0.70 11 0.71 67 0.73 2 0.78 57
62 0.69 49 0.76 53 0.72 60 0.78 60
63 0.70 34 0.81 19 0.73 27 0.83 15
64 0.70 6 0.82 11 0.73 5 0.85 2
65 0.69 40 0.84 1 0.73 16 0.85 4
66 0.70 27 0.75 54 0.73 14 0.81 36
67 0.67 68 0.69 75 0.73 45 0.79 51
68 0.70 22 0.76 44 0.73 10 0.81 35
69 0.68 53 0.76 45 0.72 62 0.79 56
70 0.71 2 0.77 40 0.73 43 0.75 70
71 0.70 29 0.80 20 0.73 54 0.81 33
72 0.67 65 0.71 70 0.72 64 0.75 71
73 0.70 32 0.80 23 0.73 37 0.82 21
(continued on next page)

8
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 4 (continued )
Firm No. Model 1 Model 2 Model 3 Model 4

APEi Rank APEi Rank APEi Rank APEi Rank

74 0.71 3 0.74 60 0.73 41 0.81 37

75 0.64 79 0.74 58 0.72 71 0.79 49
76 0.70 18 0.64 77 0.73 44 0.74 72
77 0.68 54 0.83 8 0.72 75 0.80 40
78 0.70 35 0.79 30 0.73 47 0.81 34
79 0.68 59 0.79 31 0.72 65 0.79 50
80 0.69 44 0.83 6 0.73 7 0.85 1
81 0.71 4 0.80 27 0.73 4 0.84 12
AVG—Manufacturing 0.69 0.76 0.73 0.79

Note: AVG stands for average.

Table 5
Comparisons of the APEi’s of firms at the firm and sector levels: service (Case 1).
Firm No. Model 1 Model 2 Model 3 Model 4

APEi Rank APEi Rank APEi Rank APEi Rank

82 0.70 25 0.80 10 0.73 19 0.83 7

83 0.71 17 0.59 48 0.73 29 0.61 48
84 0.71 1 0.77 22 0.73 24 0.81 15
85 0.68 32 0.77 26 0.72 43 0.78 35
86 0.67 41 0.79 16 0.73 28 0.82 8
87 0.65 46 0.75 34 0.71 46 0.78 33
88 0.70 23 0.75 30 0.73 9 0.80 26
89 0.67 42 0.74 35 0.73 21 0.79 27
90 0.69 30 0.63 47 0.72 39 0.66 46
91 0.67 40 0.80 14 0.72 34 0.84 1
92 0.68 38 0.70 43 0.73 31 0.73 43
93 0.71 4 0.77 21 0.72 40 0.74 41
94 0.70 28 0.81 9 0.72 44 0.78 31
95 0.64 47 0.72 42 0.72 41 0.76 39
96 0.69 31 0.76 27 0.73 20 0.78 34
97 0.70 29 0.75 31 0.73 17 0.81 16
98 0.71 3 0.82 4 0.73 14 0.82 12
99 0.66 43 0.68 45 0.72 35 0.73 42
100 0.67 39 0.79 15 0.73 32 0.81 14
101 0.66 45 0.73 39 0.72 37 0.77 37
102 0.71 18 0.83 3 0.73 12 0.84 2
103 0.71 5 0.81 8 0.72 36 0.81 17
104 0.68 36 0.76 28 0.71 47 0.72 45
105 0.71 8 0.77 24 0.73 1 0.80 24
106 0.71 19 0.75 33 0.73 22 0.80 25
107 0.70 27 0.75 32 0.73 30 0.75 40
108 0.62 48 0.72 41 0.69 48 0.73 44
109 0.71 16 0.74 36 0.73 27 0.79 28
110 0.68 33 0.80 12 0.73 26 0.82 9
111 0.71 9 0.82 6 0.73 10 0.84 4
112 0.68 34 0.80 11 0.73 13 0.81 19
113 0.71 6 0.84 1 0.73 16 0.84 5
114 0.70 26 0.73 40 0.73 8 0.78 30
115 0.71 14 0.74 37 0.73 6 0.80 21
116 0.71 2 0.77 25 0.73 23 0.81 13
117 0.70 24 0.80 13 0.73 11 0.82 10
118 0.71 10 0.84 2 0.73 18 0.83 6
119 0.70 20 0.79 18 0.72 45 0.78 32
120 0.71 12 0.69 44 0.73 5 0.77 38
121 0.70 21 0.79 17 0.73 2 0.82 11
122 0.68 37 0.82 5 0.73 25 0.84 3
123 0.66 44 0.66 46 0.72 42 0.66 47
124 0.71 13 0.77 23 0.73 3 0.79 29
125 0.71 7 0.79 19 0.72 38 0.80 23
126 0.71 15 0.77 20 0.73 33 0.81 18
127 0.71 11 0.81 7 0.73 4 0.80 22
128 0.68 35 0.76 29 0.73 15 0.81 20
129 0.70 22 0.73 38 0.73 7 0.77 36
AVG—Service 0.69 0.76 0.73 0.79

Note: AVG stands for average.

separately and jointly, suggesting that it is both important and necessary 4.2.1.3. Industry level analysis. Overall, the findings from the empirical
to assess the business values of IT and EC separately and simultaneously results at the industry level are consistent with those derived from the
at the same time in order to avoid under- or over-stating the power of IT results at the firm level. For manufacturing industries (Table 6), both IT
and EC. and EC contribute positively to the APE when they appear jointly and

9
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 6
Comparisons of the APE’s at the industry level: manufacturing (Case 1).
Firm No. Model 1 Model 2 Model 3 Model 4

APEi Rank APEi Rank APEi Rank APEi Rank

30 0.69 3 0.77 2 0.73 3 0.77 2

34 0.70 1 0.57 4 0.73 2 0.60 3
18 0.67 4 0.81 1 0.73 4 0.82 1
49 0.70 2 0.76 3 0.73 1 0.77 2
AVG (311) 0.69 0.72 0.73 0.74
74 0.71 1 0.74 1 0.73 1 0.81 1
AVG (313) 0.71 0.74 0.73 0.81
52 0.70 1 0.80 1 0.73 1 0.81 1
AVG (314) 0.70 0.80 0.73 0.81
77 0.68 3 0.83 2 0.72 3 0.80 2
12 0.70 1 0.83 1 0.72 2 0.81 1
44 0.70 2 0.75 3 0.73 1 0.80 2
AVG (315) 0.69 0.80 0.72 0.80
40 0.69 2 0.80 1 0.73 1 0.81 1
68 0.70 1 0.76 2 0.73 2 0.81 1
AVG (322) 0.70 0.78 0.73 0.81
11 0.70 2 0.82 2 0.73 2 0.84 2
21 0.70 3 0.84 1 0.73 1 0.84 2
70 0.71 1 0.77 3 0.73 3 0.75 1
AVG (323) 0.70 0.81 0.73 0.81
47 0.66 1 0.70 1 0.73 1 0.79 1
AVG (324) 0.66 0.70 0.73 0.79
7 0.68 11 0.73 10 0.73 8 0.80 3
8 0.68 8 0.77 7 0.73 10 0.79 4
29 0.68 13 0.68 16 0.73 9 0.77 6
2 0.69 7 0.75 9 0.73 5 0.82 2
76 0.70 2 0.64 17 0.73 12 0.74 7
23 0.67 15 0.73 11 0.73 14 0.78 5
22 0.64 16 0.69 14 0.70 16 0.57 9
67 0.67 14 0.69 15 0.73 13 0.79 4
80 0.69 6 0.83 2 0.73 2 0.85 1
43 0.68 12 0.71 12 0.73 11 0.78 5
38 0.69 4 0.78 6 0.73 7 0.78 5
50 0.64 17 0.71 13 0.70 17 0.67 8
65 0.69 3 0.84 1 0.73 6 0.85 1
32 0.68 10 0.78 4 0.73 15 0.82 2
64 0.70 1 0.82 3 0.73 1 0.85 1
14 0.69 5 0.75 8 0.73 3 0.82 2
24 0.68 9 0.78 5 0.73 4 0.85 5
AVG (325) 0.68 0.74 0.73 0.78
58 0.70 1 0.79 1 0.73 1 0.83 1
AVG (327) 0.70 0.79 0.73 0.83
75 0.64 2 0.74 2 0.72 2 0.79 1
79 0.68 1 0.79 1 0.72 1 0.79 1
AVG (331) 0.66 0.76 0.72 0.79
66 0.70 1 0.75 2 0.73 1 0.81 2
60 0.69 2 0.83 1 0.73 2 0.85 1
AVG (332) 0.69 0.79 0.73 0.83
17 0.68 4 0.76 4 0.73 4 0.80 3
54 0.67 6 0.76 5 0.72 7 0.77 5
6 0.67 7 0.73 6 0.72 6 0.76 6
61 0.70 1 0.71 7 0.73 1 0.78 4
15 0.69 3 0.78 3 0.73 3 0.81 2
28 0.67 5 0.82 1 0.72 5 0.84 1
9 0.70 2 0.80 2 0.73 2 0.84 1
AVG (333) 0.68 0.77 0.73 0.80
35 0.70 6 0.76 15 0.73 5 0.80 5
19 0.70 4 0.79 10 0.73 8 0.81 4
71 0.70 7 0.80 6 0.73 9 0.81 4
25 0.70 3 0.79 11 0.73 4 0.82 3
78 0.70 9 0.79 12 0.73 7 0.81 4
81 0.71 1 0.80 9 0.73 1 0.84 1
13 0.71 2 0.81 4 0.73 2 0.83 2
3 0.69 12 0.82 3 0.73 11 0.82 3
56 0.69 10 0.76 16 0.72 16 0.80 5
33 0.67 19 0.76 14 0.72 18 0.79 6
41 0.66 20 0.72 19 0.72 20 0.73 10
53 0.63 22 0.63 21 0.72 21 0.67 11
27 0.69 11 0.82 1 0.73 10 0.83 2
55 0.70 5 0.80 7 0.73 6 0.83 2
46 0.67 18 0.77 13 0.72 15 0.80 5
62 0.69 13 0.76 17 0.72 13 0.78 7
72 0.67 16 0.71 20 0.72 14 0.75 8
(continued on next page)

10
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 6 (continued )
Firm No. Model 1 Model 2 Model 3 Model 4

APEi Rank APEi Rank APEi Rank APEi Rank

48 0.64 21 0.73 18 0.70 23 0.74 9

51 0.67 17 0.50 23 0.72 19 0.54 13
63 0.70 8 0.81 5 0.73 3 0.83 2
57 0.68 14 0.80 8 0.72 17 0.81 4
5 0.68 15 0.82 2 0.73 12 0.83 2
37 0.57 23 0.52 22 0.70 22 0.57 12
AVG (334) 0.68 0.75 0.72 0.78
26 0.70 1 0.83 1 0.73 1 0.84 1
69 0.68 3 0.76 3 0.72 3 0.79 2
1 0.70 2 0.82 2 0.73 2 0.84 1
AVG (335) 0.69 0.80 0.73 0.82
31 0.70 3 0.76 6 0.73 2 0.79 4
59 0.68 10 0.70 10 0.72 10 0.72 8
20 0.71 1 0.83 1 0.73 1 0.84 1
4 0.70 2 0.70 9 0.73 8 0.77 6
39 0.70 4 0.73 8 0.73 6 0.76 7
16 0.69 8 0.74 7 0.73 7 0.78 5
73 0.70 6 0.80 3 0.73 4 0.82 2
10 0.70 7 0.77 5 0.73 5 0.80 3
36 0.70 5 0.81 2 0.73 3 0.82 2
45 0.69 9 0.79 4 0.73 9 0.80 3
AVG (336) 0.70 0.76 0.73 0.79
42 0.70 1 0.77 1 0.73 1 0.79 1
AVG (337) 0.70 0.77 0.73 0.79

Note: AVG stands for average.

