Electronic Commerce Research and Applications 11 (2012) 518–533

Contents lists available at SciVerse ScienceDirect

Electronic Commerce Research and Applications

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

The impact of ICT development on the global digital divide

Shing H. Doong a,1, Shu-Chun Ho b,⇑
a
Shu Te University, No. 59, Hengshan Rd., Yanchao District, Kaohsiung City 824, Taiwan
b
National Kaohsiung Normal University, No. 116, Heping 1st Rd., Lingya District, Kaohsiung City 802, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: Information and communication technology (ICT) has accelerated the growth of the global economy and
Available online 22 February 2012 improved the quality life of the world’s inhabitants. ICT has brought new ways of creating livelihoods for
people. The diffusion of ICT has also increased year by year and made it possible to reduce poverty. The
Keywords: opportunities created by ICT also may eventually decrease the ‘‘distance’’ between countries in many
Convergence other ways. Because access to ICT plays a key role in defining the global digital divide, it is important
Cross-country study to study how the ICT gaps among countries have changed. This study examines global ICT development
Data clustering
in the last decade. We collected secondary data for 136 countries from 2000 to 2008. Four relevant vari-
Digital divide
Information and communication technology
ables are used as proxies for the ICT development status of a country. Because of this multivariate nature
National development of the data, most previous studies have applied a composite index approach to represent the ICT status of
Peer effects a country. For this study, we developed a framework to reduce multivariate raw data into an ordinal
number representing a country’s ICT development level. The methodology behind the framework
involves data clustering and multi-dimensional data ranking. After applying this data reduction proce-
dure, we explored ICT development paths of different countries, and also conducted panel data analysis
based on gross national income and various fixed effects.
Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction more than half of the world’s inhabitants will have access to ICTs
by 2015. Mobile applications have been designed not only for voice
Information and communication technology (ICT) has fostered communications, but also business transactions and information
economic growth and social progress in the past few decades. Prior access (UNCTAD 2009). In developing countries, the number of
studies have shown that ICT plays a critical role in the national e- users using mobile devices to access the Internet has jumped up
commerce growth (Fathian et al. 2008, Ho et al. 2007, 2011), eco- rapidly too. As the largest developing country in the world, China
nomic growth (Hanafizadeh et al. 2009, Andrianaivo and Kpodar had 233 million mobile Internet users with an estimated annual
2011, Papaioannou and Dimelis 2007, Tcheng et al. 2007, Seo et al. growth rate of 51% in 2009 (CINIC 2010).
2009), and country development (Heeks 2008). Both developed Many countries have endeavored to develop ICT through heavy
and developing countries in the world have boosted their national resource investments over the years. Wealthier countries are con-
investments in ICT to drive their economic growth (Dewan and sidered to have more resources at their disposal in ICT develop-
Kraemer 2000, Andrianaivo and Kpodar 2011, Tcheng et al. 2007). ment and may have created a higher level of ICT development.
Heeks (2008) argued that ICT development requires new technolo- Thus, it is critical to investigate the result of the ICT investments
gies and new approaches to innovate and integrate. The diffusion of in the past decade. The objective of this paper is to explore the
ICT in recent years also has surprised many analysts who serve with global ICT development trend and to examine national wealth
leading international organizations. These include the United effects on the trend. Through regional network effects, countries
Nations, the World Bank, the Organization for Economic Cooperation with spatial proximity may influence one another in their ICT
and Development (OECD), and the International Telecommunica- development. Social influence theory (Friedkin 1998) stipulates
tion Union (ITU), as well as the governments of many countries. that an individual’s behavior may depend on the behavior of others
For example, benchmark progress in worldwide ICT access with to whom the individual is tied. The study of Agarwal et al. (2009)
an emphasis on mobile applications was realized in 2008. This was supported this social influence theory on an individual’s Internet
earlier than the prediction of 2015 made by the World Summit of adoption behavior, which may be influenced by the Internet
the Information Society (WSIS) in 2005. The estimation is that adoption behavior of the individual’s spatial peers. In addition to
national wealth effects on ICT development, we are also interested
⇑ Corresponding author. Tel.: +886 7 7172930x1707. in peer effects resulting from spatial proximity on the ICT develop-
E-mail addresses: [email protected] (S.H. Doong), [email protected] (S.-C. Ho). ment of a country. Thus, this study provides another example of
1
Tel.: +886 7 6158000x3004. social influence theory at the entity level of countries.

1567-4223/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.elerap.2012.02.002
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 519

We intend to answer the following research questions. What Techatassanasoontorn 2005a, Shirazi et al. 2009, Vicente and Lopez
are the trends of ICT development in a global sense for the last dec- 2011). We summarize these studies in terms of measures of digital
ade? Do countries with different gross national income (GNI) levels divide, data type, research method, data period, unit of analysis,
have different ICT development paths? Are there peer effects in the and variables examined in Table 1. These studies use wireless tech-
ICT development of countries with spatial proximity? nology, PC, Internet, and ICT indicators to measure the digital di-
In Section 2, we review the literature related to ICT and the vide. They also examine a large set of variables that may affect
framework in our analysis. Section 3 illustrates our model develop- the digital divide across countries. Most of these studies collected
ment and analysis approach: data clustering, cluster ranking, met- secondary data and performed cross-sectional and time-series
rics for measuring ICT gaps, and the panel study. In Section 4, we analyses (see Table 1).
describe the secondary data collected for data analysis and data Previous studies imply that opportunity to access ICT is a key
preprocessing according to our framework. Sections 5 and 6 pro- component in measuring the digital divide. In addition, Internet ac-
vide analysis and discussion, and Section 7 concludes with our cess, PC access, user digital capability, and government policy also
findings, contributions, and limitations. form a basis for measurement of the digital divide. ICT-related
measures are critical indicators that show the differences between
rich and poor countries (Chinn and Fairlie 2007, Cuervo and
2. Literature review
Menendez 2006). ICT opportunities and the digital divide have an
interesting bidirectional relationship: ICT opportunities influence
We first review studies in the digital divide. To analyze global
the digital divide, and the digital divide may hinder ICT opportuni-
ICT development, we will assess studies related to ICT that have
ties as well (OECD 2005). The ICT opportunities of a country seem
considered it as a general purpose technology. This perspective
to be closely tied to the ICT development in that country. Instead of
suggests a multivariate representation of ICT development data.
addressing the full set of issues related to the digital divide, we will
use a broad perspective to measure ICT development in this study.
2.1. Digital divide We also intend to measure ICT as a general purpose technology
(GPT).
Though more than half of the world’s inhabitants have access to
ICT, the distribution of resources has not been uniform throughout 2.2. ICT as a general purpose technology
the world. For example, there is more communication fiber in the
Asian, North American and European continents than in the African General purpose technologies (GPTs) are original ideas or tech-
continent. Even within the same continent though, there are differ- niques that have the potential to significantly influence a variety
ent levels of ICT access for different countries and regions. As ICT of industries in a country (Guerrieri and Padoan 2007). GPTs are
plays a key role in economic growth, the disparities have created characterized by their pervasiveness of use, inherent potential for
many socio-economic imbalance problems in the world. The technical improvements and innovational complementarities
phrase digital divide, in particular, has caught the attention of aca- (Bresnahanm and Trajtenberg 1995). An example is the steam engine.
demic researchers and policy-makers worldwide. The digital divide Another example is electrical systems. As GPTs improve and spread
refers to the gap between those who have access to IT and those through an economy, the economy may achieve improved produc-
who do not (Rice and Katz 2003). The OECD (2001) defined digital tivity. Since ICT is a type of GPT, measuring ICT development is a
digital divide as ‘‘the gap between individuals, households, busi- multifaceted challenge. The WSIS 2003 annual meeting provided
nesses and geographic areas at different socio-economic levels guidelines for measuring the ICT development of a country (Sciadis
with regard both to their opportunities to access information and 2005). The increasing penetration of ICT involves several critical
communication technologies and to their use of the Internet for a indices for economic growth and technology diffusion. They in-
wide variety of activities.’’ Thus, the concept of digital divide has clude mobile phone penetration, Internet penetration, PC penetra-
two key components: granularity and contents. Granularity refers tion, investment in ICT infrastructure, and so on. These indicators
to the level of entities such as individuals, businesses, countries of ICT development have been empirically tested in previous stud-
and regions where the gap occurs. Contents refer to activities that ies (Cuervo and Menendez 2006, Ho et al. 2007, Kauffman and
define the gap, for example, in terms of ICT development and use Techatassanasoontorn 2005a, UNCTAD 2010). In this study, we ar-
of the Internet. gue that the Internet and mobile phones are GPTs. We will focus on
Alleviation of the global digital divide has been a major task of these two GPTs in this study.
international organizations such as the United Nations (UN), the Technology adoption does not take place uniformly across the
World Bank, and the G8 countries (Canada, France, Germany, Italy, world. Researchers have argued that spatial proximity is likely to
Japan, Russia, the United Kingdom and the United States). These result in relational proximity when there are increased interactions
organizations have endeavored to explore how ICT impacts the (Niles and Hanson 2003). Agarwal et al. (2009) examined geo-
development of a country. They have analyzed the status quo of graphical variation in Internet use by using the spatial distribution
the development of ICT in countries, and have provided practical of individuals to define their reference group. An individual’s social
evidence year by year. In addition, various researchers have ap- or peer group refers to everyone else living in the same region – for
plied different approaches to studying the digital divide (Bélanger example, a US county. Peer effects suggest that people in the same
and Carter 2009, Cuervo and Menendez 2006, Dasgupta et al. 2001, region affect the propensity of an individual’s use of the Internet
Kauffman and Techatassanasoontorn 2005a, Sacchi et al. 2009). (Agarwal et al. 2009). Chin and Fairlie (2007) argued that income
Research on digital disparity can be divided into the study of the and the telecommunication infrastructure contribute to Internet
global digital divide (the gap between countries) and the domestic penetration. Moreover, applications of the Internet also seem to
digital divide (the gap between groups within countries). Cross- have accelerated the development of electronic commerce world-
country digital divides result from social and economic inequalities wide. The introduction of e-commerce has gradually changed the
among developed and developing countries. Some prior studies structure of global business as well (Gibbs et al. 2003). Developing
have focused on the extent of the cross-country divide (Chinn and countries can utilize the opportunities of e-commerce and ICTs to
Fairlie 2007, Crenshaw and Robinson 2006, Cuervo and Menendez increase their country’s competitiveness (UNCTAD 2001). The
2006, Dasgupta et al. 2001, Dewan et al. 2005, Dewan et al. 2010, advantages of developing ICT infrastructure are beneficial; they
Emrouznejad et al. 2010, Hanafizadeh et al. 2009, Kauffman and lead to the growth of a domestic economy and also foster economic
520 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

Table 1
Studies on the digital divide.