separately. For service industries, however, some empirical results are industry 483, the APE in the joint presence of IT and EC (Model 4: 0.78)
different from their counterparts of manufacturing industries. Consider is less than the APE as IT is invested independently of EC (Model 2:
industry 518 (Table 7) as an example. Its APE in the absence of both IT 0.79). Finally, regarding industry 622, we can observe that the joint
and EC (Model 1: 0.71) is greater than its counterparts in the presence of investments in IT and EC yield the APE 0.80 for Model 4 that is smaller
IT (standing alone, Model 2: 0.59) and in the joint presence of IT and EC than the APE of 0.81 for Model 2, as IT investment was made inde­
(Model 4: 0.61); and the contribution of EC (appearing alone, Model 3: pendently of EC investments.
0.73) is greater than its counterpart of IT (standing alone) and is the Furthermore, the results also identify the industries that benefit most
greatest among the four models. Thus, this service industry is charac­ from the IT and/or EC investments, including the non-store retailers
terized by (i) the IT PP, (ii) IT and EC are complementary, and (iii) EC is industry (454) that benefits most from the IT investment (Table 7, Model
more powerful to increase the industry’s APE than IT. 2: 0.84); the paper manufacturing industry (322) from the joint IT and
The contribution of IT to APE is greater than that of EC except for two EC investments (Table 6, Model 4: 0.81); and the building material and
manufacturing industries (311, food manufacturing, and 324, petroleum garden equipment and supplies dealers industry (444) again from the
and coal products manufacturing); and two service industries (517, joint presence of IT and EC (Table 7, Model 4: 0.84).
telecommunications; and 518, data processing, hosting, and related To sum up, the estimation results from Case 1 suggest that IT or EC
services). Consider the food manufacturing industry (311). The contri­ alone or joint IT and EC does or does not improve the firm’s and the
bution of IT (Model 2: 0.72) to APE is less than that of EC (Model 3: industry’s APEs, consistent with or in contradiction to previous studies
0.73). Similarly, in the petroleum and coal products manufacturing in­ of Peacock and Tanniru (2005), Lin and Shao (2006), Melville et al.
dustry (324), the contribution of IT (Model 2: 0.70) to APE is also less (2007), etc. Evidently, according to the results reported in Tables 4 and
than its counterpart of EC (Model 3: 0.73). For the telecommunication 5, we may observe that the impacts of IT and EC upon the APE differ,
industry (517), the contributions of IT and EC to APEs are 0.71 (Model 2) depending on whether they are invested separately or jointly. For some
and 0.73 (Model 3), respectively; and for the data processing, hosting firms, the impact of IT is greater than that of EC (e.g., firm 3 and firm 28)
and related services industry (518), the contribution of EC to APE while, for others, the influence of EC is greater than that of IT (e.g., firm
(Model 3: 0.73) is again greater than that of IT (Model 2: 0.59). Thus, in 29 and firm 37); and, for many firms, joint IT and EC investments are the
these four industries, EC is more influential than IT as far as their APEs best strategy, as measured by APEs (for instance, firms 60, 64, 80, 91,
are concerned. Also, among 16 manufacturing and 24 service industries, 102, 122, and so on). The same conclusions hold true at the industry
industry 518 is the only one that encounters the PP disease. level. These conclusions confirm our argument that the business values
Generally speaking, the COM phenomenon exists in all 16 of IT and EC could be over- or under-stated if they are valued separately
manufacturing industries and 20 out of 24 service industries, as judged rather than jointly and vice versa.
based on the estimates of the APEs in their joint presence (Model 4) that
are greater than their counterparts when IT and EC are invested sepa­ 4.2.1.4. Sector level analysis. At the sector level, both IT and EC show
rately (Model 2 and Model 3). However, there are exceptions to this positive effects on APE when they are assessed either separately or
COM phenomenon, meaning that the SUB phenomenon occurs in the jointly in both the manufacturing and service sectors (Tables 4 and 5).
clothing and clothing accessories stores industry (448), non-store re­ Consistent with the estimates at the firm level and the industry level, the
tailers industry (454), water transportation industry (483), and hospitals estimates at the sector level show that both IT and EC indeed increase
industry (622) (Table 7). In service industry 448, the APE (Model 4: the APE, no matter whether they are invested independently or jointly.
0.78) achieved when IT and EC investments were made jointly is smaller For the manufacturing sector, the APE is 0.69 (Model 1, the smallest)
than the APE (Model 2: 0.79) achieved as IT investment was made alone. without deploying IT and EC investments. In contrast, when IT and EC
For industry 454, the APE of the joint presence of IT and EC (Model 4: investments are assessed separately, the APEs are 0.76 (Model 2, the
0.83) is also smaller than 0.84 of Model 2 when IT appears alone. For second largest) and 0.73 (Model 3, the third one), respectively; while

11
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 7
Comparisons of the APE’s at the industry level: service (Case 1).
Firm No. Model 1 Model 2 Model 3 Model 4

APEi Rank APEi Rank APEi Rank APEi Rank

105 0.71 1 0.77 3 0.73 1 0.80 3

117 0.70 5 0.80 1 0.73 4 0.82 1
121 0.70 3 0.79 2 0.73 2 0.82 2
106 0.71 2 0.75 6 0.73 5 0.80 4
90 0.69 6 0.63 8 0.72 7 0.66 8
92 0.68 8 0.70 7 0.73 6 0.73 7
85 0.68 7 0.77 4 0.72 8 0.78 6
88 0.70 4 0.75 5 0.73 3 0.80 5
AVG (423) 0.70 0.74 0.73 0.78
91 0.67 1 0.80 1 0.72 1 0.84 1
87 0.65 2 0.75 2 0.71 2 0.78 2
AVG (424) 0.66 0.77 0.72 0.81
84 0.71 1 0.77 1 0.73 1 0.81 1
AVG (441) 0.71 0.77 0.73 0.81
111 0.71 1 0.82 1 0.73 1 0.84 1
AVG (444) 0.71 0.82 0.73 0.84
93 0.71 1 0.77 3 0.72 3 0.74 3
103 0.71 2 0.81 1 0.72 1 0.81 2
125 0.71 3 0.79 2 0.72 2 0.80 1
AVG (448) 0.71 0.79 0.72 0.78
124 0.71 1 0.77 1 0.73 1 0.79 1
AVG (453) 0.71 0.77 0.73 0.79
118 0.71 1 0.84 1 0.73 1 0.83 1
AVG (454) 0.71 0.84 0.73 0.83
86 0.67 3 0.79 2 0.73 3 0.82 2
96 0.69 1 0.76 3 0.73 1 0.78 3
99 0.66 4 0.68 5 0.72 4 0.73 4
108 0.62 5 0.72 4 0.69 5 0.73 5
122 0.68 2 0.82 1 0.73 2 0.84 1
AVG (481) 0.66 0.75 0.72 0.78
97 0.70 1 0.75 1 0.73 1 0.81 1
AVG (482) 0.70 0.75 0.73 0.81
119 0.70 1 0.79 1 0.72 1 0.78 1
AVG (483) 0.70 0.79 0.72 0.78
110 0.68 1 0.80 1 0.73 1 0.82 1
AVG (484) 0.68 0.80 0.73 0.82
101 0.66 1 0.73 1 0.72 1 0.77 1
AVG (492) 0.66 0.73 0.72 0.77
102 0.71 4 0.83 2 0.73 2 0.84 1
116 0.71 1 0.77 4 0.73 4 0.81 3
113 0.71 2 0.84 1 0.73 3 0.84 2
94 0.70 5 0.81 3 0.72 5 0.78 5
115 0.71 3 0.74 5 0.73 1 0.80 4
AVG (511) 0.71 0.80 0.73 0.82
95 0.64 1 0.72 1 0.72 1 0.76 1
AVG (515) 0.64 0.72 0.72 0.76
123 0.66 2 0.66 2 0.72 2 0.66 2
128 0.68 1 0.76 1 0.73 1 0.81 1
AVG (517) 0.67 0.71 0.73 0.73
83 0.71 1 0.59 1 0.73 1 0.61 1
AVG (518) 0.71 0.59 0.73 0.61
120 0.71 1 0.69 2 0.73 1 0.77 1
104 0.68 2 0.76 1 0.71 2 0.72 2
AVG (532) 0.69 0.72 0.72 0.75
100 0.67 2 0.79 1 0.73 1 0.81 1
126 0.71 1 0.77 2 0.73 2 0.81 2
AVG (541) 0.69 0.78 0.73 0.81
82 0.70 1 0.80 1 0.73 1 0.83 1
AVG (561) 0.70 0.80 0.73 0.83
129 0.70 1 0.73 1 0.73 1 0.77 1
AVG (562) 0.70 0.73 0.73 0.77
127 0.71 1 0.81 1 0.73 1 0.80 1
AVG (622) 0.71 0.81 0.73 0.80
109 0.71 1 0.74 1 0.73 1 0.79 1
AVG (623) 0.71 0.74 0.73 0.79
107 0.70 2 0.75 2 0.73 3 0.75 3
112 0.68 3 0.80 1 0.73 2 0.81 1
114 0.70 1 0.73 3 0.73 1 0.78 2
AVG (721) 0.69 0.76 0.73 0.78
89 0.67 2 0.74 2 0.73 2 0.79 2
98 0.71 1 0.82 1 0.73 1 0.82 1
AVG (722) 0.69 0.78 0.73 0.81

Note: AVG stands for average.

12
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 8
Comparisons of the APEi’s of firms at the firm and sector levels: manufacturing (Case 2).
Firm No. Model 1 Model 5 Model 6

APEi Rank APEi Rank APEi Rank

1 0.70 28 0.82 21 0.84 18

2 0.69 48 0.76 54 0.83 26
3 0.69 42 0.82 20 0.83 25
4 0.70 12 0.75 58 0.81 43
5 0.68 62 0.83 15 0.85 16
6 0.67 70 0.71 68 0.77 65
7 0.68 60 0.75 57 0.81 47
8 0.68 55 0.80 33 0.82 39
9 0.70 24 0.81 25 0.85 11
10 0.70 33 0.78 40 0.80 50
11 0.70 13 0.79 37 0.86 8
12 0.70 9 0.87 2 0.83 27
13 0.71 5 0.87 1 0.87 3
14 0.69 43 0.76 53 0.83 30
15 0.69 50 0.79 36 0.82 34
16 0.69 46 0.74 60 0.78 61
17 0.68 52 0.77 46 0.81 44
18 0.67 64 0.84 10 0.84 22
19 0.70 15 0.81 26 0.81 42
20 0.71 1 0.84 11 0.86 9
21 0.70 14 0.87 3 0.86 7
22 0.64 76 0.69 75 0.53 81
23 0.67 71 0.73 61 0.79 59
24 0.68 57 0.77 48 0.85 14
25 0.70 8 0.82 22 0.84 20
26 0.70 17 0.85 9 0.85 10
27 0.69 38 0.85 7 0.85 12
28 0.67 66 0.85 8 0.86 6
29 0.68 63 0.68 76 0.77 66
30 0.69 37 0.78 42 0.77 63
31 0.70 16 0.77 50 0.80 53
32 0.68 58 0.78 39 0.82 35
33 0.67 73 0.76 51 0.80 52
34 0.70 10 0.60 79 0.63 78
35 0.70 23 0.78 43 0.81 41
36 0.70 26 0.82 19 0.83 32
37 0.58 81 0.50 80 0.56 79
38 0.69 41 0.82 23 0.81 46
39 0.70 25 0.72 63 0.76 70
40 0.69 39 0.83 12 0.83 28
41 0.66 74 0.72 66 0.73 74
42 0.70 7 0.77 47 0.79 55
43 0.68 61 0.69 72 0.76 69
44 0.70 20 0.73 62 0.79 54
45 0.69 47 0.80 34 0.80 49
46 0.67 72 0.81 28 0.84 21
47 0.66 75 0.71 69 0.80 48
48 0.64 77 0.72 64 0.73 73
49 0.70 31 0.77 49 0.78 62
50 0.64 78 0.69 74 0.64 77
51 0.67 69 0.48 81 0.53 80
52 0.70 30 0.82 17 0.84 23
53 0.63 80 0.62 78 0.66 76
54 0.67 67 0.77 45 0.78 60
55 0.70 19 0.82 18 0.85 15
56 0.69 36 0.75 55 0.82 36
57 0.68 51 0.81 30 0.82 37
58 0.70 21 0.81 27 0.85 13
59 0.68 56 0.69 73 0.71 75
60 0.69 45 0.86 4 0.86 5
61 0.70 11 0.72 65 0.79 58
62 0.69 49 0.74 59 0.76 68
63 0.70 34 0.82 24 0.84 19
64 0.70 6 0.83 13 0.87 2
65 0.69 40 0.85 6 0.87 4
66 0.70 27 0.76 52 0.81 40
67 0.67 68 0.70 71 0.81 45
68 0.70 22 0.79 38 0.83 31
69 0.68 53 0.78 41 0.80 51
70 0.71 2 0.80 31 0.77 64
71 0.70 29 0.83 16 0.83 29
72 0.67 65 0.71 67 0.75 71
73 0.70 32 0.81 29 0.83 24
(continued on next page)

13
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 8 (continued )
Firm No. Model 1 Model 5 Model 6

APEi Rank APEi Rank APEi Rank

74 0.71 3 0.78 44 0.82 38

75 0.64 79 0.71 70 0.77 67
76 0.70 18 0.64 77 0.73 72
77 0.68 54 0.86 5 0.83 33
78 0.70 35 0.75 56 0.79 57
79 0.68 59 0.79 35 0.79 56
80 0.69 44 0.83 14 0.87 1
81 0.71 4 0.80 32 0.84 17
AVG—Manufacturing 0.69 0.77 0.80

Note: AVG stands for average.

when IT and EC investments are evaluated jointly, the APE is 0.79

(Model 4, the largest). The empirical evidence suggests that when IT and
Table 9 EC are invested separately, the APEs increase by 10% (from 0.69 to 0.76)
Comparisons of the APEi’s of firms at the firm and sector levels: service (Case 2). and 5.8% (from 0.69 to 0.73), respectively; however, when IT and EC are
Firm No. Model 1 Model 5 Model 6 assessed jointly, the APE increases by 14.5% (from 0.69 to 0.79). The
results at the manufacturing sector implying that the sector is the PP free
APEi Rank APEi Rank APEi Rank
and the contribution of IT to APE is greater than that of EC, suggesting
82 0.70 25 0.77 24 0.79 30 that IT is more important than EC. Additionally, the COM phenomenon
83 0.71 17 0.52 48 0.54 48
generated by the joint presence of IT and EC does exist in this sector
84 0.71 1 0.72 41 0.80 24
85 0.68 32 0.80 15 0.80 26 since the APE of their joint presence (0.79 from Model 4) is larger than
86 0.67 41 0.80 18 0.83 12 their counterparts when IT and EC are assessed separately (0.76 from
87 0.65 46 0.77 26 0.80 27 Model 2 and 0.73 from Model 3).
88 0.70 23 0.74 35 0.78 32 For the service sector, the APE is 0.69 (Model 1) in the absence of IT
89 0.67 42 0.72 40 0.75 38
90 0.69 30 0.57 47 0.61 47
and EC investments; but, as IT and EC appear separately, the APEs are
91 0.67 40 0.79 19 0.85 7 0.76 (Model 2) and 0.73 (Model 3), respectively; and when IT and EC are
92 0.68 38 0.70 44 0.73 42 invested jointly, the APE is 0.79 (Model 4, the largest). In other words, if
93 0.71 4 0.73 36 0.71 45 IT and EC are invested separately, the APEs would increase by 10%
94 0.70 28 0.83 7 0.80 25
(from 0.69 to 0.76) and 6% (from 0.69 to 0.73), respectively; and when
95 0.64 47 0.73 38 0.76 34
96 0.69 31 0.75 31 0.78 33 IT and EC investments are assessed jointly, the APE increases by 15%
97 0.70 29 0.79 21 0.82 15 (from 0.69 to 0.79). The same conclusions as the manufacturing sector
98 0.71 3 0.85 4 0.81 20 can readily be drawn for the service sector. IT is capable of making a
99 0.66 43 0.68 46 0.73 41 greater contribution to the service sector’s APE than EC, again, implying
100 0.67 39 0.81 12 0.84 9
101 0.66 45 0.71 42 0.75 37
the absence of the PP from the service sector; and the COM phenomenon
102 0.71 18 0.86 3 0.86 3 also exists in the service sector.
103 0.71 5 0.82 10 0.80 23
104 0.68 36 0.72 39 0.68 46 4.2.2. Case 2: three factors (traditional capital, traditional labor, and IT
105 0.71 8 0.76 28 0.79 29
capital)
106 0.71 19 0.74 34 0.79 31
107 0.70 27 0.80 14 0.72 44
108 0.62 48 0.73 37 0.73 43 4.2.2.1. Model estimates. Differing from Case 1 where we have treated
109 0.71 16 0.70 45 0.76 36 IT as a firm characteristic (i.e., a determinant of inefficiency), Case 2
110 0.68 33 0.84 5 0.85 4
provides settings where IT capital serves as an input alongside tradi­
111 0.71 9 0.83 8 0.85 5
112 0.68 34 0.76 27 0.74 40 tional capital (non-IT capital) and traditional labor in the CES produc­
113 0.71 6 0.87 1 0.86 2 tion function. Table 3 reports the estimation results of the two CES-based
114 0.70 26 0.77 25 0.83 11 SPF models under the Case 2 setting.
115 0.71 14 0.76 30 0.82 14 First, according to Table 3, the estimation coefficient (− 0.14) of EC
116 0.71 2 0.80 13 0.84 8
117 0.70 24 0.80 16 0.81 17
(Model 6, or equation (13)) is statistically significant, with a negative
118 0.71 10 0.86 2 0.87 1 sign, implying a positive effect of EC upon the productive efficiency
119 0.70 20 0.83 6 0.81 16 since originally it bears a negative sign in equation (13). All the co­
120 0.71 12 0.70 43 0.81 22 efficients of IT (1.01 from Model 5 and 0.89 from Model 6) are signifi­
121 0.70 21 0.78 22 0.81 18
cant and positive, implying that IT also exerts a positive effect on the
122 0.68 37 0.83 9 0.85 6
123 0.66 44 0.76 29 0.75 39 output and, hence, the productive efficiency. Recall that in Case 1,
124 0.71 13 0.78 23 0.76 35 traditional capital (Kit ) and traditional labor (Lit ) are substitutable in
125 0.71 7 0.80 17 0.82 13 Models 1 to 4 (Table 2). Now, in Case 2, the appearance of IT capital (Iit )
126 0.71 15 0.74 33 0.81 21 as an input along with Kit and Lit changes the status of the relationship
127 0.71 11 0.82 11 0.79 28
128 0.68 35 0.75 32 0.81 19
between Kit and Lit from substitution to complement (Model 5) because
129 0.70 22 0.79 20 0.83 10 ρKL = − 0.69 is negative and greater than − 1 (see Table A). However, the
AVG—Service 0.69 0.77 0.79 presence of EC in Model 6 leads to ρKL = − 0.10, implying that the
Note: AVG stands for average. complement relationship between Kit and Lit is weakened. Besides, the
substitution relationship exists not only between Lit and Iit but also be­
tween Kit and Iit in Model 5 and Model 6 as well.
Second, similar to the results in Case 1, both IT and EC increase the

14
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 10
Comparisons of the APE’s at the industry level: manufacturing (Case 2).
Firm No. Model 1 Model 5 Model 6

APEi Rank APEi Rank APEi Rank

30 0.69 3 0.78 2 0.77 3

34 0.70 1 0.60 4 0.63 4
18 0.67 4 0.84 1 0.84 1
49 0.70 2 0.77 3 0.78 2
AVG (311) 0.69 0.76 0.75
74 0.71 1 0.78 1 0.82 1
AVG (313) 0.71 0.78 0.82
52 0.70 1 0.82 1 0.84 1
AVG (314) 0.70 0.82 0.84
77 0.68 3 0.86 2 0.83 2
12 0.70 1 0.87 1 0.83 1
44 0.70 2 0.73 3 0.79 3
AVG (315) 0.69 0.83 0.82
40 0.69 2 0.83 1 0.83 1
68 0.70 1 0.79 2 0.83 2
AVG (322) 0.70 0.81 0.83
11 0.70 2 0.79 3 0.86 2
21 0.70 3 0.87 1 0.86 1
70 0.71 1 0.80 2 0.77 3
AVG (323) 0.70 0.82 0.83
47 0.66 1 0.71 1 0.80 1
AVG (324) 0.66 0.71 0.80
8 0.68 8 0.80 5 0.82 8
7 0.68 11 0.75 10 0.81 11
29 0.68 13 0.68 16 0.77 13
2 0.69 7 0.76 9 0.83 5
76 0.70 2 0.64 17 0.73 15
23 0.67 15 0.73 11 0.79 12
22 0.64 16 0.69 15 0.53 17
67 0.67 14 0.70 12 0.81 9
80 0.69 6 0.83 3 0.87 1
38 0.69 4 0.82 4 0.81 10
43 0.68 12 0.69 13 0.76 14
50 0.64 17 0.69 14 0.64 16
65 0.69 3 0.85 1 0.87 3
32 0.68 10 0.78 6 0.82 7
64 0.70 1 0.83 2 0.87 2
14 0.69 5 0.76 8 0.83 6
24 0.68 9 0.77 7 0.85 4
AVG (325) 0.68 0.75 0.79
58 0.70 1 0.81 1 0.85 1
AVG (327) 0.70 0.81 0.85
75 0.64 2 0.71 2 0.77 2
79 0.68 1 0.79 1 0.79 1
AVG (331) 0.66 0.75 0.78
66 0.70 1 0.76 2 0.81 2
60 0.69 2 0.86 1 0.86 1
AVG (332) 0.69 0.81 0.84
17 0.68 4 0.77 5 0.81 4
54 0.67 6 0.77 4 0.78 6
6 0.67 7 0.71 7 0.77 7
61 0.70 1 0.72 6 0.79 5
15 0.69 3 0.79 3 0.82 3
28 0.67 5 0.85 1 0.86 1
9 0.70 2 0.81 2 0.85 2
AVG (333) 0.68 0.78 0.81
35 0.70 6 0.78 13 0.81 13
71 0.70 7 0.83 4 0.83 10
19 0.70 4 0.81 9 0.81 14
25 0.70 3 0.82 7 0.84 7
78 0.70 9 0.75 16 0.79 16
13 0.71 2 0.87 1 0.87 1
81 0.71 1 0.80 12 0.84 5
3 0.69 12 0.82 6 0.83 9
56 0.69 10 0.75 15 0.82 11
33 0.67 19 0.76 14 0.80 15
41 0.66 20 0.72 19 0.73 20
53 0.63 22 0.62 21 0.66 21
27 0.69 11 0.85 2 0.85 2
55 0.70 5 0.82 5 0.85 3
46 0.67 18 0.81 10 0.84 8
62 0.69 13 0.74 17 0.76 17
72 0.67 16 0.71 20 0.75 18
(continued on next page)

15
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 10 (continued )
Firm No. Model 1 Model 5 Model 6

APEi Rank APEi Rank APEi Rank

48 0.64 21 0.72 18 0.73 19

51 0.67 17 0.48 23 0.53 23
63 0.70 8 0.82 8 0.84 6
57 0.68 14 0.81 11 0.82 12
5 0.68 15 0.83 3 0.85 4
37 0.57 23 0.50 22 0.56 22
AVG (334) 0.68 0.76 0.78
26 0.70 1 0.85 1 0.85 1
69 0.68 3 0.78 3 0.80 3
1 0.70 2 0.82 2 0.84 2
AVG (335) 0.69 0.81 0.83
31 0.70 3 0.77 6 0.80 7
59 0.68 10 0.69 10 0.71 10
20 0.71 1 0.84 1 0.86 1
4 0.70 2 0.75 7 0.81 4
39 0.70 4 0.72 9 0.76 9
16 0.69 8 0.74 8 0.78 8
73 0.70 6 0.81 3 0.83 2
10 0.70 7 0.78 5 0.80 6
36 0.70 5 0.82 2 0.83 3
45 0.69 9 0.80 4 0.80 5
AVG (336) 0.70 0.77 0.80
42 0.70 1 0.77 1 0.79 1
AVG (337) 0.70 0.77 0.79

Note: AVG stands for average.

overall average productive efficiency no matter whether they appear manufacturing firm 37 as a second example. The pattern of the impacts
separately or jointly in the models. In the absence of both IT as an input of IT and EC upon the firm’s performance is similar to firm 34. The firm’s
and EC as a firm characteristic (or a determinant of inefficiency), the APEs are, respectively, 0.58, 0.50, and 0.56 recorded from Model 1 (in
APE is 0.69 (Model 1); as IT is treated as an input in the absence of EC, the absence of IT and EC), Model 5 (in the presence of IT alone), and
the APE is 0.77 (Model 5); and when IT is treated as an input and EC as Model 6 (in the joint appearance of IT and EC). Thus, the APE estimates
an inefficiency correction (or an inefficiency determinant) jointly, the suggest that firms 34 and 37 suffer from the PP. Moreover, in both firms,
APE is 0.80 (Model 6). These results imply that when IT is assessed the combination of IT and EC investments in Model 6 generates a higher
alone, the APE increases by 11.6% (from 0.69 to 0.77); but, when IT and APE than IT appearing alone in Model 5, implying that there exists the
EC are assessed simultaneously, the APE increases by 15.9% (from 0.69 COM phenomenon created by the joint presence of IT and EC. Obviously,
to 0.80). As such, we may argue that overall, the IT PP is out of the we can also observe from Table 8 that the PP is absent from, but the SUB
picture; and there exists the COM phenomenon since the APE of 0.80 in phenomenon is existent in, some manufacturing firms (e.g., firms 70 and
the joint presence of IT and EC (Model 6) is greater than the APE of 0.77 77 listed in Table 8) because the IT investment alone (Model 5) leads to
when IT appears alone (Model 5). higher APEs than their counterparts of the joint presence of IT and EC
(Model 6) and because the appearance of EC along with IT reduces the
4.2.2.2. Firm level analysis. At the firm level, we can analyze the APEs APEs of these firms.
of firms based on the empirical results reported in Table 8 (for Likewise, we can observe from Table 9 that there are service firms
manufacturing firms) and 9 (for service firms). In general, the estimates that show the varying effects of IT and/or EC investments upon the
demonstrate that both IT and EC contribute to the increase of APEs at the firm’s performance as measured by productive efficiency. For example,
firm level, and the APEs are affected by IT and EC to varying extents, service firms 83 and 90, among others, are two obvious examples where
depending on whether IT and EC are assessed simultaneously (Model 6) the PP appears along with the COM phenomenon because of the joint
or separately (Model 5). The empirical evidence of individual firms investments of IT and EC, just like manufacturing firms 34 and 37. Of
greatly supports the conclusions of the model estimates as discussed course, we also can find the service firms where the PP disappears with
above. In addition, we are able to identify the firms that benefit most the existence of the SUB phenomenon (e.g., firms 93, 94, etc., listed in
from the IT and/or EC investments by carefully looking into Tables 8 and Table 9, because the APEs, 0.73 and 0.83, are larger than those, 0.71 and
9. 0.80, in the joint appearance of IT and EC.
On the one hand, among the manufacturing firms, firm 13 benefits Nevertheless, based on the APE estimates reported in Tables 8 and 9,
most from the IT adoption (Model 5: 0.87), whereas firm 80 captures the we can summarize several points of interest as stated below to answer
highest APE in the joint appearance of IT and EC (Model 6: 0.87). On the research questions (iv) and (v).
other, among the service firms, firm 113 benefits most from the IT in­ One, in total, we can identify 5 manufacturing firms (i.e., Nos. 34, 37,
vestment (Model 5: 0.87), while firm 118 wins the largest APE in the 51, 53, and 76) along with 4 service firms (Nos. 83, 90, 109, and 120)
joint presence of IT and EC (Model 6: 0.87). which have faced the PP and the COM relationship between IT and EC
Again, although both IT and EC have positive effects on the firm’s since their APEs from Model 1 are greater than their counterparts from
APE when they are evaluated either independently or simultaneously, Model 5 which in turn are smaller than those APEs generated by Model
there are exceptions. Take manufacturing firm 34 (in Table 8) as an 6.
illustrative example. In the absence of both IT as an input and EC as a Two, there are 8 manufacturing firms (12, 21, 22, 30, 38, 50, 70, and
determinant of inefficiency (i.e., as a firm characteristic), the firm’s APE 77) and 13 service firms (93, 94, 98, 103, 104, 107, 112, 113, 118, 119,
is 0.70 (Model 1); however, when IT is treated as an input alone but 123, 124, and 127) which are immune from the PP disease but experi­
without EC as a firm characteristic, its APE is 0.60 (Model 5); and the ence the SUB relationship between IT and EC. The reason is that their
firm’s APE is 0.63 (Model 6) in the joint presence of IT and EC. Consider APEs estimates from Model 1 are smaller than those produced by Model

16
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 11 5 which are greater than those obtained from Model 6.