Authors Measurement Data/method Year/countries Variables

of digital divide
Kauffman and Digital wireless Cross-countries, 1992–1999, 43 Wealth, telecom infrastructure, market
Techatassanasoontorn technology diffusion time-series data countries competition, access cost, standards
(2005a)
Dewan et al. (2005) IT penetration Cross-countries, 1985–2001, 40 Density of main telephone line, average
(mainframes, PCs, time-series data countries monthly telephone subscription cost, average
Internet) cost of local call, size of urban population, GDP
per capita, average year of schooling, size of
trade in goods in the economy
Crenshaw and Robinson (2006) Internet Panel data analysis 1995–2000, 58 Internet hosts, telephone mainlines,
countries employment in service sector, political
openness, global urban share
Cuervo and Menendez (2006) ICT-related indicators Factor analysis and cluster 2001, 15 European Computers, main telephone lines, broadband
analysis Union countries connections, secure servers, business with a
website, business buying online, Internet dial
up access cost, households connected to the
Internet, public service online, active
population using a computer for professional
purposes
Dewan et al. (2010) PC and Internet Cross-countries, secondary 1991–2005, 26 GDP, PCs, Internet users, average PC unit price,
data analysis countries average monthly cost of telephone access
Emrouznejad et al. (2010) ICT opportunity index Data envelopment analysis 2007, 183 economies Main telephone lines, mobile cellular
subscribers, International Internet bandwidth,
adult literacy rates, gross enrolment rates,
primary secondary tertiary, Internet users,
household with a TV, computers, broadband
Internet subscribers, International outgoing
telephone traffic
Banker et al. (2011) Digital trading platform Secondary data, regression 881 transactions Raw grade indicator, premium grade indicator,
model coefficient of variation, sell transaction
indicator, click and book indicator, order book
management indicator, seller/buyer is a trader,
number of seller/buyer transaction
Talukdar and Gauri (2011) Internet access and Telephone surveys 2000 and 2008, US Internet adoption level at home, annual
usage household income, education level

development and infrastructure upgrades (Meso et al. 2009, Okoli of ICT (e.g., percentage of businesses buying online). In order to
et al. 2010). minimize the number of variables, Cuervo and Menendez (2006)
used factor analysis to extract two meaningful factors: ICT infra-
2.3. Analysis of multifaceted ICTs structure and diffusion (Factor 1), and e-government and Internet
access cost (Factor 2). The authors then applied cluster analysis
To achieve a more comprehensive understanding, we will use to group together countries with similar factor scores (UNESCO
multiple indicators to represent the ICT development of a country. 2003). Digital gaps among the fifteen EU countries were then
Many studies have attempted to capture the multifaceted nature of analyzed based on the cluster results.
ICT by using composite indicators known as indices (Fazio et al.
2000, Crenshaw and Robinson 2006, Cuervo and Menendez 3. Model development
2006). An index refers to a mathematical combination of several
primitive variables measuring different aspects of ICT develop- Researchers have been endeavoring to understand the implica-
ment; popular ICT indices include, for example, the Information tions of ICT development in various countries. This requires mea-
Society Index (IDC 1995), and the Digital Access Index (ITU sures and analysis of the relationship between ICT and national
2003). Determining the proper forms and weights with which to development. Several methodological approaches and statistical
combine primitive variables is a key challenge for these index techniques have been applied and developed to meet this need.
approaches. Our multivariate ICT development data contain four indicators
Corrocher and Ordanini (2002) provided a multivariate statis- (mobile phone, Internet, Telecommunication investment, and
tics-based method to compose an index for measuring the digital Telecommunication revenue) for each country that we will study,
divide. The authors used principal component analysis to aggre- for the study years between 2000 and 2008. To analyze global
gate elementary indicators into six digitization factors: market, ICT development trend, one can compose a representative index
diffusion, infrastructure, human resources, competition and com- from these four indicators and study the path of this index. Princi-
petitiveness. They then aggregated these six factors to obtain a pal component analysis is often used to combine multiple raw
synthetic index that can be used to gauge the digital divide. Cuervo variables into a composite index (Johnson and Wichern 2007).
and Menendez (2006) provided a factor analysis and cluster anal- We used the first principal component score to observe the layout
ysis-based methodology to analyze the digital divide among fifteen of the first 10 countries in our data set (see Fig. 1).
European Union (EU) countries. The authors collected ten ICT-re- From the ten curves, it is difficult to say whether the composite
lated indicators from 2001 that were useful for the analysis of index has converged throughout the period. If we had plotted the
the digital divide. These indicators included ICT infrastructure (e.g., curves from all 136 countries from our data set, the chart would look
number of computers per 100 inhabitants) and the pervasiveness much messier. With this short time-series data set, the definition of
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 521

3 Albania

1st Principal Component Score

2.5 Algeria
2
Argentina
1.5
Armenia
1
0.5 Australia

0 Austria

-0.5 Bahamas
-1 Bahrain
-1.5
Bangladesh
-2
2000 2001 2002 2003 2004 2005 2006 2007 2008 Barbados

Fig. 1. First principal component score for ten countries.

convergence cannot be literally applied though. The limit of each well-defined inter-cluster distance, such as the average linkage
score curve cannot be detected with only nine different points of or the complete linkage, the two closest clusters can be merged
data over time. However, we found that gaps among the countries to form a bigger cluster. This procedure is repeated until all of
narrowed from 2000 to 2008. Applying the r-convergence concept the data belong to the one and only cluster. A dendrogram showing
from Barro and Sala-i-Martin (2004), we can compute the variance the merging process is generated at the end. Application users then
of component scores in each year, and study their variations with decide where to cut the dendrogram to create different clusters.
the passage of time. We expect to see a fluctuating variance curve The cut points represent users’ preferences for inter-cluster and in-
with declining amplitudes in general over the study period. tra-cluster data similarity levels, and determine the number of
clusters. The selection of an appropriate data similarity level is
3.1. Extending r-convergence with data clustering sometimes a subjective choice. One drawback of the hierarchical
algorithm is its procedure to merge the closest clusters at each
To define the convergence of ICT development across countries, intermediate step. This type of greedy procedure may result in a
we apply the Barro and Sala-i-Martin (2004) r-convergence ap- bad dendrogram in the end. Improved algorithms using stochastic
proach to compute the variance of a cross-sectional data at each methods to merge clusters have been introduced in many studies
P
point in time. This variance ni¼1 ðwi  lÞ2 =n is based on a common (Witten and Frank 2005).
mean l for all data wi in a cross-section. Here, wi represents a gen- Unlike hierarchical algorithms, a partitioning algorithm such as
eric univariate variable with n data points. When we examine the the k-means algorithm uses representative points to segment data
curves more carefully, it seems more reasonable to assert that and aims to optimize an objective function measuring the quality
there are multiple means in the process of measuring data disper- of data segmentations. Assuming the number of clusters has been
sion. For example, we could probably assume a three-mean or determined, a k-means algorithm will start with a random set of k
four-mean state. Assume that four means l1, l2, l3, l4 are to be representative points called initial centroids g1, g2, . . . , gk. Depend-
used and each data point wi is associated with a properly selected ing on the original data points, these centroids may be multivariate
mean lj(i), where j(i) e {1, 2, 3, 4} denotes the association. j(i) as- as well. The k-means algorithm then repeats the following two
signs wi to mean lj(i) in a multiple means based variance formula. steps in sequence until some preset conditions are met:
Then, a multiple mean based formula capturing local variances may
P
be defined as v ar ¼ ni¼1 ðwi  ljðiÞ Þ2 : The association of a mean lj(i)  Partition step: Assign each data point to the cluster represented
to the data point wi is not trivial when the means themselves are to by the nearest centroid.
be determined at the same time. Using multiple means to measure  Expectation step: Update the centroid of each cluster by averag-
data variances has the advantages of accommodating local data ing data points belonging to the same cluster.
dispersion, and it should result in a smaller variance than a single
mean based formula can provide. Though simultaneous determina- These two steps are iterated to minimize a dissimilarity mea-
tion of means l1, l2, l3, l4 and the association function j(i) to make sure, indicating the quality of data segmentation. A commonly
the variance as small as possible is a computationally difficult used dissimilarity measure is defined based on the Euclidean dis-
problem, heuristic algorithms such as k-means clustering algo- tance (Han and Kamber 2007).
rithms exist that converge quickly to local minimums. Here k
stands for the number of means to be used in the partitioning of X
k X
Dissimilarity ¼ kxi  gj k2E ð1Þ
data. For example, k = 4 in the previous example.
j¼1 xi 2Pj
Cuervo and Menendez (2006) argued that data clustering pro-
vides a useful tool to handle multivariate data related to ICT devel- In the above equation, the inner summation is taken over all points
opment data. Data clustering has been widely used in data mining to xi belonging to cluster Pj represented by the centroid gj, and the
explore the implicit structure embedded in data (Han and Kamber outer summation is taken over all clusters. The Euclidean distance
2007). The goal of a clustering algorithm is to partition the data ||xi  gj||E between a data point and its cluster centroid measures
into non-overlapping subsets called clusters, such that intra-cluster the local data dispersion. This dissimilarity measure is the multivar-
data similarity is maximized while inter-cluster data similarity is iate analogy of the variance that captures local variances with
minimized. Two popular algorithms can be applied to segment multiple means. Here, i represents the index of a data point, j is
data: hierarchical clustering algorithms and partitional clustering the index of a cluster and k denotes the number of clusters in data
algorithms (Han and Kamber 2007). An agglomerative hierarchical clustering or the number of means in a multiple means based
algorithm starts with each data point as its own cluster. Using a variance formula.
522 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