Comparisons of the APE’s at the industry level: service (Case 2). Three, it can be observed that 10 manufacturing firms (13, 18, 19,
Firm No. Model 1 Model 5 Model 6 26, 27, 40, 45, 60, 71, and 79) and 3 service firms (85, 102, and 108) are
free from the PP but the relationship between IT and EC is either com­
APEi Rank APEi Rank APEi Rank
plementary or substitutable. This is because the APEs from Model 1 are
105 0.71 1 0.76 4 0.79 4 less than those APEs obtained from Model 5 which are equal to the APEs
117 0.70 5 0.80 2 0.81 1
121 0.70 3 0.78 3 0.81 2
produced by Model 6.
106 0.71 2 0.74 5 0.79 5 Four, there is only one manufacturing firm (that is, No.29) whose
90 0.69 6 0.57 8 0.61 8 APE (0.68) from Model 1 is equal to the APE (0.68) from Model 5 which
92 0.68 8 0.70 7 0.73 7 is less than the APE (0.77) estimated by Model 6. The results clearly
85 0.68 7 0.80 1 0.80 3
imply that, for this firm, we cannot tell whether or not it suffers the PP
88 0.70 4 0.74 6 0.78 6
AVG (423) 0.70 0.74 0.77 but, clearly, the COM relationship is with the firm.
91 0.67 1 0.79 1 0.85 1 Five and final, a significant majority of 57 manufacturing firms and
87 0.65 2 0.77 2 0.80 2 more than half (28) of the service firms studied do not suffer from the IT
AVG (424) 0.66 0.78 0.82 PP disease coupled with the COM relationship between IT and EC. It is
84 0.71 1 0.72 1 0.80 1
AVG (441) 0.71 0.72 0.80
because, for these firms, their APEs produced by Model 1 < (less than)
111 0.71 1 0.83 1 0.85 1 the APEs obtained from Model 5 < the APEs estimated based on Model 6.
AVG(444) 0.71 0.83 0.85 To summarize, we conclude that, based on the empirical results at
93 0.71 1 0.73 3 0.71 3 the firm level, except for firms 34, 37, 51, 53,76, 83, 90 109 and 120
103 0.71 2 0.82 1 0.80 2
(some of which were discussed above), both IT and EC contribute
125 0.71 3 0.80 2 0.82 1
AVG (448) 0.71 0.79 0.78 significantly and positively to the APE, and their contributions to the
124 0.71 1 0.78 1 0.76 1 APE differ when they are evaluated separately and jointly. The empirical
AVG (453) 0.71 0.78 0.76 evidence strongly suggests that from the strategic point of view, it is
118 0.71 1 0.86 1 0.87 1 important to evaluate the business values of IT and EC separately and
AVG (454) 0.71 0.86 0.87
86 0.67 3 0.80 2 0.83 2
simultaneously at the same time, as far as the firm’s performance is
96 0.69 1 0.75 3 0.78 3 concerned, in order to determine whether IT and EC investments should
99 0.66 4 0.68 5 0.73 4 be engaged jointly or separately.
108 0.62 5 0.73 4 0.73 5
122 0.68 2 0.83 1 0.85 1
AVG (481) 0.66 0.76 0.78
4.2.2.3. Industry level analysis. We can observe from Tables 10 and 11
97 0.70 1 0.79 1 0.82 1 that the business values of IT and EC differ if they are evaluated inde­
AVG (482) 0.70 0.79 0.82 pendently (Model 5) and simultaneously (Model 6) and, consequently,
119 0.70 1 0.83 1 0.81 1 their contributions to the manufacturing and service industries also
AVG (483) 0.70 0.83 0.81
differ. Generally speaking, both IT and EC contribute positively to their
110 0.68 1 0.84 1 0.85 1
AVG (484) 0.68 0.84 0.85 APEs of the manufacturing and service industries when they are assessed
101 0.66 1 0.71 1 0.75 1 jointly and separately, with only two exception associated with industry
AVG (492) 0.66 0.71 0.75 311-food manufacturing and industry 315-apparel manufacturing
102 0.71 4 0.86 2 0.86 2 (Table 10) and eight service industries associated with industries 448-
116 0.71 1 0.80 4 0.84 3
113 0.71 2 0.87 1 0.86 1
clothing and clothing accessories stores, 453-miscelleneous store re­
94 0.70 5 0.83 3 0.80 5 tailers, 483-water transportation, 518-data processing, hosting, and
115 0.71 3 0.76 5 0.82 4 related services, 622-hosptials, 623-nursing and residential care facil­
AVG (511) 0.71 0.82 0.84 ities, 721-accommodation, and 722-food services and drinking places
95 0.64 1 0.73 1 0.76 1
(Table 11). For 311 and 315, IT and EC together have yielded the APEs
AVG (515) 0.64 0.73 0.76
123 0.66 2 0.76 1 0.75 2 (0.75 and 0.82 from Model 6) which are smaller than the APEs (0.76 and
128 0.68 1 0.75 2 0.81 1 0.83 from Model 5) when IT stands alone or when EC is absent. For 448,
AVG (517) 0.67 0.75 0.78 453, 483, 622, 721, and 722, their APEs from Model 1 are less than those
83 0.71 1 0.52 1 0.54 1 from Model 5 which are greater than those from Model 6, whereas for
AVG (518) 0.71 0.52 0.54
518 and 623, their APEs in the absence of both IT as an input and EC as a
120 0.71 1 0.70 2 0.81 1
104 0.68 2 0.72 1 0.68 2 determinant of inefficiency (i.e., from Model 1) are greater than those
AVG (532) 0.69 0.71 0.74 APEs in the presence of IT (Model 5) which are smaller than the APEs in
100 0.67 2 0.81 1 0.84 1 the joint presence of IT and EC (Model 6). Therefore, the empirical ev­
126 0.71 1 0.74 2 0.81 2
idence suggests that the IT PP and the COM phenomenon exist in in­
AVG (541) 0.69 0.77 0.82
82 0.70 1 0.77 1 0.79 1 dustries 518 and 623. In contrast, the IT PP is absent from, and the SUB
AVG (561) 0.70 0.77 0.79 phenomenon is present in, eight industries, namely, 311, 315, 448, 453,
129 0.70 1 0.79 1 0.83 1 483, 622, 721, and 722.
AVG (562) 0.70 0.79 0.83 More specifically, the COM relationship between IT and EC exists in
127 0.71 1 0.82 1 0.79 1
an overwhelming majority of industries since the APEs of Model 6 in the
AVG (622) 0.71 0.82 0.79
109 0.71 1 0.70 1 0.76 1 joint presence of both IT and EC investments are greater than their
AVG (623) 0.71 0.70 0.76 counterparts of Model 5 with the sole presence of IT. However, the
112 0.68 3 0.76 3 0.74 1 empirical estimates indicate that, when IT is treated as an input along
107 0.70 2 0.80 1 0.72 2
with Kit and Lit , and EC does not exist (Model 5), the APEs of the above-
114 0.70 1 0.77 2 0.83 1
AVG (721) 0.69 0.78 0.76
stated eight industries are 0.76, 0.83, 0.79, 0.78, 0.83, 0.82, 0.78, and
98 0.71 1 0.85 1 0.81 1 0.79, respectively; but, when IT is treated as an input along with EC
89 0.67 2 0.72 2 0.75 2 (Model 6), the APEs of these eight industries decline to 0.75, 0.82, 0.78,
AVG (722) 0.69 0.79 0.78 0.76, 0.81, 0.79, 0.76, and 0.78. Hence, the SUB phenomenon of the
Note: AVG stands for average. joint IT and EC takes place because the APEs from Model 6 in the joint

17
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 12
Performance comparisons of relevant models with and without inputs substitution and complement: S or W and % (percentage changes).
Firm No. (a) Model 5 vs Model 1 (b) Model 5 vs Model 4 (c) Model 6 vs Model 4

S/W % Rank S/W % Rank S/W % Rank

1 S 17.35 38 S 0.08 91 S 0.92 68

2 S 10.01 82 S 1.49 54 S 2.03 37
3 S 18.72 24 S 0.55 82 S 1.63 46
4 S 6.14 106 S 5.79 6 S 5.49 6
5 S 22.54 11 S 1.55 52 S 1.31 59
6 S 6.58 103 W − 2.00 108 S 1.12 62
7 S 10.28 76 S 2.07 41 S 1.32 57
8 S 17.80 34 S 3.66 23 S 3.06 17
9 S 16.04 45 S 0.98 68 S 1.50 53
10 S 12.45 69 S 1.02 65 S 0.53 77
11 S 12.30 71 W − 3.44 117 S 1.93 40
12 S 24.18 5 S 4.59 13 S 2.49 29
13 S 23.81 6 S 7.47 3 S 5.51 5
14 S 10.24 77 S 1.01 66 S 0.98 66
15 S 15.39 52 S 1.54 53 S 1.38 56
16 S 7.95 93 S 0.30 88 W − 0.04 91
17 S 13.25 63 S 1.74 48 S 1.31 58
18 S 24.82 3 S 3.86 20 S 1.98 38
19 S 15.67 49 S 1.90 46 S 0.24 84
20 S 18.19 29 S 1.47 55 S 1.48 54
21 S 23.54 8 S 3.17 26 S 2.33 34
22 S 6.75 101 S 0.16 89 W − 7.45 126
23 S 9.98 83 S 1.00 67 S 0.73 71
24 S 13.82 58 W − 0.75 99 S 0.42 80
25 S 16.24 43 S 2.85 32 S 2.30 36
26 S 21.03 17 S 2.33 38 S 1.54 48
27 S 23.05 10 S 4.12 16 S 2.96 19
28 S 26.60 1 S 3.12 29 S 2.39 31
29 S 1.28 119 S 0.11 90 S 0.51 78
30 S 11.75 73 S 0.53 84 S 0.09 88
31 S 9.30 89 S 0.39 85 S 0.06 89
32 S 15.90 46 S 0.94 69 S 1.10 64
33 S 14.75 55 S 0.37 87 S 1.11 63
34 W − 14.75 126 S 5.27 10 S 5.72 3
35 S 11.11 74 S 2.17 40 S 1.52 51
36 S 17.98 32 S 1.69 51 S 0.87 69
37 W − 13.51 125 W − 3.68 118 W − 2.02 112
38 S 18.28 28 S 5.32 9 S 3.81 11
39 S 3.74 112 W − 0.59 98 W − 1.08 102
40 S 20.38 19 S 3.78 21 S 3.22 15
41 S 7.60 97 W − 0.20 92 W − 0.68 98
42 S 9.59 88 S 0.87 74 S 0.65 75
43 S 2.67 116 W − 2.57 112 W − 2.23 115
44 S 4.07 111 W − 3.15 115 W − 0.34 94
45 S 15.60 51 S 1.33 58 S 0.54 76
46 S 21.20 16 S 4.35 14 S 4.02 10
47 S 6.67 102 S 1.10 63 S 1.74 44
48 S 12.31 70 W − 0.75 100 W − 0.60 97
49 S 10.15 80 S 1.17 61 S 0.80 70
50 S 7.44 98 W − 3.03 113 W − 3.56 118
51 W − 27.84 129 W − 3.06 114 W − 1.43 110
52 S 18.37 27 S 3.37 24 S 2.73 23
53 W − 1.95 123 W − 2.14 109 W − 1.37 109
54 S 15.30 53 S 1.96 45 S 1.23 60
55 S 17.79 35 S 3.14 28 S 1.53 49
56 S 8.39 92 W − 0.86 102 S 2.89 22
57 S 18.03 31 S 0.84 76 S 1.53 50
58 S 15.74 48 S 2.85 31 S 2.31 35
59 S 2.22 118 W − 0.32 94 W − 0.76 100
60 S 24.67 4 S 3.25 25 S 1.73 45
61 S 2.74 115 S 0.94 70 S 0.70 73
62 S 8.47 91 W − 1.47 106 W − 1.86 111
63 S 17.31 39 S 1.30 59 S 1.09 65
64 S 18.15 30 S 1.71 50 S 2.35 33
65 S 23.64 7 S 1.83 47 S 2.58 27
66 S 9.29 90 S 0.93 72 S 0.71 72
67 S 4.17 110 S 1.72 49 S 1.86 43
68 S 12.66 67 S 3.14 27 S 2.55 28
69 S 14.70 56 S 2.19 39 S 1.61 47
70 S 13.29 61 S 3.78 22 S 2.39 32
71 S 18.53 26 S 2.73 34 S 2.61 26
72 S 6.16 105 S 0.89 73 S 0.34 82
73 S 15.89 47 S 0.86 75 S 1.42 55
(continued on next page)

18
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Table 12 (continued )
Firm No. (a) Model 5 vs Model 1 (b) Model 5 vs Model 4 (c) Model 6 vs Model 4

S/W % Rank S/W % Rank S/W % Rank

74 S 9.70 86 S 4.81 12 S 1.91 41

75 S 10.16 78 W − 4.93 124 W − 3.66 119
76 W − 8.40 124 W − 0.49 96 W − 0.75 99
77 S 26.40 2 S 3.96 17 S 3.29 14
78 S 7.36 99 W − 5.64 125 W − 2.17 114
79 S 17.00 40 S 0.62 79 W − 0.42 95
80 S 20.49 18 S 0.59 81 S 2.41 30
81 S 13.10 64 S 0.53 83 S 0.46 79
AVG—Manufacturing S 12.31 - S 1.26 - S 1.14 -
82 S 10.47 75 W − 3.81 119 W − 4.59 121
83 W − 26.01 128 W − 10.92 129 W − 10.88 129
84 S 1.14 120 W − 6.84 127 W − 0.78 101
85 S 16.49 42 S 3.91 18 S 3.10 16
86 S 19.15 22 S 0.93 71 S 0.38 81
87 S 17.58 37 S 2.48 36 S 2.93 21
88 S 4.42 109 W − 2.43 110 W − 2.07 113
89 S 7.93 94 W − 3.36 116 W − 5.80 124
90 W − 16.23 127 W − 8.15 128 W − 8.04 128
91 S 18.70 25 W − 0.45 95 S 0.21 86
92 S 3.41 114 S 0.75 78 W − 0.10 92
93 S 3.46 113 W − 4.87 123 W − 4.75 123
94 S 18.78 23 S 2.04 43 S 2.98 18
95 S 13.27 62 S 0.80 77 S 0.11 87
96 S 10.09 81 W − 0.30 93 S 0.33 83
97 S 13.73 59 S 5.42 8 S 1.17 61
98 S 19.92 20 S 3.87 19 W − 1.21 106
99 S 2.23 117 W − 0.84 101 W − 0.27 93
100 S 19.33 21 S 1.99 44 S 2.63 25
101 S 7.86 95 W − 2.52 111 W − 2.70 116
102 S 21.34 15 S 3.02 30 S 1.93 39
103 S 16.15 44 S 1.42 56 W − 0.49 96
104 S 6.42 104 W − 4.51 121 W − 6.01 125
105 S 7.73 96 W − 1.01 103 W − 1.09 104
106 S 4.81 108 W − 1.44 105 W − 1.23 107
107 S 14.80 54 S 6.75 4 W − 4.70 122
108 S 17.62 36 S 1.04 64 S 0.05 90
109 W − 1.45 122 W − 5.89 126 W − 4.54 120
110 S 23.35 9 S 5.20 11 S 3.49 13
111 S 16.82 41 S 1.29 60 S 1.52 52
112 S 11.84 72 W − 4.84 122 W − 7.99 127
113 S 22.42 12 S 2.66 35 S 2.65 24
114 S 9.82 84 S 5.87 5 S 5.53 4
115 S 7.04 100 S 2.81 33 S 1.87 42
116 S 12.93 65 S 4.28 15 S 3.70 12
117 S 13.31 60 W − 0.53 97 W − 1.08 103
118 S 21.50 14 S 2.07 42 S 4.42 9
119 S 17.90 33 S 5.62 7 S 4.73 8
120 W − 0.78 121 S 2.34 37 S 4.93 7
121 S 10.16 79 W − 1.36 104 W − 1.14 105
122 S 21.84 13 S 1.15 62 S 0.66 74
123 S 14.15 57 S 14.48 1 S 13.38 1
124 S 9.73 85 S 0.60 80 W − 3.43 117
125 S 12.59 68 S 1.39 57 S 2.94 20
126 S 4.92 107 W − 4.33 120 S 0.21 85
127 S 15.66 50 S 0.37 86 W − 1.29 108
128 S 9.60 87 W − 1.47 107 S 0.92 67
129 S 12.68 66 S 8.62 2 S 7.60 2
AVG—Service S 10.81 - S 0.54 - S 0.06 -

Note: AVG stands for average, S stands for strengthen, and W stands for weaken.

Table 13
The chi-square tests on the endogeneity of inputs.
Model DF The observed value of H The critical value at the 5% level Conclusion

1 3 7.083 7.815 H0
2 4 8.799 9.488 H0
3 4 8.651 9.488 H0
4 5 10.081 11.070 H0
5 6 12.083 12.592 H0
6 7 12.317 14.067 H0

Notes: H stands for the statistic of the Hausman test and the critical value is based on the chi-square distribution at the 5% level.

19
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

presence of IT and EC are smaller than those from Model 5 with sole the state their values and to make correct decisions on IT and EC
appearance of IT. . investments.
Finally, we have considered a total of 40 industries, dividing into 16 Second, the influence of IT and EC upon the performance of US
manufacturing and 24 service industries. It is apparent that (i) the APEs manufacturing and service firms and industries also differ, depending on
of all 16 manufacturing and 22 out 24 service industries fall into the whether they are assessed separately or jointly. Therefore, again, it is of
group composed of these 38 (= 16 + 22) industries that are immune crucial importance to assess the business values of IT and EC separately
from the IT PP disease; (ii) meanwhile, there are only two service in­ and simultaneously and, then, compare the results to avoid over- or
dustries (Nos. 518 and 623) that suffer from the IT PP; (iii) 14 under-stating the values of IT and EC.
manufacturing and 17 service industries encounter the COM relation­ For a significant majority of manufacturing and service firms, IT and
ship between IT and EC; and (iv) only 2 manufacturing industries (Nos. EC contribute positively to their APEs, as they are present jointly and
311 and 315) and 6 service industries (Nos. 448, 453, 483, 622, 721, and separately, thereby rendering the importance of IT and EC investments,
722) appear to face the SUB relationship between IT and EC. no matter whether they are assessed in Case 1 or Case 2 with different
The COM and SUB phenomena and the PP could be explained by the models at the firm, industry and sector levels. The empirical evidence
distinctive business functions and implementation management of IT suggests that overall, the power of IT and EC to impact the performance
and EC investments in different industries (and firms as well). Accord­ of the service sector is slightly (less than 1%) weaker than that of the
ingly, the PP of IT and the SUB phenomenon caused by the joint presence manufacturing sector. One possible explanation could be the different
of IT and EC are managerial and strategic issues of great importance and focuses in the business functions. For example, the manufacturing
interest. companies focus on the production processes of tangible goods while the
service companies tend to focus on intangible services, thereby leading
4.2.2.4. Sector level analysis. At the sector level, we can observe from to different value-creating processes and applications of IT and EC.
Tables 8 and 9 that the results in Case 2 are consistent with those in Case Third, the complement and substitution relationships of different
1 and that IT and EC contribute positively to the APEs of both sectors inputs indeed have something to do with the performance of US
regardless of whether they are assessed jointly or separately. manufacturing and service firms. The usage of IT also impacts the per­
For the manufacturing sector (Table 8), the sector’s APE is 0.69 formance of other inputs such as ordinary capital (non-IT capital) and
(Model 1) in the absence of both IT as an input and EC as a firm char­ ordinary labor. Under the Case 1 setting, the estimates (Table 2) of the
acteristic (or a determinant of inefficiency); when IT is treated as an substitution parameters suggest that the substitution relationship be­
input without EC (Model 5), the APE increases to 0.77; and when IT is tween traditional capital and traditional labor exists in all four models (i.
treated as an input and EC as a firm characteristic jointly (Model 6), the e., Models 1 to 4). In Case 2 (Table 3), the complement relationship
APE is 0.80. Consequently, when IT is assessed alone, the APE increases between Kit and Lit exists in Model 5 (− 0.69) and Model 6 (− 0.10); but
by 11.6% (from 0.69 to 0.77); and when IT and EC are assessed simul­ the relationships between Kit and Iit and between Lit and Iit in Models 5
taneously, the APE increases by 15.9% (from 0.69 to 0.80). Thus, the and 6 are substitutable. These different relationships among inputs
empirical evidence implies that the COM phenomenon, taking place imply the importance of treating IT as an input factor in the production
because of the joint presence of IT and EC, exists since the APE of Model function and how the deployment of IT changes the cooperation of Kit
6 in the joint presence of IT and EC is greater than that of Model 5 in the and Lit with Iit . A further discussion of this issue is provided in Section
sole presence of IT. The empirical evidence further implies that the PP of 5.2 below.
IT is absent from the manufacturing sector due to the fact that the APE Fourth, we find that the IT productivity paradox (PP) does exist in
(0.77) of Model 5 in the presence of IT is larger than its counterpart two manufacturing firms (34 and 37), three service firms (83, 90, and
(0.69) of Model 1 in the absence of IT. Therefore, we conclude that the 120), 0 manufacturing industries, and two service industries (518 and
PP is absent from, and the COM phenomenon is existent in, the 623); but both the manufacturing and service sectors are immune from
manufacturing sector. the PP disease.
For the service sector (Table 9), the APE of Model 1in the absence of Fifth, our empirical results indicate that the SUB phenomenon takes
IT as an input and EC as an inefficiency determinant (or correction) is place in 12 manufacturing firms (e.g., firms 12, 18, 37, etc.), 13 service
0.69; yet, the APE of Model 5 with IT as an input is 0.77; and the APE of firms (e.g., firms 94, 112, 124, etc.), 0 manufacturing industries, and 6
Model 6 with both IT and EC investments is 0.79. Again, like the service industries (448, 453, 483, 622, 721, and 722). In other words,
manufacturing sector, the empirical evidence implies that the COM the rest of firms, industries, and the manufacturing and service sectors
phenomenon exists along with the absence of the IT PP in the service are facing or experiencing the phenomenon of COM.
sector. Sixth and final, the same ranking of the six CES-based research
models, as demonstrated by the Wald (or Z) and LM tests, implies that
5. Findings and managerial implications making no IT and EC investments (Model 1) in the production process
and management is the firm’s worst operations strategy and that
5.1. Major findings: answers to the five major research questions employing IT capital as an input alongside traditional capital and
traditional labor and, simultaneously, engaging EC investments as an
Based on the detailed discussion of the vast amount of the empirical inefficiency-correcting instrument (Model 6) is the best strategy, as far
results at the firm, industry, and sector levels conducted in the preceding as the firm’s performance (measured by productive efficiency) is con­
Section 4, we are able to summarize major findings and answers to the cerned. This conclusion is obviously consistent to the APE values of the
research questions as follows. manufacturing and service sectors. For the manufacturing sector, the
First, the business values of IT and EC at the firm, industry, and sector ranking of the APEs of the six models is 0.80 (Model 6) > 0.79 (Model 4)
levels differ, depending largely on whether they are evaluated sepa­ > 0.77 (Model 5) > 0.76 (Model 2) > 0.73 (Model 3) > 0.69 (Model 1).
rately or jointly. Accordingly, it is instructive to evaluate the values of IT For the service sector, the ranking of the APEs of the six models is exactly
and EC independently and also jointly in order not to under- or over- the same, that is, 0.79 (Model 6) ≥0.79 (Model 4) > 0.77 (Model 5) >
0.76 (Model 2) > 0.73 (Model 3) > 0.69 (Model 1).