It is a computationally hard task to find the centroids a totally-ordered set. After ranking all centroids in each year,
g1, g2, . . . , gk that minimize the objective function globally (Maha- we use ordinal cluster labels to represent the ICT development
jan et al. 2009). The iterative steps for the k-means algorithm can status of a country and track the movement of these ordinal
frequently converge to a local minimum. Random restarts with labels.
different centroids are proposed to alleviate this local optimum
problem. Another issue with k-means algorithms is the number 3.3. Entropy: comparing the distribution of cluster members
of clusters. Finding a way to determine the number k such that
intra-cluster dissimilarity is minimized while inter-cluster dissim- After using data clustering to segment countries into different
ilarity is maximized is still an active research topic in data mining. ICT development groups and assigning cluster labels with ordinal
Objective goals for using cross-validation principles or clustering numbers, we can gain insights into the ICT development trend of
validation indices have been proposed (Han and Kamber 2007). each country. For example, using the same ten countries in Fig. 1
In some cases, subjective opinion also matters in the determination and their ordinal cluster labels (1–4), we can plot the curve of their
of the cluster number. We will explain later how we determine this ICT development history (see Fig. 2).
critical number for our data set. The clustering is done for all 136 countries in our data set. Fig. 2
shows the results for ten countries. In each year, we partitioned the
3.2. Ranking k-means clusters 136 countries into four clusters based on their ICT relevant attri-
butes. The clustering and ranking procedure transforms the inter-
There are two advantages from using data clustering to summa- val type multivariate data into ordinal cluster labels called cluster
rize the internal data structure of ICT development. First, the scores. With this data discretization, ICT development stabilizes
technique can easily handle multivariate data sets that are quite for these countries near the end of the study period. Since the clus-
common for ICT-relevant data. Thus, we do not need to use a ter scores for each country do not change after 2006, convergence
composite index approach to analyze the data set. Second, by seems to have occurred. However, the following questions
categorizing ICT data with discrete cluster labels, we reduce the remained unanswered. Do all 136 countries have a convergent
fluctuation and noise normally associated with interval type data. ICT development path? Do they converge mostly to high cluster
For example, in Fig. 1, it is hard to analyze the trend of component scores or low cluster scores? How do we measure the dispersion
score curves. If we discretize the component scores into different of cluster scores in each year and how does this dispersion move
levels, we would have a better chance to detect the trajectory of along the time axis? We will use the concept of entropy to tackle
ICT development over the years. this issue.
By using data clustering on a cross-sectional data of multivari- Entropy is a concept in thermodynamics, but it was applied by
ate ICT indicators, we can summarize the development status of Shannon to study information theory in the late 1940s (Witten
countries with discrete cluster labels. Since we also obtained a and Frank 2005). It can be used to gauge the diversity level of mul-
time-series of these cross-sectional data, finding a way to track tiple categories in a data set. Assume that a set contains k different
the evolution of the cluster labels is also important. Centroids from types of objects and the probability of seeing each type of object is
k-means algorithms are natural choices for representing cluster p1, p2, . . . , pk. Objects refer to cluster labels, thus, using our previous
labels. If we can rank centroids, we can use the natural numbers example of 4 clusters, k = 4. Then the heterogeneous level of
1, 2, and 3, etc. as cluster labels. Using numbers as cluster labels information content of this set is computed as an entropy value:
allows us to track the ICT development status of a country over Entropy = p1 log2p1      pk log 2pk.
the study period. Unfortunately, there is no easy way to rank This entropy value reaches a maximum for log2k when the
multivariate centroids, though we can easily compare two real probability distribution on different types of objects is uniform,
numbers. i.e. p1 = p2 =    = pk = 1/k. If the distribution is skewed to one single
The non-dominated solutions concept in multi-objective type of object though, for example, p1 = 1, then it has the minimal
programming may provide a solution to our problem (Deb 2001). value of 0. The size of the entropy value is often used to measure
Using two-dimensional data as an example, we say that a point the homogeneity of the elements in a data set. The larger the entro-
(a1, b1) is dominated by another point (a2, b2) if and only if their py value, the more heterogeneous the set, and the smaller the en-
components satisfy conditions a1 6 a2 ; b1 6 b2 and at least one of tropy, the more homogeneous the set.
these inequalities is strict. In other words, the second point is as
good as the first one in every component and it must be better in 3.4. A research framework for studying trends of ICT development
at least one component. For example, (10, 20) is dominated by
(15, 20). However, it may appear that two data points are not We next develop a framework to analyze ICT development
comparable with this domination relation. For example, (10, 20) trends at the country level. This framework applies to a cross-sec-
and (15, 15) are not dominated by each other in any direction. tional time-series data set of different ICT development attributes.
Therefore, the domination relation is only a partially-ordered It consists of a cluster score assignment step and a trend analysis
relation for sets of multivariate data. Though Zorn’s (1935) lemma step. The cluster score assignment step is applied to all countries
may be helpful to deal with partially-ordered sets, we adopt a in our data set for every year. The trend analysis step may be ap-
simpler approach to make the set of centroids a totally-ordered plied to a restricted subset of countries that do not change over
set. the study period. For example, we may restrict the subjects to all
Two centroids can be compared based on the values of a chosen Asian countries or members of certain economic organizations.
component in each. For example, assume that the mobile penetra- Assuming that ICT development attributes have been collected
tion rate is a preferred variable to compare two centroids. Then, the for a fixed set of countries over successive years, for the cross-sec-
one with a higher mobile penetration rate ranks higher than the tional data in each year we apply multivariate clustering algo-
other one. In very rare cases when two centroids have exactly rithms to partition the set of countries, with each cross-section
the same mobile penetration rate, then we can use the second of data to be partitioned into the same number of clusters. A con-
preferred component to make the comparison. To maintain a sistent ranking method will be applied to the yearly clustering re-
consistency in comparison, the order of component preferences is sult to denote cluster labels as cluster scores. Using the ordinal
the same for every yearly cross-sectional data. With this definition cluster scores, we can track how the ICT development level of a
of order comparison, the multivariate centroids data set becomes country changes over the years. The cluster scores approach
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 523

Albania
4.5
Algeria
4
3.5 Argentina

Cluster labels
3 Armenia

2.5 Australia
2 Austria
1.5
Bahamas
1
Bahrain
0.5
Bangladesh
0
2000 2001 2002 2003 2004 2005 2006 2007 2008
Barbados

Fig. 2. ICT cluster labels of the ten countries in Fig. 1 during the study period. Note: In 2000, Cluster Label 1 included Albania, Algeria, Bangladesh, Cluster Label 3 included
Argentina, Armenia, Bahamas, Bahrain, Barbados, and Cluster Label 4 included Australia, Austria. In 2008, Cluster Label 1 included Algeria, Armenia, Bangladesh, Cluster Label
3 included Albania, and Cluster Label 4 includes Argentina, Australia, Austria, Bahamas, Bahrain and Barbados.