20
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

5.2. The impacts of the substitution and complement relationships of According to the analysis in the three aspects, this study demon­
inputs upon productive efficiency strates the vital and substantial effects of substitution and complement
on the IT and EC values and, hence, the performance of the firms in both
As mentioned above, capital and labor are substitutable based on the manufacturing and service sectors. In order to improve the firm’s
Models 1, 2, 3, and 4 in which IT-capital does not enter the expected performance, we strongly suggest that managers analyze the possible
(maximum) output function. However, when IT-capital is treated as an substitution and complement relationships among the three inputs (K, L,
input, capital and labor are no longer substitutable but are comple­ and, I) prior to making IT and/or EC investment decisions in the oper­
mentary; labor and IT-capital are substitutable, and capital and IT- ations and production process. Also, our empirical results support the
capital are also substitutable according to Models 5 and 6. In fact, the notion that it is necessary to treat IT in two different ways (as an input or
substitution and complement among capital, labor, and IT-capital are as a determinant or correction) of inefficiency) in order to accomplish
not only very complex (Lin and Shao 2006; Chen and Lin 2009) but also the precise and reliable outcomes. Similarly, when assessing the values
impact IT and EC values substantially in the operations and production of IT and EC simultaneously and their impacts upon the firm’s perfor­
processes. Hence, in order to address this interesting and critical issue, mance, one may consider the Case 1 and Case 2 settings to decide
we discuss it in three highly relevant aspects as follows. whether IT should be treated as a determinant of inefficiency (Models 2
The first aspect is relative to whether or not the substitution and and 4) or as an input factor in the production function (Models 5 and 6).
complement relationships strengthen (S) or weaken (W) the value of IT
in the absence of EC (i.e., IT is assessed separately from EC) when IT is 5.3. Specification tests on the identification assumptions (Ahn et al.,
treated as an input (Case 2) along with capital and labor, compared to 2007)
when IT is not used (absent). Table 12 shows the performance com­
parisons of relevant models with and without inputs substitution and There are fundamental differences in the identification assumptions
complement. From the results presented in Columns (a) of Table 12, we and specification tests between a classical multiple linear regression
find that there were 120 firms which have earned S and only 9 firms model (CMLRM) and a nonlinear (time-varying) stochastic production
(firms 34, 37, 51, 53, 76, 83, 90, 109, and 120) which have secured W. In frontier model (NSPFM). The nonlinearity underlying a production
this aspect, a positive percentage increase means that treating IT as an function f( ⋅) and a normal-half normal assumption on wit − uit (or wit −
input contributes (leads) to a higher productive efficiency than without εit ) are notable. In the first place, an important specification error test
IT. Firm 28 won the first place with a 26% increase compared to the last relates to the existence (absence) of the one-sided error (Schmidt and
place achieved by firm 51 with a 27.84% decrease. The empirical evi­ Lin, 1984). Consider Model 1. Under the half-normal assumption, the
dence clearly suggests that the IT productivity paradox does not exist for hypothesis to test is H0 : σ 2u = 0 against H1 : σ 2u > 0 (Coelli et al., 2005;
a firm in both the manufacturing and service sectors under the impacts Schmidt and Lin, 1984). According to Ahn et al. (2007), the hypothesis is
of substitution and complement. On the contrary, the 9 firms mentioned equivalent to testing H0 : δ = 0 against H1 : δ > 0, where δ = σ u /σw . The
above have faced the productivity paradox when IT is treated as a
test statistic is called a Wald or Z statistic = ̂ δ) ∼ N(0, 1),
δ/S.D.(̂
production factor (Case 2) along with ordinary capital and ordinary
labor. where ̂δ is the nonlinear maximum likelihood (NML) estimate of δ and S.
The second aspect concerns whether or not the substitution and D.(̂
δ) is the NML estimate of S.D.(δ) (Coelli, 1995; Coelli et al., 2005).
complement strengthen (S) or weaken (W) the value of IT in the absence The observed value of the test statistic is 7.91 which is greater than the
of EC (i.e., IT is assessed separately from EC) when IT is treated as an critical value of 2.326 at the 1% level and 1.645 at the 5% level,
input factor (Case 2), compared to when IT is not treated as an input implying that the presence of uit in Model 1 is justified statistically and
(Case 1). We now consider the results of Model 5 vs Model 4 reported in confirmed by the data.
Column (b) of Table 12 and observe that S is associated with 91 firms The same procedures are applied to justify the presence of uit in
and W with 38 firms. In this aspect, a positive percentage increase means Model 5 and the presence of εit in Models 2 to 4 and Model 6. Consider
that treating IT as an input contributes to a higher productive efficiency Model 3 as an example. The hypothesis to test is H0 : σ2ε = 0 vs.
than treating IT as an inefficiency determinant (adjustment/correction). H1 : σ 2ε > 0. The Z test leads to concluding that the alternative hypoth­
Firm 123 ranks the first with an increase of 14.48% while firm 83 ranks esis is not rejected because the observed value of the test statistic is
the last with a decrease of 10.92%. The empirical evidence suggests that, 10.80 which is greater than the critical values at the 1% and 5% levels.
under the impacts of substitution and complement, the conclusions For Models 2, 4, 5 and 6, we have also calculated the observed values of
reached by Models 5 and 6 where IT is treated as an input (Case 2) differ their test statistics, 11.02, 12.63, 11.18, and 12.75, respectively.
from those reached by Models 2 and 4 where IT is treated as an in­ To sum up, the test conclusions strongly suggest that the presence of
efficiency determinant (Case 1). uit in Models 1 and 5 and εit in Models 2, 3, 4 and 6 are confirmed by the
The third (final) aspect has to do with whether or not the substitution data and justified statistically; and that, according to the observed
and complement strengthen (S) or weaken (W) the values of IT and EC in values of the test statistics, Model 6 is the most desired setting while
the joint presence of IT and EC (i.e., IT and EC are assessed simulta­ Model 1 is the least desired one, or the ranking is Model 6 > Model 4 >
neously) when IT is treated as an input factor (Case 2), compared to Model 5 > Model 2 > Model 3 > Model 1.
when IT is considered as a determinant (correction) of inefficiency (Case In the second place, a test of the half normal distribution assumption
1). The results of Model 6 vs Model 4 shown in Column (c) of Table 12 underlying uit in Models 1 and 5 and εit in Models 2, 3, 4, and 6 is called
indicate that S goes to a majority of 90 firms and W to 39 firms. In this for. Lee (1983) has developed a Lagrangian multiplier (LM) test on H0 :
aspect, a positive percentage increase means that treating IT as an input uit (or εit ) is half-normally distributed against H1 : uit (or εit ) is not. The
contributes to a higher productive efficiency than treating IT as an in­ LM test statistic is shown to be chi-square distributed with three degrees
efficiency determinant (correction) when IT and EC are assessed jointly. of freedom under H0 . Comparing the observed values of the test statistic,
Again, firm 123 wins the first place with a 13.38% improvement in its 5.70, 4.05, 3.61, 3.34, 2.77, and 2.58 for Models 1 to 6, respectively, to
productive efficiency, and firm 83 remains at the last place with a the critical values, 7.815 and 11.348 at the 5% and 1% levels, we can
decrease of 10.88%. That is, firm 123 would benefit from the substitu­ conclude that the half-normal distribution (H0 ) is not rejected, and that
tion and complement relationships among capital, labor, and IT-capital Model 6 > Model 4 > Model 5 > Model 2 >Model 3 >Model 1, the same
when IT is treated as a production factor and IT and EC are assessed ranking as concluded in the Z test.
jointly. In contrast, treating IT as an inefficiency determinant (correction In the third place, to claim that information technology capital has
or adjustment) and assessing IT and EC jointly would be a better strategy the power to influence inefficiency, we have computed the (grand
for firm 83. average) correlations over i and t between information capital and

21
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

inefficiency for each of the research models involving information strong theoretical foundations, and (v) the foremost condition estab­
technology capital. These are 0.91, 0.89, 0.86, and 0.80 for Models 6, 4, lished by the specification tests to confirm the identification assump­
5, and 2, respectively. The evidence suggests that information technol­ tions (i.e., the normal-half normal and the absence of the endogeneity of
ogy capital and inefficiency are highly correlated through an economic inputs) throughout the six NSPFM’s.
model. First, to achieve our multiple research goals, using IT and EC, as
In the fourth place, the interrelationships among the inputs are spelled out in Section 1, we do these: (a) Adding Iit to Model 1 of one
relative to the issues of nonlinearity (as specified by CES) and multi­ equation and two factors, we get Model 5 of one equation and three
collinearity. Nonlinearity has been transformed into linearity (see factors, with all the estimation parameters remaining significant at the
Appendices A and B). As far as multicollinearity is concerned, we have 1% level and R2 increasing from 0.78 to 0.89 (see Table 3); (b) Adding Iit
computed the simple correlations using the Pearson’s formula as given to Model 1 yields Model 2 of two equations and two factors, with all the
below: r (K,Y) = 0.78, r (L,Y) = 0.84, r (I,Y) = 0.90, r (EC,Y) = 0.56, r (K, five estimation parameters being significant at the 1% level and R2
L) = 0.61, r (L,I) = 0.65 = r (K,I), r (EC,K) = 0.30, r (EC,L) = 0.32, and r increasing from 0.78 to 0.88 (see Table 2); (c) Introducing Eit into Model
(EC,I) = 0.45. Thus, the empirical evidence suggests that multi­ 1 to create Model 3 of two equations and two inputs, we can observe that
collinearity is of little concern. all the five parameters are significant at the 1% level and R2 increases
In the fifth and final place, as mentioned above, the endogeneity of from 0.78 to 0.86; (d) Introducing Iit to Model 3 leads to Model 4 of two
inputs is at the heart of modern discussions on inefficiency. Fortunately, equations and two inputs, and we find that all the six parameters are still
the Wu’s (1973) F test or the Hausman’s (1978) chi-square test is valid significant at the 1% level and R2 increases to 0.92 from 0.86; and (e)
for testing whether or not the endogeneity of inputs problem exists in Finally, adding Eit to Model 5 yields Model 6 of two equations and three
our CES-based SPF models as shown in Subsection 5.4 below. factors; and it can be observed that all the eight parameters are signif­
icant at the 1% level and R2 increases from 0.89 to 0.92 (the highest and
5.4. The Hausman’s chi-square tests on the endogeneity of inputs the same as Model 4).
Second, as pointed out in Section 3.8 above,the 2SNML method of
One big advantage of deploying the CES functions is that the six CES- estimation is a valid and legitimate procedure, especially under the
based research models are linear in both unknown coefficients and re­ condition where the endogeneity of inputs is absent, to obtain the biased
gressors (data), thereby enabling us to perform the Hausman’s (1978) but consistent results.
chi-square test which turns out to be the same tests of Durbin (1954) and Third, the theoretical foundations underlying the six models of one
Wu (1973), as mentioned in Greene (2012) and Karakaplan and Kutlu equation with 2 factors or three factors and two equations with 2 factors
(2017). However, in the Hausman test (and the Wu test as well), the or three factors are strong. The models and factors as well are delineated
cross-sectional data and the OLS method are used. Here, our data are a by two theories, namely, the theory of production and the theory of
panel set and, hence, OLS is no longer valid. Therefore, we apply Avery’s errors in variables, as detailed in Sections 3.2 and 3.3.
(1977) error components-seemingly unrelated regression (EC-SUR) Fourth and final, the most critical and vital condition for the undis­
procedure (cf., Lin et al., 1987, for an application) to pool the puted evidence of robustness is that the identification assumptions un­
cross-sectional and time-series (or panel) data in order to estimate each derlying the six NSPFM’s are satisfied as proved by the various
of the six research models. specification tests undertaken in the preceding Subsections 5.3 and 5.4.
The Hausman test also requires an instrumental variables (IV) On the basis of the tests, we are also able to rank the six models.
equation for each research model. This requirement is met by using the In sum, the various robustness checks strongly suggest that our
same research models with variables lagged one period as suggested by research models are sound and justified by the various specification tests
Greene (2012). For example, for Model 1, the IV equation involves and that the 2SNML estimation procedure is appropriate and valid.
Consequently, from the statistical and econometric points of view, the
ln Yi,t− 1 , ln Ki,t− 1 , ln Li,t− 1 , and (lnKi,t− 1 − lnLi,t− 1 )2 = [ln(Ki,t− 1 /Li,t− 1 )]2 =
results are robust and reliable. Furthermore, the results of the six models
Z1i,t− 1 (say). Similarly, (ln Li,t− 1 − ln Ii,t− 1 )2 = Z2i,t− 1 and provide a highly comprehensive comparative analysis. As a note,
2
(ln Ki,t− 1 − ln Ii,t− 1 ) = Z3i,t− 1 . Also, note that the hypothesis to be tested virtually, econometrics texts like Johnston (1963), Goldberger (1964),
consists of the null hypothesis (H0): the exogeneity of inputs vs. the Theil (1971), Zellner (1971), Kmenta (1997), Pindyck and Rubinfeld
alternative hypothesis (H1): the endogeneity of inputs. The test con­ (1998), Stewart (2005), and Greene (2012) contain no sensitivity
clusions are summarized in Table 13 and, the conclusions suggest that analysis. However, sensitivity analysis with forecasting is needed to
the endogeneity of inputs is absent from our data. evaluate forecast errors (Pindyck and Rubinfeld, 1998).
The test results imply that the problem of inputs endogeneity is more
or less relative to the data used and that it has something to do with 5.6. Managerial and strategic implications
model specifications. In other words, the high quality and accuracy of
the data used play a crucial role and the high quality of data depends In the first place, the joint and separate assessments of IT and EC are
heavily on the reliability and credibility of data sources; and the necessary and equally important. Under the Case 1 setting, we evaluated
generalized two-equation SPF approach (Chen and Lin, 2009) is more and compared the contributions of IT and EC through Models 1, 2, 3, and
favorable than the original one-equation SPF model as far as the issue of 4. We found that the contributions of IT and EC differ. Specifically, the
inputs endogeneity is concerned. effects of IT upon the performance of US firms, industry, and sectors are
generally greater than that of EC; and, comparing the separate evalua­
5.5. A robustness analysis tion (Model 2 and Model 3) to the joint presence of IT and EC (Model 4),
we find that Model 4 exerts greater contributions to the APE than Models
In this section, we discuss the robustness of the six NSPFM’s and the 2 and 3. The power of the joint presence of IT and EC is also confirmed
results (parameter estimates) obtained from the 2SNML method, based under the Case 2 setting as well. That is, the APEs of Model 6 are
on (i) the statistical significance of the estimates, (ii) the models’ generally greater than that of Model 5. These results imply that to gauge
explanatory power gauged by R2, (iii) the estimation method, (iv) the the influence of IT and EC upon the performance of the firm, industry