simplifies data constructs embedded in a clustering result, and it 3.5. A panel data analysis of ICT development based on a nation’s
enables us to track the ICT development trend more easily. wealth
Tracking the trend of a country is a micro-application of the
cluster scores. Let st(c) represent the cluster score of country c The previous section describes a framework for implementing
in the year t. We sort countries in our data set by the names an exploratory study of ICT development on a global basis. There
alphabetically, with Albania as country 1 (c = 1) and Zambia as is no causal factor for explaining why development has evolved.
country 136 (c = 136) in the analysis. The study period runs from ICT development is a capital-intensive development process for
2000 to 2008, thus t can be 2000 up to 2008. The trend of a coun- most countries. A country must invest its own funds in this area
try can be explored by the absolute change of the score over suc- or invite foreign parties to conduct development projects. Rich
cessive years dt(c) = |st+1(c)  st(c)|. If the cluster score st(c) countries usually have more resources at their disposal, and they
converges to a limit, that is, country c has a convergent ICT devel- often invest more money than poor countries in ICT projects. Do
opment path, then dt(c) should converge to zero as t becomes rich countries always have a high level of ICT development
large. We have seen that the ICT development status of the ten though? Is there any causal link between a country’s national
countries in Fig. 2 converged to their respective limits near the wealth and its ICT development trend? A country’s wealth may
end of the study period. be represented by its gross national income (GNI), which is avail-
At a macro-level, we also want to explore the trends of all coun- able as a time-series data set for the country. Collecting GNI
P
tries as a whole. By aggregating the individual changes Dt = cdt(c) time-series for a cross-section of countries allows us to make a pa-
from all countries, we can explore whether ICT development has nel data set. Cluster scores used to denote the ICT development sta-
converged for them as a group. In addition to this convergence tus of countries over years make another panel data set. A causal
metric, we would also like to measure how the diversity of cluster study linking GNI and ICT development may be conducted by using
scores changes over the years. Let Et denote the entropy of cluster established regression tools in panel data research. A properly
scores in year t. With these metrics, four interesting scenarios arise established panel study model can explain why trends have
as t approaches the end of the study period: happened.
Let ICTct = st(c) and GNIct respectively denote the ICT cluster
 Scenario I: Both Dt and Et go to zero as t approaches the end of score and the World Bank GNI classification of country c at year
the study period. In this case, the ICT development of most t. If we consider the cluster score as an interval-type data, we
countries will converge to a common level. can regress ICTct against GNIct with fixed-effects or random-effects
 Scenario II: Dt goes to zero, and Et remains large. The ICT devel- linear models as follows:
opment of most countries will converge, and the convergence
will be uniform on every level. Fixed effects : ICT ct ¼ b0 þ b1 GNIct þ a2 C 2 þ a3 C 3 þ   
 Scenario III: Dt remains large, and Et goes to zero. This indicates þ aN C N þ c2 T2 þ c3 T 3 þ    þ cM T M þ uct
that cluster scores will be concentrated at a single level near
ð2Þ
the end of the study period, and some countries will not have
a convergent ICT development path. A possible cause for this
Random effects : ICT ct ¼ b0 þ b1 GNIct þ uct ð3Þ
could be a large fluctuation of cluster scores for many coun-
tries. For example, some countries may belong to the low where C2, C3, . . ., and CN are dummy variables for the country effects,
ICT level in a year, and then move to the high ICT level in T2, T3, . . ., and TM are dummy variables for the time effects, and uit is
the following year. the idiosyncratic error. A dummy variable takes the value of 0 or 1
 Scenario IV: Both Dt and Et remain large as t approaches the end only, and is assumed 0 unless its special fixed effect is being consid-
of the study period. In this case, some countries do not have a ered. For example, C2 = 1 only when we are considering the country
convergent ICT development path, and cluster scores are likely fixed effect of Country 2, and T2 = 1 only when we are considering
to be distributed uniformly on every level. A possible cause the time fixed effect of Year 2 in our study. The panel data contain
for this phenomenon might be that two nearby cluster levels N countries and M years. (In this study, Country 2 (c = 2) is Algeria
swap their members in successive years. and Year 2 is 2001 (t = 2001); also, N = 136 and M = 9.) The first
 All other cases are intermediate cases that fall between the four country and the first year are used as the base to compute country
extreme cases. Our data will show the actual pattern of the or time fixed effects, thus their dummy variables are omitted in the
countries collected for this study. fixed effects model. Using the Hausman (1978) test, we can
524 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

determine whether the random-effects model is consistent and moved from a low GNI status to a lower middle GNI status in
should be preferred over the fixed-effects model. 2002. The United States has kept its high GNI level throughout the
On the other hand, if we consider the cluster score to be an ordi- entire period.
nal number, then we can use an ordered probit or logit regression
model to estimate the expected value of ICT given GNI with an 4.2. Data description
appropriate link function F (see Appendix B).
The raw data in Table 2 were further normalized to account for
EðICT ct jGNIct Þ ¼ Fðb  GNIct þ a2 C 2 þ a3 C 3 þ    þ aN C N þ c2 T 2 size differences among countries. We divided national mobile sub-
þ c3 T 3 þ    þ cM T M Þ ð4Þ scription number by the population of a country to get the mobile
penetration variable. We obtained the Internet user penetration by
A satisfactory panel regression model with fixed country or year dividing the number of Internet users in each country by the pop-
effects may help us explain why a shift in the ICT development ulation. We adjusted two financial indicators (capital investment
trend has occurred for a country or a given year. This type of casual and telecom revenue) according to the GDP of each country.
model complements our exploratory study by attempting to Descriptive statistics of these new variables for the end years are
discover what the changes are while reasoning why changes may shown in Table 3.
be happening. From the descriptive statistics, we see that the two financial
indicators are several orders smaller than the other two penetra-
4. Data preprocessing and cluster score computation tion variables. Since data clustering algorithms using Euclidean
distance to measure dissimilarity level may ignore the impact of
To investigate global ICT development, we collected secondary variables with less strength, we further normalized each variable
data published by international organizations. The data collection by its z-score transformation z = (x  l)/r. Here l and r are the
procedure involved some tradeoffs between the breadth and depth sample mean and sample standard deviation of a variable. We per-
of data. When it comes to data breadth, the concern is with the formed this normalization procedure for the cross-sectional data of
number of countries. In contrast, data depth refers to the number every year so that each variable will have a sample mean of 0 and a
of variables associated with each country. Since our focus is on a sample standard deviation of 1. For example, a z = (x  .1886)/
global study, we decided to place more emphasis on breadth than .2479 transformation is applied to the mobile penetration variable
depth. Overall, we collected the four ICT relevant variables for 136 for 2000, and for 2008 the formula is z = (x  .8024)/.4419. The fi-
countries from the Year 2000 to the Year 2008. nal variables used in this study are shown in Table 4.

4.1. Data collection 4.3. Cluster score computation

To analyze a large sample of 136 countries, we selected four According to the proposed research framework, we need to
variables to represent the ICT development of each country. We ar- compute a cluster score for every country and every year. The clus-
gued that mobile phones and the Internet are ICT-related general ter score assignment begins with a data clustering procedure to
purpose technologies (GPTs). Capital investment in telecommuni- cluster the multivariate ICT variables (MP, IUP, INV, and REV) in each
cations and total telecommunication revenues represent the year. We chose k-means algorithms rather than hierarchical algo-
overall ICT market in a country. Table 2 shows the definitions of rithms for the following reasons:
the variables and sources of our data.
We do not claim that the above variables are exhaustive, but  A hierarchical algorithm needs to make a subjective selection of
we believe they represent a country’s ICT development to some the cut point on its dendrogram to determine the number of
extent. The first two variables measure the ICT infrastructure clusters. This may be difficult. Also, its greedy procedure for
(info-use of WSIS 2003), while the last two variables measure merging the closest clusters at every intermediate step may
the financial results of ICT development (infodensity of WSIS not produce a good dendrogram.
2003).  The objective function used by a k-means algorithm is a natural
0In addition to the ICT-related variables, we also collected the extension of the concept of r-convergence in Barro and Sala-i-
gross national income (GNI) classification from the World Bank Martin (2004). In addition, centroids from the algorithm are
(2009) from the year 2000 to 2008. Each year, the World Bank deter- good representatives for the clusters, and they can be ranked
mines boundary points to stratify countries into four categories. with an appropriate procedure. Ranking clusters allows us to
They include: high GNI, upper middle GNI, lower middle GNI, and convert cluster labels into ordinal numbers as cluster scores,
low GNI. High GNI countries typically include developed countries and also enables the analyses that follow.
such as Austria, Japan and the U.S., and low GNI countries include
developing countries such as Bangladesh, Gambia and Zambia. There are some issues associated with the k-means algorithm
Though a country’s GNI category may change from year to year, though. For example, how do we stop the iterative steps and find
the changes are infrequent. For example, during the period from adequate centroids? We performed the partition-expectation cycle
2000 to 2008, Armenia changed its status only once: the country 100 times for each run. In most cases, the centroids stopped

Table 2
Definitions of variables and data sources.

Variables Definition Sources

Mobile phone penetration Subscriptions to a public mobile telephone service ITU, Euromonitor Intl
Internet user Number of total fixed-line Internet subscriptions Euromonitor Intl
Capital investment in telecom Gross annual investment in telecom including fixed, mobile and Internet services Euromonitor Intl
Total telecom revenue Gross annual telecom revenue earned from fixed, mobile and Internet services Euromonitor Intl
Gross national income (GNI) per capita Economies classified as low, middle (subdivided into lower middle, upper middle), or high income World Bank (2009)

Note: We converted values to international dollars in capital investment for telecom/GDP and total telecom revenue/GDP to avoid the variation in purchasing power.
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 525

Table 3
Descriptive statistics for 2000 and 2008.

Variables Min Max Mean Std. dev.