22
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

and sector, the business values of IT and EC should be evaluated not only The empirical evidence confirms that the contributions of IT and EC
separately but also jointly. Otherwise, the contributions and effects of IT to the firm’s performance as measured by productive efficiency differ as
and EC on the firm’s performance could be miscalculated, resulting in their values are assessed independently and simultaneously. Therefore,
misleading conclusions. a useful managerial implication is that it is of vital importance to
In the second place, the effects of IT and EC on the performance of US compare the values of IT and EC as they are evaluated separately and
business organizations, as measured by productive efficiency, are posi­ jointly. Also, as an input, IT capital does contribute to productive effi­
tive. Regardless of whether the Case 1 or the Case 2 setting is chosen, IT ciency; consequently, the productivity paradox has not been found to
and EC generally contribute positively to the APEs at the firm, industry, exist, in most of the firms and industries considered and the two sectors
and sector levels. This evidence provides a second important implication as well. Furthermore, according to the substitution parameters, the
for US business organizations to think carefully about how to deploy the substitution relationship has been discovered to exist among ordinary
investments of IT and EC in the value-creating process. Nowadays, the capital, labor, and IT capital in both Cases 1 and 2. However, when IT is
degree of environment uncertainty is higher than ever before. It is both treated as an input (Case 2), the complement relationship between or­
important and necessary for managers to respond to the fast changing dinary capital and ordinary labor has been found (Model 5 and Model 6).
environment flexibly and accurately. Under this situation, the applica­ In other words, the deployment of Iit in the production process changes
tion of Case 1 and Case 2 may provide a great solution to the IT and EC the relationship between Kit and Lit from substitution to complement.
investments deployment decision-making problem for business organi­ Finally, the substitutability phenomenon has taken place in an insig­
zations. In short, the second important implication is that carefully nificant number of firms and industries. In other words, the comple­
analyzing and comparing the results of Case 1 and Case 2 are a must. mentarity relationship between IT and EC occur in most of the firms and
In the third place, each and every piece of our empirical evidence industries under study and in the two sectors as well.
leads to suggesting that the generalized two-equation SPF model (Chen As a first note, a question may be raised as to whether there are any
and Lin, 2009) prefers to the traditional one-equation approach (e.g., contextual variables that are needed to be taken into account. Our
Aigner et al., 1997). answer is no. We don’t have any contextual variables that are needed to
In the fourth (last, but not the least) place, IT plays a crucial role in be considered, according to the definition of contextual variables
the complex complement and substitution relationships between tradi­ (Worchel, 1986). There are at least five contextual variables, namely,
tional (non-IT) capital and traditional labor. The empirical results of environment, corporate strategy, technology, organization size, and
Case 1 indicate that, when IT is treated as a determinant (correction) of workforce. These factors in the firm’s context determine the appropriate
inefficiency in Case 1, all of the pairwise substitution parameters of Kit managerial strategy and structure of an organization. Strategy and or­
and Lit (ρKL, for Model 1: 0.82, Model 2: 0.26, Model 3: 0.67, and Model ganization are not the focus of this research. However, in this study,
4: 0.22) are greater than zero, meaning that the relationship between Kit information technology (Iit) and e-commerce (Eit) may be referred to as a
and Lit is substitutable. However, when IT is treated as an input factor in proxy of technology; and traditional labor (Lit) as a proxy of workforce. It
Case 2, the relationship between Kit and Lit has changed from substitu­ seems to us that contextual variables are frequently used in sociology,
tion to complement (Model 5: ρKL = -0.69 and Model 6: ρKL = -0.10), psychology, and human relationships. Thus, the contextual variables are
while the relationship between Iit and Kit and between Iit and Lit are delineated from personal and interpersonal characteristics (Worchel,
substitutable because ρKL and ρLI are positive in both Models 5 and 6. 1986). If we use any contextual variables incorrectly, our results will be
These interesting results provide an answer to the controversial issue distorted.
associated with the IT deployment. Some people may insist that IT in­ As a second note, we suggest three extensions for future research.
vestment usually involves a huge amount of money and could possibly One extension has to do with the substitutability phenomenon between
crowd out the existing labor force, while others may argue that IT in­ IT and EC that was discovered in some manufacturing and service firms,
vestment could play the role of a facilitator in the US business organi­ as well as some manufacturing and service industries. This phenomenon
zations. As such, our empirical results imply that the US business may be caused by the distinctive business functions and environments or
organizations deploying IT could work, process, and cooperate much others. A further investigation to explain this phenomenon is called for.
more efficiently internally and externally than deploying no IT, provided Second, it is both instructive and interesting to compare the results
that people are convinced by referring to our empirical results that based on the CES-based SPF approaches with their counterparts based
confirm that IT contributes positively to APEs in both Cases 1 and 2 at on the CES-based partial adjustment valuation (PAV) approaches (Lin
the firm, industry, and sector levels. et al., 2010, 2019; Lin and Kao, 2014). Third, the present study may be
moved a step forward by considering efficiency change in relation with
6. Conclusion technological progress and productivity growth (Lin, 2013) or by un­
dertaking a Bayesian PAV analysis in a similar framework as Kim et al.
While the adoptions of IT and EC have caught a lot of attention in (2015).
research and practice since 1990’s, the assessment of the business values
of IT and EC has become increasingly prominent and important not only Acknowledgements
in business practice but also in academic research, since the investments
of IT and EC have increased consistently over time. The valuable comments and suggestions from two anonymous ref­
However, virtually the existing IT/EC-related studies assessed the erees of this journal are greatly appreciated. Earlier versions of this
values of IT and EC separately instead of jointly, leading to the conse­ paper were presented at the Joint Weekly Doctoral Seminars of the
quences of under- or over-stated values of IT and EC, as well as the Departments of Operations Management and Strategy and Management
misleading or compromised conclusions. To bridge the gap, this study Science and Systems in the School Management at The State University
evaluates empirically the values of IT and EC separately and jointly of New York at Buffalo, the Annual Meeting of the Production and Op­
based on six CES-based (time-varying) stochastic production frontier erations Management Society, and the International Conference on En­
(SPF) models derived from the theory of production and the theory of gineering, Technology, and Applied Science (ICETA) in Taipei, Taiwan.
errors in variables. The authors thank the participants of the Weekly Seminars and the

23
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

conferences for their helpful comments and suggestions. Funding: The School of Management, The State University of New York at Buffalo, for
authors gratefully acknowledge that this work was supported by the a two-year period; and by the College of Management at National Sun
faculty fellowships granted by the Faculty Research Committee in the Yat-sen University.

Appendix A

The two-factor CES-based SPF specification (Arrow et al., 1961; Kmenta, 1997; Lin and Shao, 2006; Greene, 2012)

Step 1: The two-factor CES production function for panel data can be written
( )− α ( ′ )
(A1)
′
Yit = γ δKit− ρ + (1 − δ)L−it ρ ρ exp − uit + wit

where γ is the efficiency parameter; δ is the distribution parameter, α is the return to scale parameter, and ρ is the substitution (or complement)
parameter. In Arrow et al. (1961), α is set equal to 1. In equation (A1), α is allowed to be determined by the data.

Step 2: Take the natural logarithm of equation (A1) to get

(α) ( )
ln Yit = ln γ − ln δKit− ρ + (1 − δ)L−it ρ − uit + wit (A2)
ρ

Step 3: Using a Taylor series expansion approximation to equation (A2) in Step 2 around the point ρ = 0 gives the restricted (structural, implicit)
model as follows:
( )
1
lnYit = ln γ + αδlnKit + α(1 − δ)lnLit − αδρ(1 − ρ)(lnKit − lnLit )2 − uit + wit (A3)
2

which is the structural two-factor CES-based SPF specification.

Step 4: The unrestricted (non-structural, explicit, estimable) counter part of the restricted equation given in

Step 3 is

lnYit = βo + β1 lnKit + β2 lnLit + β4 (lnKit − lnLit )2 − uit + wit

(A4)
= β0 + β1 X1it + β2 X2it + β4 X4it − uit + wit

()
where β0 = lnγ, β1 = αδ, β2 = α(1 − δ), β4 = − 1
2 αδρ(1 − δ), Xit = lnKit , X2it = lnLit , and X4it = (lnKit − lnLit )2. Note that equation (A4) is Model 1 in
Subsection 3.7.

Step 5: Therefore, the coefficients of structural Model (A3) correspond uniquely to the coefficients of non-structural Model (A4) as follows.
/
β1
γ = anti ln β0 = exp(β0 ), α = β1 + β2 , δ = , and ρ = − 2β4 (β1 + β2 ) (β1 β2 ). (A5)
β1 + β2
Thus, we can first estimate the coefficients of Model (A4) and then the estimation values of the coefficients of Model (A3) can be obtained uniquely,
using the relationships in (A5). There is no identification problem, that is, the parameters in (A3) are exactly identified.

Step 6: Assuming g(Zit ; α) = α0 + α1 Eit and combining the two-factor two equation model into a single restricted (structural) model and a single
unrestricted (non-structural) model. The unrestricted model is Model 3 or equation (10) in Subsection 3.7. The same procedure can be applied to
Models 2 and 4 or equations (9) and (11).

Table A
Correspondences between ρ or ρi and σ in the CES production models.

Parameter of substitution: ρ or ρi (i = KL, LI, KI) Elasticity of substitution: σ Economic meaning

Range: [-1, ∞] Range: [-∞, ∞]

0 1 The CES reduces to the Cobb-Douglas function with constant returns to scale
∞ ∞ The CES reduces to fixed proportions (a straight line)
>0 >0 Inputs substitution
− 1≤ρ, ρi < 0 (i = KL, LI, KI)) <0 Inputs complement

Note: This Table A can be used to determine the substitution and complement relationships among Kit , Lit , and Iit .
Source: Chen and Lin (2009, Table 1, P. 558).

24
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Appendix B

The three-factor CES-based SPF specification (Chen and Lin, 2009)

Step 1: The three-factor CES production function for panel data can be described by
( )− α ( ′ ′ )
Yit = γ δ1 Kit− ρ + δ2 L−it ρ + (1 − δ1 − δ2 )Iit− ρ ρ exp − uit + wit ,
(B1)
γ > 0, 1 > δ1 , δ2 > 0, α > 0, ρ ≥ − 1, i = 1, …, m and t = 1, …, n

where γ is the efficiency parameter; δ1 , δ2 , and 1 − δ1 − δ2 are the distribution parameters for Kit , Lit , and Iit , respectively; α is the return to scale
parameter; and ρ is the substitution or complement parameter.

Step 2: Take the natural logarithm of equation (B1) to get

(α) [ ]
ln Yit = ln γ − ln δ1 Kit− ρ + δ2 L−it ρ + (1 − δ1 − δ2 )Iit− ρ − uit + wit (B2)
ρ

Step 3: A Taylor series expansion approximation to equation. (B2) around the point ρ = 0 is given by
ln Yit = ln γ + αδ1 ln Kit + αδ2 ln Lit + α(1 − δ1 − δ2 )ln Iit
( ) ( )
−
1
ραδ1 δ2 (ln Kit − ln Lit )2 −
1
ραδ2 (1 − δ1 − δ2 )(ln Lit − ln Iit )2 (B3)
2 ( ) 2
1
− ραδ1 (1 − δ1 − δ2 )(ln Kit − ln Iit )2 − uit + wit ,
2

which is the restricted (structural) three-factor CES-based SPF specification.

Step 4: The unrestricted (non-structural) counterpart of the restricted equation (B3) is

ln Yit = β0 + β1 ln Kit + β2 ln Lit + β3 ln Iit + β4 (ln Kit − ln Lit )2 + β5 (ln Lit − ln Iit )2 + β6 (ln Kit − ln Iit )2 − uit + wit
(B4)
= β0 + β1 X1it + β2 X2it + β3 X3it + β4 X4it + β5 X5it + β6 X6it − uit + wit

() () ()
1 1 1
where β0 = lnγ, β1 = αδ1 , β2 = αδ2 , β3 = α(1 − δ1 − δ2 ), β4 = − 2 ραδ1 δ2 , β5 = − 2 ραδ2 (1 − δ1 − δ2 ), and β6 = − 2 ραδ1 (1 − δ1 − δ2 ).
Model (B4) is Model 5 or equation. (12) in Subsection 3.7.