2000 2008 2000 2008 2000 2008 2000 2008
Mobile penetration .0002 .0074 .8173 2.0781 .1886 .8024 .2479 .4419
Internet user penetration .000002 .0040 .4802 .8710 .0897 .3012 .1337 .2681
Capital investment in telecom/GDP .000012 .000023 .0276 .0392 .0082 .0069 .0059 .0075
Total telecom revenue/GDP .0010 .0009 .0695 .1192 .0278 .0319 .0141 .0193

Table 4 The FPC accounted for about 44.6–49.5% of the data variance in
List of final variables.
every year. Using r-convergence, we computed the standard devi-
Variables Description Use ation of the FPC scores to measure the extent of data dispersion
GNI Gross national income Panel data (see Fig. 3).
analysis The standard deviation declined initially, increased in 2002–
a
Mobile penetration (MP) Mobile subscription/ Clustering 2004 and continued to diminish until 2008. If we ignore the minor
population
fluctuations from 2002 to 2004, this seems to indicate that global
Internet user penetration (IUP)a Internet subscription/ Clustering
population ICT development tended to achieve r-convergence in the end.
Capital investment in telecom Investment/GDP Clustering Thus, the FPC scores of all countries become less dispersed from
a
(INV) their common mean 0 as a whole; but the result does not tell us
Total telecom revenue (REV)a Telecom revenue/GDP Clustering
whether a country has convergent FPC scores. (The mean of all
a
z-Score of the variables in the description field. FPC scores in a year is 0.) We have seen in Fig. 1 that it is difficult
to check the convergence of the FPC scores of a country because of
the complexity involved in interval type scores. Although we can
moving before the 100th cycle, so this is a good termination crite- apply the research framework to cluster the univariate variable
rion. To guarantee good final centroids, we ran the algorithm 100 (FPC score) and discretize each country’s ICT development status
times using different initial centroids in each run. The run with with its cluster score, we did not do this because the FPC accounted
the minimum total dissimilarity as described in Eq. (1) produced for less than 50% of the total variance in each year. Adding the sec-
the final data clusters for a year in our procedure. ond principal component score will increase the explanatory
A more critical problem related to the k-means algorithm is the power to more than 50% of total variance, but it also creates diffi-
determination of the parameter k, the number of intended clusters. culties associated with a multivariate trend analysis.
Several cluster validation indices have been discussed in literature.
Among these indices, the Davis and Bouldin 1979) index measures 5.2. Using multivariate ICT development indicators
the intra-to-inter clustering effect. This index is defined as:
The cross-sectional multivariate data (MP, IUP, INV, REV) are
1X k
partitioned into a proper number of clusters in each year. We first
DB ¼ Dj
k j¼1 use a cross-validation approach to determine the number of clus-
ters. The open-source software package Weka 3.6 has an expecta-
where Dj ¼ maxj–j0 fDj;j0 g and Dj;j0 is the within-to-between cluster tion–maximization clustering algorithm that can help determine
spread for the involved clusters j and j0 . That is Dj;j0 ¼ ðdj þ the number of clusters (Witten and Frank 2005). Three to five clus-
dj0 Þ=distj;j0 , where dj and dj0 are the average within-cluster distances ters are appropriate for yearly cross-sections from 2000 to 2008. So
of clusters j and j0 , and distj;j0 is the distance between the centroids we decided to try a k-means algorithm with three to five centroids.
of two involved clusters. (In other words, dj measures the intra-clus- The total dissimilarity and Davis–Bouldin index for each cross-sec-
ter dissimilarity of a cluster and distj;j0 measures the inter-cluster tional data are shown in Table 5.
dissimilarity between two clusters.) Since our goal is to minimize For each year, the total dissimilarity decreases as the number
the intra-cluster dissimilarity and maximize the inter-cluster of centroids increases. So the more centroids we have, the higher
dissimilarity at the same time, we prefer to choose a value of k with the within-cluster similarity can be obtained. However, the be-
a smaller Davis–Bouldin index. tween-cluster dissimilarity can become smaller as well. This is
the reason why the Davis–Bouldin index includes a between-
5. Data analysis cluster distance (dist j;j0 ) in its formula. After calculating the Davis–
Bouldin index for each year, we see that the five-cluster partition
The first principal component (FPC) score is initially used as a has a smaller index only for 2001. Though it also has a minimum
composite index approach to explore the trend of global ICT devel- value of .86 in 2002, unfortunately, one of the clusters contains
opment. Then, the results of cluster scores approach are presented. only one country. Thus, we viewed the next smaller index (.93)
The cluster scores approach is also applied to a fixed subset of as the preferred partition for this year. Bold-faced numbers in
countries (OECD members) as an example to check the viability the table indicate the preferred number of clusters according to
of the approach. Finally, panel data analysis is presented to assess the Davis–Bouldin index.
whether there is a causal relationship between national wealth and According to the Davis–Bouldin index, a four-cluster partition is
ICT development status. Peer effects from spatial proximity are preferable for the five yearly cross-sectional data sets (2002, 2003,
also presented as a modified panel regression model. 2006, 2007 and 2008), while a three-cluster partition is appropriate
for the three yearly data sets (2000, 2004 and 2005). In contrast, a
5.1. A composite index approach with FPC scores five-cluster partition makes sense for the cross-sectional data set
of 2001. There are two reasons for not using different numbers of
The FPC score from the four ICT variables (MP, IUP, INV and REV) clusters for different years. First, this practice creates difficulties
was computed for every country in each year from 2000 to 2008. in tracking cluster movements from year to year. For example, if
526 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

1.42

Standard deviation
1.4
1.38
1.36
1.34
1.32
1.3
2000 2001 2002 2003 2004 2005 2006 2007 2008

Fig. 3. Standard deviation of the first principal component scores in each cross-section.

Table 5
Total dissimilarity and Davis–Bouldin index.

Clusters 2000 2001 2002 2003 2004 2005 2006 2007 2008
Total dissimilarity 3 189.88 197.91 220.25 208.20 209.77 205.21 214.13 225.92 239.48
4 154.27 160.11 166.24 163.69 182.41 171.74 172.89 180.93 190.55
5 130.51 131.84 142.06 134.88 157.00 146.59 143.01 144.21 151.68
Davis–Bouldin index 3 0.96 1.01 1.13 0.98 0.99 1.00 1.02 1.07 1.07
4 1.05 1.14 0.93 0.86 1.10 1.08 0.98 0.97 0.98
5 1.01 1.00 0.86 0.99 1.18 1.12 1.05 1.05 1.07

2.5

2
D & M metrics

1.5 D
E
1

0.5

0
2000 2001 2002 2003 2004 2005 2006 2007 2008

Fig. 4. Standardized convergence metric Dt and entropy metric Et for all countries.

we use four clusters for 2003 and three clusters for 2004, there will Accordingly, cluster scores for the US have converged, and so
be a problem of tracking changes in the four clusters in 2003 to the have the scores for Armenia. To measure the global trend, we use
three clusters in 2004. Second, the maximum entropy value of clus- the aggregate cluster score changes Dt introduced before. To min-
ter scores in a year depends on the number of clusters. In a year imize the effects of big jumps (e.g., from 1 to 4), we refine our pre-
with three clusters, this maximum value is log23, and the maximum vious definition of the metric as follows:
value will be log24 for a year with four clusters. Due to these rea-
sons, we preferred to choose the same number of clusters for each X
136
Dt ¼ dt ðcÞ; dt ðcÞ ¼ Minð1; jstþ1 ðcÞ  st ðcÞjÞ;
year. Since a four-cluster partition had more years with a preferred c¼1
Davis–Bouldin index, we used a four-cluster partition for every
t ¼ 2000; . . . ; 2008
year. Last, a subjective reason to partition countries into four differ-
ent ICT development clusters is because the World Bank practices a In other words, regardless of the extent of the gaps, we only count
four-stratum classification of GNI. how many times a country has changed its ICT status over the study
After ranking the centroids, we may label them as high, upper period. The entropy value for measuring the diversity of cluster
middle, lower middle and low ICT development centroids like the scores in a year is computed as:
World Bank has done for the GNI classification. The correspond-
ing clusters are marked with ordinal numbers of 4, 3, 2 and 1. X
4
Et ¼ pt;s log2 ðpt;s Þ; t ¼ 2000; . . . ; 2008
Using cluster scores to denote the ICT development status of a
s¼1
country, we can obtain useful insights about its ICT development
path over time. For example, cluster scores of the US for the years where pt,s is the proportion of cluster scores s e {1, 2, 3, 4} in year t.
from 2000 to 2008 are all 4s, meaning the US belonged to the Using these two metrics, we plot the changes over years in Fig. 4.
high ICT development cluster during the whole study period. The convergence metric Dj has been normalized to the range (0, 1)
Cluster scores for Armenia over the same period are 3, 2, 1, 1, by dividing each number by 92. The largest migration of cluster
2, 1, 1, 1 and 1, indicating that Armenia belonged to the upper scores occurred from 2003 to 2004. The smallest number of score
middle cluster in 2000, the lower middle cluster in 2001 and changes, three, appeared in the period from 2006 to 2007. The per-
2004, and the low cluster for all other years. Notice that the clus- iod from 2007 to 2008 had five score changes. Cluster scores migra-
ter score st(c) for country c in year t is a relative measure for that tion occurred before 2006 but slowed down after that year. Near the
year only, because we conducted the data clustering procedure end of the study period, most countries’ ICT development levels
independently every year. Thus, a country may have had cluster stabilized.
scores s2000(c) = 3 and s2001(c) = 2 when it could not keep pace How these countries development levels converged is another
with the global ICT development in 2001. story. With four categories or clusters in a data set, the maximum
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 527