Step 5: Thus, the coefficients of restricted equation (B3) relate uniquely to those of the unrestricted equation (B4) as follows:
β1 β2 β2
γ = anti ln β0 = exp(β0 ), δ1 = , δ2 = , δ2 =
(β1 + β2 + β3 ) (β1 + β2 + β3 ) (β1 + β2 + β3 )
2β4 (β1 +β2 +β3 )] 2β5 (β1 +β2 +β3 )]
α = β1 + β2 + β3 , and three substitution (or complement) parameters: ρKL = [− β1 β2
between Kit and Lit , ρLI = [− β2 β3
between

[ − 2β6 (β1 + β2 + β3 )]
Lit and Iit , and ρKI = between Kit and Iit . (B5)
β 1 β3
The restricted model (B3) may be called the structural (implicit) model while the unrestricted model (B4) may be referred to as the non-structural
(explicit or estimable) model. We can observe that the explicit coefficients in Model (B4) correspond to the implicit coefficients in Model (B3) under
exact identification. Therefore, we can use the LIMDEP software to estimate the coefficients of Model (B4) and, then, the coefficients of Model (B3) can
be solved uniquely using the relationships in (B5). It is just identified and, therefore, there is no identification problem.

Step 6: Again, assuming g(Zit , α) = α0 + α2 Eit and combing the three factor-two equation model into a single restricted (structural) model and a
single unrestricted (non-structural) model. The unrestricted model is Model 6 or equation (13) in Subsection 3.7.

References Brynjolfsson, E., 1993. The productivity paradox of information technology. Commun.
ACM 36, 67–77.
Brynjolfsson, E., Hitt, L., 1996. Paradox lost? Firm-level evidence on the returns to
Aigner, D., Lovell, C.A.K., Schmidt, P., 1997. Formulation and estimation of stochastic
information systems spending. Manag. Sci. 42, 541–558.
frontier production function models. J. Econom. 6, 21–37.
Chen, Y.H., Lin, W.T., 2009. Analyzing the relationships between information
Ahn, S.C., Lee, Y.H., Schmidt, P., 2007. Stochastic frontier models with multiple time-
technology, inputs substitution and national characteristics based on CES stochastic
varying individual effects. J. Prod. Anal. 27, 1–12.
frontier production models. Int. J. Prod. Econ. 120, 552–569.
Arrow, K.J., Chenery, H.B., Minhas, B.S., Solow, R.M., 1961. Capital-labor substitution
Chou, Y.C., Shao, B.B.M., Lin, W.T., 2012. Performance evaluation of production of IT
and economic efficiency. Rev. Econ. Stat. 43, 225–250.
capital goods across OECD countries: a stochastic frontier approach to Malmquist
Avery, R.B., 1977. Error components and seemingly unrelated regression. Econmetrica
index. Decis. Support Syst. 54, 173–184.
45, 199–209.

25
W.T. Lin et al. International Journal of Production Economics 241 (2021) 108269

Coelli, T., 1995. Estimators and hypothesis tests for a stochastic frontier function: a Lin, W.T., Shao, B.B.M., 2006. The business value of information technology and inputs
Monte Carlo analysis. J. Prod. Anal. 6, 247–268. substitution: the productivity paradox revised. Decis. Support Syst. 42, 493–507.
Coelli, T., Rao, D.S.P., O’Donnell, C.J., Battese, G.E., 2005. An Introduction to Efficiency Lovell, C.A.K., 1993. Production frontiers and productive efficiency. In: Fried, H.O.,
and Productivity Analysis, second ed. Springer Science & Business Media, New York, Lovell, C.A.K., Schmidt, S.S. (Eds.), The Measurement of Productive Efficiency:
N.Y. Techniques and Applications. Oxford University Press, New York, pp. 3–67.
Cullen, A.J., Webster, M., 2007. A model of B2B e-commerce, based on connectivity and McAfee, A., 2002. The impact of enterprise information technology adoption on
purpose. Int. J. Oper. Prod. Manag. 27, 205–225. operational performance: an empirical investigation. Prod. Oper. Manag. 11, 33–53.
Devaraj, S., Fan, M., Kohli, R., 2002. Antecedents of B2C channel satisfaction and McAfee, A., Brynjolfsson, E., 2008. Investing in the IT that makes a competitive
preference: validating e-commerce metrics. Inf. Syst. Res. 13, 316–333. difference. Harv. Bus. Rev. 86, 98–107.
Dewan, S., Min, C., 1997. The substitution of information technology for other factors of Melville, N., Kraemer, K., Gurbaxani, V., 2004. Review: information technology and
production: a firm level analysis. Manag. Sci. 43, 1660–1675. organizational performance: an integrative model of IT business value. MIS Q. 28,
Disney, S.M., Naim, M.M., Porter, A., 2004. Assessing the impact of e-business on supply 283–322.
chain dynamics. Int. J. Prod. Econ. 89, 109–118. Melville, N., Gurbaxani, V., Kraemer, K., 2007. The productivity impact of information
Durbin, I., 1954. Errors in variables. Rev. Int. Stat. Inst. 22, 23–32. technology across competitive regimes: the role of industry concentration and
Gefen, D., Straub, D., 2000. The relative importance of perceived ease of use in IS dynamism. Decis. Support Syst. 43, 229–242.
adoption: a study of e-commerce adoption. J. Assoc. Inf. Syst. Online 1, 1–28. Article Napoleon, K., Gaimon, C., 2004. The creation of output and quality in services: a
8. framework to analyze information technology-worker systems. Prod. Oper. Manag.
Geoffrion, A.M., Krishnan, R., 2003. E-business and management science: mutual 13, 245–259.
impacts. Manag. Sci. 49, 1275-1286–1445-1456. Ord, K., Fildes, R., Kourentzes, N., 2017. Principles of Business Forecasting, second ed.
Goldberger, A.S., 1964. Econometric Theory. Wiley & Sons, New York, NY. South- Western, Cengage Learning, Boston, MA.
Grandon, E.E., Person, J.M., 2004. Electronic commerce adoption: an empirical study of Peacock, E., Tanniru, M., 2005. Activity-based justification of IT investments. Inf. Manag.
small and medium US businesses. Inf. Manag. 42, 197–216. 42, 424–425.
Greene, W.H., 2012. Econometric Analysis, seventh ed. Pearson Education, Inc., Upper Pindyck, R.S., Rubinfeld, D.L., 1998. Econometric Models and Economic Forecasts,
Saddle River, NJ. fourth ed. Irwin/McGraw–Hill, New York, N.Y.
Gunasekaran, A., Marri, H.B., McGaughey, R.E., Nebhwani, M.D., 2002. E-commerce and Prasad, S., Babbar, S., 2000. International operations management research. J. Oper.
its impact on operations management. Int. J. Prod. Econ. 75, 185–197. Manag. 18, 209–247.
Hausman, J.A., 1978. Specification tests in econometrics. Econometrica 46, 1251–1271. Ramanathan, R., 2011. An empirical analysis on the influence of risk on relationships
Johnston, J., 1963. Econometric Methods. McGraw Hill, New York, NY. between handling of product returns and customer loyalty in E-commerce. Int. J.
Karakaplan, M.U., Kutlu, L., 2017. Handling endogeneity in stochastic frontier analysis. Prod. Econ. 130, 225–261.
Econ. Bull. 37, 889–901. Sánchez-Rodríguez, C., Dewhurst, F.W., Martínez-Lorente, A.R., 2006. IT use in
Kim, G., Lin, W.T., Simpson, N.C., 2015. Evaluating the performance of US supporting TQM initiatives: an empirical investigation. Int. J. Oper. Prod. Manag.
manufacturing and service operations in the presence of IT: a Bayesian stochastic 26, 486–504.
production frontier approach. Int. J. Prod. Res. 53, 5500–5523. Schmidt, P., 1985. Frontier production functions. J. Econom. 4, 289–328.
Kmenta, J., 1997. Elements of Econometrics, second ed. The University of Michigan, Ann Schmidt, P., Lin, T.F., 1984. Simple tests of alternative specifications in stochastic
Arbor, MI. frontier models. J. Econom. 24, 349–361.
Kutlu, L., 2010. Battese-Coelli estimator with endogenous regressors. Econ. Lett. 109, Shee, A., Stefanou, S.E., 2015. Endogeneity corrected stochastic production frontier and
79–81. technical efficiency. Am. J. Agric. Econ. 97, 939–952.
Kutlu, L., Tran, K.C., Tsionas, M.G., 2019. A time-varying true individual effects model Stender, M., Ritz, T., 2006. Modeling of B2B mobile commerce processes. Int. J. Prod.
with endogenous regressors. J. Econom. 211, 539–559. Econ. 101, 128–139.
Lee, H.G., Clark, T.H., 1996. Impacts of the electronic marketplace on transaction cost Stewart, K.G., 2005. Introduction to Applied Econometrics. Brooks/Cole Thomson
and market structure. Int. J. Electron. Commer. 1, 127–149. Learning, Belmont, CA.
Lee, L.F., 1983. A test for distribution assumptions for the stochastic frontier functions. Subramani, M., Walden, E., 2001. The impact of e-commerce announcements on the
J. Econom. 22, 245–267. market value of firms. Inf. Syst. Res. 12, 135–154.
Lin, W.T., Chen, Y.H., Chatov, R., 1987. The demand for natural gas, electricity and Swaminathan, J.M., Tayur, S.R., 2003. Models for supply chains in e-business. Manag.
heating oil in the United States. Resour. Energy 9, 233–258. Sci. 49, 1387–1486.
Lin, W.T., 2009. The business value of information technology as measured by technical Thatcher, M.E., Oliver, J.R., 2001. The impact of technology investments on a firm’s
efficiency: evidence from country-level data. Decis. Support Syst. 46, 865–874. production efficiency, product quality, and productivity. J. Manag. Inf. Syst. 18,
Lin, W.T., 2013. Assessing the impacts of the integration of the ICT investments of 17–42.
Taiwan and China upon economic growth in Taiwan. In: Chow, Peter C.Y. (Ed.), Theil, H., 1971. Principles of Econometrics. Wiley & Sons, New York, NY.
Economic Integration across the Taiwan Strait: Global Perspectives. Edward Elgar Tran, K.C., Tsionas, E.M., 2013. GMM estimation of stochastic frontier model with
Publishing Limited Inc., Cheltenham, UK and Northampton, MA, USA, pp. 56–80. endogenous regressors. Econ. Lett. 118, 233–236.
Ch. 3. Triplett, J.E., Bosworth, B.P., 2004. Productivity in the U.S Service Sector: New Sources
Lin, W.T., Chen, Y.H., Shao, B.B.M., 2015. Assessing the business values of information of Economic Growth. Brookings Institute Press, Washington, D.C.
technology and e-commerce independently and jointly. Eur. J. Oper. Res. 245, Vakharia, A.J., 2002. E-business and supply chain management. Decis. Sci. J. 33,
815–827. 495–503.
Lin, W.T., Chen, Y.H., Chou, C.C., 2016. Assessing the values of EC and IT separately and Waldman, D.M., 1982. A stationary point for the stochastic frontier likelihood.
simultaneously and inputs substitution and complement. Inf. Manag. 53, 109–121. J. Econom. 19, 275–279.
Lin, W.T., Chuang, C.H., Choi, J.H., 2010. A partial adjustment approach to evaluating Warren, L.H., Nicholls, M.G., 1999. Dynamic benchmarks for operations evaluation using
and measuring the business value of information technology. Int. J. Prod. Econ. 127, soft computing methods. Decis. Sci. J. 30, 19–45.
158–172. Worchel, S., 1986. The influence of contextual variables on interpersonal spacing.
Lin, W.T., Chiang, C.Y., 2011. The impact of country characteristics upon the value of J. Nonverbal Behav. 10, 230–254.
information technology as measured by productive efficiency. Int. J. Prod. Econ. Wu, D.M., 1973. Alternative tests of independence between stochastic regressors and
132, 13–33. disturbances. Econometrica 41, 733–750.
Lin, W.T., Chuang, C.H., 2013. Investigating and comparing the dynamic patterns of the Zellner, A., 1971. An Introduction to Bayesian Inference in Econometrics. Wiley & Sons,
business value of information technology over time. Eur. J. Oper. Res. 228, 249–261. New York, NY.
Lin, W.T., Kao, T.W., 2014. The partial adjustment valuation approach with dynamic and Zhu, K., 2004. The complementarity of information technology infrastructure and e-
variable speeds of adjustment to evaluating and measuring the business value of commerce capacity: a resource-based assessment of their business value. J. Manag.
information technology. Eur. J. Oper. Res. 238, 208–220. Inf. Syst. 21, 167–202.
Lin, W.T., Chen, Y.H., Hung, T.S., 2019. A partial adjustment valuation approach with Zhu, K., Kraemer, K.L., 2002. E-commerce metrics for net-enhanced organizations:
stochastic and dynamic speeds of partial adjustment to measuring and evaluating the assessing the value of e-commerce to firm performance in the manufacturing sector.
business value of information technology. Eur. J. Oper. Res. 272, 766–779. Inf. Syst. Res. 13, 275–295.
Lin, W.T., Shao, B.B.M., 2000. Relative sizes of information technology investments and
productive efficiency: their linkage and empirical evidence. J. Assoc. Inf. Syst.
Online 1, 1–35. Article 7.

26