entropy value is 2. In 2001, the entropy value was close to 2, thus variables (INV and REV) had similar trends of movement to one an-
the cluster scores distribution should be pretty uniform for that other. Fig. 7 shows the time series of the INV variable of all cen-
year. Unfortunately, we did not see an entropy (Et) curve that goes troids. The capital investment rate moved up and down over the
to zero like the convergence (Dt) curve. Thus, we cannot conclude years, indicating no clear trend.
that countries converged to the same ICT development level.
Charting the cluster scores distribution in Fig. 5, we can observe 5.3. ICT development trend for OECD countries
the trend of each cluster over the years. In 2000, more than half of
the countries (70 to be exact) were in the lower middle and low ICT With the ICT development cluster scores at our disposal, we are
development clusters. Later, more than half of the countries (70) ready to explore the ICT development trends for a subset of
were in the upper middle or high clusters in 2008. The number countries. Out of the 136 countries in our data, there are 29 OECD
of countries in the ICT low cluster stayed about the same in 2000 members (Australia, Austria, Belgium, Canada, Czech Republic,
and 2008, but the number in the ICT high cluster grew. This growth Denmark, Finland, France, Germany, Greece, Hungary, Iceland, Ire-
came at the cost of the upper middle cluster. land, Italy, Japan, Luxembourg, Mexico, Netherlands, New Zealand,
Fig. 6 shows the movement of centroids in the mobile phone Norway, Poland, Portugal, South Korea, Spain, Sweden, Switzer-
penetration (MP) variable. The mobile phone penetration rate land, Turkey, United Kingdom, and the United States). During our
increased for all clusters over the years. The lower three clusters study period of 2000–2008, there were no OECD membership
(upper middle, lower middle, and low) grew together, while the changes in our data set. Thus, this group of countries can serve
high cluster maintained a substantial lead during the study period. as a pivotal example to examine our proposed framework.
Even though the lower clusters had a higher rate of growth, it may Fig. 8 shows the convergence (Dt) and entropy (Et) curves of the
take them many more years to catch up to the level enjoyed by the OECD group over the study period. The Dt curve was normalized by
leading countries. The growth trend of the Internet user dividing the raw number of cluster score changes by 6; this is the
penetration (IUP) variable is similar. Both telecom financial-related largest change in cluster scores, which happened from 2000 to

80
70
Number of countries

60
50 4
40 3
30 2

20 1

10
0
2000 2001 2002 2003 2004 2005 2006 2007 2008
(Note: 4= high, 3=upper middle, 2= lower middle, and 1= low ICT development status)

Fig. 5. Number of countries in each cluster (Note: 4 = high, 3 = upper middle, 2 = lower middle, and 1 = low ICT development status).

1.4
1.2
MP variable

1
0.8 4
0.6 3
2
0.4
1
0.2
0
2000 2001 2002 2003 2004 2005 2006 2007 2008
Note: 4= high, 3=upper middle, 2= lower middle, and 1= low ICT development status

Fig. 6. The MP (mobile penetration) variable of four centroids. Note: 4 = high, 3 = upper middle, 2 = lower middle, and 1 = low ICT development status.

0.06

0.05
INV variable

0.04

0.03 4
3
0.02 2
0.01 1

0
2000 2001 2002 2003 2004 2005 2006 2007 2008
Note: 4= high, 3=upper middle, 2= lower middle, and 1= low ICT development status.

Fig. 7. The INV (Capital Investment in Telecom) variable of four centroids. Note: 4 = high, 3 = upper middle, 2 = lower middle, and 1 = low ICT development status.
528 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

1.2

D & E metrics
0.8
D
0.6
E
0.4

0.2

0
2000 2001 2002 2003 2004 2005 2006 2007 2008

Fig. 8. Standardized convergence metric Dt and entropy metric Et of 29 OECD countries.

2001. Because Dt was zero in 2007 and 2008, we see that the ICT Considering only significant factors at the p < .05 level in the or-
development of OECD members converged. The entropy value dered probit model (Table 8), we have a regression model as fol-
has fallen from 0.994 (2000) to 0.575 (2008). Cluster scores appear lows (see Eq. (4) and Appendix B).
in different primary clusters. Indeed, most OECD countries con-
verged to the high ICT development status at the end of the study EðICT ct jGNIct Þ ¼ Fðfct Þ ¼ 4  Uð2:113  fct Þ  Uð3:025  fct Þ
period. In 2008, there were 26 members in the high ICT cluster, one  Uð4:264  fct Þ
member (Canada) in the upper middle cluster and two members fct ¼ 1:241GNIct  :430T 4 þ 1:028T 5
(Mexico and Turkey) in the low ICT cluster. Z u pffiffiffiffiffiffiffi 2
UðuÞ ¼ ð1= 2pÞeðw =2Þ dw
5.4. Panel data study of GNI and ICT causal effects 1

The GNI factor has a significant positive impact on the ICT status of
Up to this point, we have explored the trends in global ICT a country. A higher GNI value increases the expected value of ICT.
development, but we have not studied why they happened. As Thus, wealthier countries are expected to have a higher ICT
we argued before, a nation’s wealth may be an important factor development level. In 2000, out of 35 GNI high countries, 27 of them
affecting its ICT development trend, a panel data analysis using belonged to the ICT high cluster (ratio = .77), and in 2008, 42 of the
GNI and ICT cluster scores as the cause and the effect was con- 44 GNI high countries were in the ICT high cluster (ratio = .95). The
ducted next. We used the software package STATA 9.2 for this pur- ratios of GNI upper middle countries belonging to the ICT high clus-
pose. First, we considered simple linear regression models with ter were .08 and .30 in 2000 and 2008.
random effects and fixed effects (see Table 6). There is a significant negative impact on ICT in 2003 and a sig-
According to Hsiao (2003), the fixed-effects model is always nificant positive impact in 2004. The estimation model for 2003 is
consistent, but may not be as efficient as the random-effects mod- fc,2003 = 1.241GNIc,2003  .430. Thus, countries migrated to a lower
el. On the other hand, the random-effects model is efficient, but ICT level unless they had a higher GNI value to counteract the
may not be consistent because it may fail to recognize correlations specific year fixed-effect. On the other hand, the estimation model
between regressors and idiosyncratic errors. The Hausman test for 2004 is fc,2004 = 1.241GNIc,2004 + 1.028, reflecting a reverse
shows that there is a significant difference between the fixed-ef- trend of global ICT development in that year (see Fig. 5). Little
fects coefficients and the random effects coefficients (p = .000). change in the cluster scores distribution was observed after
Thus, the fixed-effects model should be used to model our panel 2004. This can be explained by the estimation model fct =
data. After using dummy variables for countries and years, we 1.241GNIct, t = 2005, 2006, 2007, 2008, and the relatively stable
ran the linear regression model in Eq. (2) again with the results re- GNI classification.
ported in Table 7. Many countries have a significant fixed effect
that controls for the specific time-invariant environment of a coun-
5.5. Panel data models with region fixed-effects
try. For example, the relatively stable human capital of a country is
part of the country fixed effect that has an important impact on its
If we want to study whether the ICT development of countries
ICT development. Although model fitness has increased substan-
in a geographical region has converged, we can apply procedures
tially (R2 = .802), the fixed effects of country and year together ren-
used for the investigation of OECD countries. For example, we
der our main regressor (GNI) insignificant (p = .056) with the
may ask whether Asian counties, as a whole, have convergent ICT
significance level of a = .05. The year dummy variable for 2003
development paths by considering the convergence metric Dt and
has a significant negative impact, which may explain why many
the entropy metric Et. On the other hand, if we want to investigate
countries shifted to the low cluster in 2003 (see Fig. 5).
how a specific geographical region may impact the ICT develop-
Although linear models provide useful insights into the rela-
ment of a country, we can use a panel data model by considering
tionship between regressors and output, our cluster scores are
region fixed-effects. The United Nations defines the world in terms
actually ordinal numbers. Thus, it is more natural to use panel data
of five geographical regions. The region of a country is fixed. To see
analysis with an ordered categorical outcome variable. The above
region fixed-effects on ICT development, we introduce a time-
analysis shows that fixed-effects models are more appropriate for
invariant dummy variable Regionk for each region (see Table 9,
our data. Thus, we conducted an ordered probit analysis with fixed
Eq. (4) and Appendix B).
effects using a STATA add-on package (reoprob) (Frechette 2001).
Since a full ordered probit model with both the country and year EðICT ct jGNIct Þ ¼ Fðb  GNIct þ a2 Region009 þ a3 Region019
fixed-effects failed to converge, we ran the analysis with the year
þ a4 Region142 þ a5 Region150 þ uct Þ
fixed-effects only. Table 8 shows that the main regressor GNI has
a significant positive impact on the dependent variable and the The ordered probit model with Region fixed effects is reported in
dummy variable for 2003 has a significant negative impact. Table 10.
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 529

Table 6
Panel data GNI-ICT regressions with a linear model.

Fixed-effects regression Random-effects regression

Estimate (Std. dev.) p Estimate (Std. dev.) p
GNI .240 (.076) .002 .620 (.043) .000
Constant 1.80 (.190) .000 .859 (.122) .000
R2 = .496, model sig. = .002 R2 = .496, model sig. = .000

Note: Standard statistical test notations are used: R2 denotes the coefficient of determination and p denotes the probability of obtaining a test
statistic at least as extreme as the one actually observed, assuming that the null hypothesis is true. The popular significance level a = .05 was used
in the study.

Table 7
Panel data linear model with fixed effects.

t GNI 2001 2002 2003 2004 2005 2006 2007 2008 Constant
Estimated .153 .068 .108 .180 .404 .045 .192 .162 .175 2.031
(Std. dev.) (.080) (.073) (.073) (.073) (.073) (.073) (.074) (.075) (.076) (.258)
P .056 .350 .140 .014 .000 .537 .009 .030 .021 .000

Note: R2 = .802, model sig. = .000. Country effects are not shown.

Table 8
Ordered probit model with year fixed-effects.

t GNI 2001 2002 2003 2004 2005 2006 2007 2008 cut1 cut2 cut3 q
Estimated 1.241 .242 .180 .430 1.028 .038 .293 .171 .143 2.113 3.025 4.264 .627
(Std. dev.) (.072) (.164) (.169) (.176) (.161) (.171) (.172) (.173) (.172) (.226) (.231) (.248) (.031)
P .000 .139 .287 .015 .000 .826 .088 .321 .404 .000 .000 .000 .000

Note: Log likelihood = 962.20, model sig. = .000.

Table 9 utilizations of national wealth in ICT development for countries

Geographical regions according to the United Nations. of Americas. In contrast, countries in Europe managed to improve
UN code Region Examples of countries Number of countries their ICT development by taking advantages of positive region
in our study fixed-effects during the study period.
002 Africa Algeria, Egypt, South Africa 36
009 Oceania Austria, Tonga 6 5.6. Peer effects from spatial proximity
019 Americas Barbados, Peru, US 23
142 Asia China, Japan, Taiwan 38
Do countries with spatial proximity tend to have similar ICT
150 Europe France, Germany, UK 33
development paths? A cross-sectional study by Agarwal et al.
(2009) demonstrated the peer effects of the Internet: people in
It appears that only Region019 (the Americas) and Region150 the same region affect the propensity of Internet use by other
(Europe) have significant region fixed-effects. The base region is individuals. Are these peer effects applicable on a country-wide
Africa, and so its fixed effect is not omitted. A negative region effect level? The last section of region fixed-effects partially investigates
was observed for the Americas region, while a positive effect was this issue. We should keep in mind that the region fixed-effects are
observed for the Europe region. Thus, if a country is in the time invariant. In this section, we study how the ICT development
Americas, then it had a negative impact on its ICT development of peer countries in the same UN defined region as a focal country
during the study period. This seems to contradict the common may impact the focal country’s ICT development during the study
sense that countries in Americas generally have a better ICT period. To examine peer effects, we need to create a variable to
development than countries in Africa. The estimation model aggregate the ICT levels of peer countries in a region. Peer effects
fct = 1.01GNIct  .544Region019 + 1.636Region150 involves the GNI are possible causes that may increase the likelihood of a particular
factor, so we need to interpret the result more carefully. The GNI action based on the incidence of actions by peers (Agarwal et al.
classifications, according to the World Bank, differ substantially 2009). Unlike social norms that are frequently perceived
for countries of different regions. For example, in 2000, the average subjectively, peer effects can be more objectively measured. Agar-
GNI level for Americas was 2.61, while the average GNI level for wal et al. (2009) used the average Internet adoption rate of peer
Africa was only 1.47 in the same year. Thus, the negative impact respondents residing in the same county as the focal respondent
of the dummy variable Region019 can be interpreted as inefficient

Table 10
Ordered probit model for region fixed-effects.

GNI Region009 Region019 Region142 Region150 cut1 cut2 cut3 q

Estimated 1.010 .100 .544 .048 1.636 1.731 2.591 3.800 .536
(Std. dev.) (.089) (.631) (.272) (.302) (.316) (.187) (.193) (.215) (.047)
P .000 .874 .046 .875 .000 .000 .000 .000 .000

Note: Log likelihood = 987.19, model sig. = .000.

530 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

to represent peer effects. Thus, peer effects are not measured by optimistic to see that the key indicator MP (mobile penetration
the perception of the focal respondent. rate) increased for all four clusters; but the unstable capital invest-
To calculate peer effects, we define ICTPEct as the average ICT ments in telecom (INV) for the lower two levels makes the goal of
cluster score of peer countries residing in the same U.N. defined re- eliminating worldwide digital divides questionable. These results
gion as the focal country c in year t. Similar to Agarwal et al. (2009), imply that the digital divide is alleviated somewhat if we measure
this peer influence is individualized within a region and updated it from the mobile phone penetration rate but capital investments
for every year: in telecom cannot tell us an absolute answer. Since telecom invest-
X ments of a country usually allocate resources in the infrastructure,
ICTPEct ¼ ICT mt =ðjRj  1Þ
the consequent development from the ICT infrastructure can only
m2Rfcg
been realized after a few years.
where R denotes the set of countries in the region from Table 9 that After exploring the trend in global ICT development, we then
contains country c, |R| is the number of countries in the region, and analyzed the relationship between a nation’s wealth and its ICT
m e R  {c} indicates all other countries in the region except c. The development. Research has proved that these national wealth ef-
ICT cluster score is regressed against the main regressor GNI and fects are significant in a country’s ICT development. Thus, we used
the peer effects variable ICTPE. Table 11 shows that the peer effects a panel data model to analyze the causal effect of GNI on the ICT
variable is significant. Thus, peer effects are observed in our data. development of a country. We found that GNI had a significant po-
sitive impact on the ICT cluster score through ordered probit
6. Discussion regression model. That is, countries with higher national income
tended to enjoy a higher ICT development status. This finding is
ICTs have been viewed as impacting the economic growth of a consistent with prior research.
country and improving the quality of life of the poor. During the Our second question asked whether countries with different
past decades, countries have increasingly invested in their ICT GNI levels have different ICT development paths. Based on the re-
infrastructure to enhance development. How ICT development var- sults from our panel data modeling, the answer is affirmative in
ies across countries becomes a critical indicator for global digital general. For example, Albania and Argentina had GNI level se-
divide. It is critical to investigate the trends of global ICT develop- quences of 2, 2, 2, 2, 2, 2, 2, 2, 2 and 3, 3, 3, 3, 3, 3, 3, 3, 3, and ICT devel-
ment for better resource allocation at the national level. opment paths were 1, 2, 3, 2, 3, 2, 3, 3, 3 and 3, 2, 3, 1, 2, 1, 4, 4, 4 during
Our first research question was to explore global ICT develop- our study period from 2000 to 2008. This example shows the posi-
ment trends in the last decade. To answer the question, we use tive impact of GNI on ICT. In contrast, Armenia had a GNI level se-
data clustering and centroids ranking to summarize the ICT devel- quence of 1, 1, 2, 2, 2, 2, 2, 2, 2, similar to that of Albania, but its ICT
opment status of a country into an ordinal number called a cluster development path was 3, 2, 1, 1, 2, 1, 1, 1, 1, which is quite different
score. This idea is similar to the World Bank’s classification of a na- from Albania’s. This discrepancy can only be explained by a coun-
tion’s wealth (GNI) in four strata. Differences in cluster scores try fixed-effect resulting from Armenia. Because of a convergence
implicitly define digital divides between two countries. By plotting issue in model estimation, the ordered probit model was not run
cluster scores of a country from different years, we can detect the with country fixed-effects. With a simple linear regression model,
trend of the country’s ICT development. We apply a data clustering there was a significant and negative country fixed-effect of .947
approach to demonstrate that each of the listed countries in Fig. 1 (p = .001) for Armenia.
had a convergent ICT development path over the study period. The To sum up, GNI always shows a positive impact on ICT develop-
cluster scores curve stabilized after 2006 for each country. In addi- ment, and thus countries of higher GNI levels tended to have a
tion, by aggregating cluster score changes from all countries into a higher ICT development path. Individual country fixed-effects
convergence metric Dt, we can examine the convergence trend of may also play a role in the ICT development path of a country. Cur-
the global ICT development. Most countries had a convergent ICT rently, these country fixed-effects can only be examined through
development path during the study period. Using the entropy met- simple linear regression models. The ordered probit model did
ric Et, we see that the convergence happened at non-unique ICT show year fixed-effects that match our observation of ICT develop-
development levels. Indeed, our cluster scores approach shows ment trends with cluster scores.
that there were two heavy clusters (high and low clusters) near Finally, our third research question on peer effects investigated
the end of the study period, and the high ICT development cluster the social influence theory at the entity level of countries. The spa-
grew at the expense of the upper middle ICT development cluster tial proximity was determined with respect to the UN defined geo-
in the last decade. graphical regions. Two countries are considered peers if they
Since a country’s ICT development level changed from year to belong to the same region. Before considering the behavior in
year, we cannot say that ICT investments were ineffective for those terms of ICT development status of peer countries, we first investi-
countries initially belonging to the lower two levels (lower middle gated region fixed effects, which were a natural extension of coun-
and low) in 2000. Some of these countries have moved to the upper try fixed-effects. We found that region fixed-effects did exist. For
two levels through well-designed development projects. Using the example, Americas had a negative region fixed-effect on ICT devel-
scores distribution, we also see that about half of the countries opment, while Europe had a positive region fixed-effect. The Amer-
were in the upper two levels and the other half in the lower two icas regional effect was negative with respect to its members’ GNI
levels from 2000 to 2008. Thus, the severity of global digital divides contributions on ICT development, and the Europe regional effect
was not reduced from this simple count of cluster members. It is was the opposite. What caused these regional effects remains an
issue to be studied further.
Table 11
Following the idea of social influence from Agarwal et al. (2009),
Ordered probit model for peer effects. we constructed a peer influence variable to measure peer effects
based on regional spatial proximity. This peer influence variable
GNI ICTPE cut1 cut2 cut3 q
averages the ICT development cluster scores of peers in the same
Estimated .837 1.349 4.127 5.037 6.327 .644 region. We expected higher peer effects to bring a higher ICT devel-
(Std. dev.) (.099) (.117) (.289) (.297) (.325) (.029)
opment path to a country, due to social influences from spatial
p .000 .000 .000 .000 .000 .000
peers. Countries in the same regions may have more interactions
Note: Log likelihood = 956.20, model sig. = .000. in terms of international trade, culture exchange activities, or
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 531

business transactions. The higher national interactions among to a pattern of results that allow us to learn more about the empir-
countries may lead to higher social learning with respect to ical patterns from our collected data.
government policies and investments. Our findings validate this Second, our empirical contributions are that we empirically
expectation by exhibiting a positive and significant impact of the examined the trend of ICT development of 136 countries from
peer effects. Thus social influence theory has also been validated 2000 to 2008. We also examined the convergence and peer effects
at the country level for the behavior of ICT development. of ICT development across countries. These empirical results pro-
vide practical insights for government policy makers as well as
7. Conclusions the international organizations, such as the United Nations and
the OECD. Governments and international organizations can allo-
The rapid development of ICT has diminished the socio-eco- cate government and international financial resources and funding
nomic gap between countries. Although it is too early to claim more effectively and institute related national policies and plans
that ICT development will eliminate the digital divide across for ICT development.
countries altogether, ICT is certainly an influential factor that Third, the theoretical contributions are that we applied differ-
has affected the global economy. The purpose of this research ent approaches to examine the convergence and peer effects of
was to investigate ICT development across countries in the last global ICT development. Whether ICT development across coun-
decade. In this study, we collected secondary data of ICT relevant tries has been converging or diverging has been a critical research
variables of 136 countries spanned five continents from 2000 to issue. Most of the previous studies have built theoretical models
2008. This is large data set among cross-country studies. It has and have applied modeling approaches to predict convergence or
been a great challenge to collect secondary and country level data divergence. Few studies empirically examined global ICT develop-
because of the differences in data capture practices and the time ment by a large data set. We also empirically validated peer influ-
lags of the data reports. Prior studies (Kauffman and Techatassan- ence theory for ICT development across countries. We applied
asoontorn 2004, 2005b, 2005c; Kauffman and Kumar 2005, Ho social influence theory (Friedkin 1998, Agarwal et al. 2009) to ini-
et al. 2007, 2011) have showed that effective data collection is tially examine the peer effects among countries. We measured
critical to create the possibilities to study global technology the peer effects based on regional spatial proximity, which allows
diffusion. identification of the peer effects from regions. We have argued
Our major findings are as follows. First, the first principle com- that disparate ICT developments created global digital divides.
ponent score and data clustering approaches show that most coun- In this study, our focus was on the analysis of global ICT develop-
tries had a convergent ICT development path during from 2000 to ment. Thus, it is natural to ask: how can we apply the knowledge
2008. The convergence occurred at non-unique ICT development we learn from the analysis of ICT development to issues related to
levels. Both the high and low clusters have more numbers of coun- global digital divides? This knowledge gap between ICT develop-
tries in ICT development after 2005 to 2008. From mobile phone ment and digital divides should be alleviated for the theory
penetration, we are optimistic to see the global digital divide was developed in the study to have practical use in improving the
somewhat less by the end of study period. Overall, ICT develop- quality of life.
ment of most countries converged near the end of the study period. There are several limitations of this study that are worthwhile
Second, the panel data analysis results demonstrate that countries to point out. First, due to the lack of available data, this study is
with different GNI levels have different ICT development paths. limited in that it only analyzed four ICT indicators. Although cov-
Countries with higher GNI levels tend to invest more in the ICT ered more countries and a longer period of time, missing data con-
infrastructure and have higher ICT development path, which has strained our ability to examine other variables in our data set. It is
been confirmed in prior studies. We can infer that ICT infrastruc- always a tradeoff to struggle with the coverage and size of a data
ture and investment have a positive association with the level of set. Since our focus in this study is to explore the trends over time,
country wealth. Third, country and region fixed-effects have also we chose to use the larger data set. Future research can focus on
been found which indicate why countries with similar GNI time collecting and analyzing more indicators. Second, we used an heu-
series had different ICT development paths. By considering the ristic method based on subjective judgments of the importance of
ICT development of peer countries in the same region, we also val- various indicators to rank multivariate centroids from the k-means
idated peer effects on the ICT development of a country. algorithm. Future research may develop more objective procedure
We next discuss the methodological, empirical, and theoretical to compare two centroids. Third, peer effects in this study were
contributions. First, our methodological contribution is that we ap- limited to peers defined by spatial proximity. As social network
plied different approaches to probe the trends of our data. The dif- analysis becomes more prevalent in general social studies, we
ferent methodological approaches allow us to perform empirical may need to consider peer effects resulting from different social
estimations and with alternative assessments of multiple analysis. ties or culture differences. For example, peers may be defined in
Our approach includes the use of k-means algorithm to cluster terms of economic or cultural proximity. Do countries involved
multivariate ICT data, which gives us a way to interpret a cluster in a social network created by economic relations also exhibit peer
through its centroid. We selected an appropriate number of clus- effects from this network? Future studies can also investigate these
ters based on the Davis–Bouldin cluster validation index. Cluster types of peer effects from non-spatial proximity.
scores simplified our investigation of the ICT development trend
of different countries and they made the ICT development path Acknowledgments
of a country easier to comprehend. Cluster scores also enabled a
panel data study connecting the GNI level and the ICT development The authors thank the program chairs, the anonymous review-
status of a country. Our proposed framework reduces multivariate ers, and audience for helpful remarks on our research at the 2011
ICT data into a manageable format of ordinal cluster scores. Cluster International Conference on Electronic Commerce in Liverpool, UK
scores gave us practical insights into the ICT development paths of in August 2011, where an earlier version of this research was pre-
many countries. With cluster scores, we were also able to inspect sented. We benefited from the comments of Fred Riggins and Rob
the impact of national wealth and various fixed effects on the ICT Kauffman, who provided helpful comments and inputs. Shu-Chun
development of a country. We last applied the panel data analysis Ho and Shing H. Doong thank the National Science Council in Tai-
approach to examine whether GNI impacts the national ICT devel- wan for partially supporting the research under Grant Nos. NSC97-
opment. Overall, different methodological approaches contributed 2410-H-017-034-MY2 and NSC99-2410-H-366-006-MY2.
532 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533

Appendix A. Notation table Appendix A (continued)

Notation Definition Comments
Notation Definition Comments
yct fct + uct Sum of the
wi A generic univariate The index i runs from 1
contribution from
variable to n
regressors and
K The number of clusters In the data clustering
idiosyncratic errors
in data clustering or step, we set k = 4 to get
Dj maxj–j0 {Dj;j0 } Dj;j0 is the within-to-
the number of means 4 clusters
in a multiple means between cluster spread
based variance formula for clusters j and j0 in
j, j0 A generic cluster Must be an integer Davis–Bouldin index
number between 1 and k distj;j0 Distance between the It measures the
j(i) Assigns wi to mean lj(i) Must be an integer centroids of clusters j between-cluster
in a multiple means between 1 and k and j’ dissimilarity in Davis–
based variance formula Bouldin index
xi A generic multivariate Conceptually, this Regionk Dummy variable (0 or k can be 002, 009, 019,
variable used in data denotes the 1) for region fixed 142 and 150 according
clustering multivariate variable effects to the UN
for ICT development ICTPEct A panel data variable Similar to the peer
gj The j-th multivariate j is an integer between representing ICT peer effect used in Agrawal
centroid 1 and k effects for country c in et al. (2009)
Pj The cluster represented j is an integer between year t.
by centroid gj 1 and k
c A generic index used In our data set, c is
for alphabetically between 1 and 136 Appendix B. Derivation of Eq. (4)
sorted countries
t A generic index used In our data set, t is The derivation of Eq. (4) runs as follows. If we assume four
for years in the study between 2000 and levels of ICT development status (1 = low, 2 = low middle,
period 2008 3 = upper middle, 4 = high), that is, ICTct e {1, 2, 3, 4}, then an
st(c) The cluster score of With our data, ordered probit or logit regression model is given by a hidden
country c in year t st(c) e {1, 2, 3, 4} variable yct = b  GNIct + a2C2 +    + aNCN + c2T2 +    + cTTM + uct,
dt(c) The absolute difference This measures the three cut points cut1 < cut2 < cut3 and the following formulae
|st+1(c)  st(c)| change of cluster (Hsiao 2003):
scores in two
PrðICT ct ¼ 1Þ ¼ Prðyct < cut1 Þ
successive years
Dt The aggregate cluster The convergence PrðICT ct ¼ 2Þ ¼ Prðcut1 6 yct < cut2 Þ
score change from all metric used to check PrðICT ct ¼ 3Þ ¼ Prðcut2 6 yct < cut3 Þ
countries in year t, i.e., for the convergence of PrðICT ct ¼ 4Þ ¼ Prðcut3 6 yct Þ
P
cdt(c) cluster scores of many
countries When the idiosyncratic error uit assumes a standard logistic proba-
Et The entropy of cluster The entropy metric bility density function eu/(1 + eu)2, we have the logit model. In the
scores distribution in used to detect following, we assume a probit model. uit assumes pffiffiffiffiffiffiffi a 2standard normal
year t concentrations of probability density function uðuÞ ¼ ð1= 2pÞeðu =2Þ . Let UðuÞ ¼
Ru
cluster scores 1
uðwÞdw denote the cumulative distribution function. Denoting
ICTct The cluster score of Same as st(c) fct = b  GNIct + a2C2 + . . . + aNCN + c2T2 + . . . + cMTM and from yct <
country c in year t cut1, we obtain an equivalent event uct < cut1  fct. Therefore,
GNIct The GNI classification Used as the main Pr(ICTct = 1) = Pr (yct < cut1) = Pr(uct < cut1  fct) = U(cut1  fct).
from the World Bank of regressor in most panel Similarly, we obtain Pr (ICTct = 2) = U(cut2  fct)  U(cut1  fct),
country c in year t data models Pr(ICTct = 3) = U(cut3  fct)  U(cut2  fct) and Pr(ICTct = 4) = 1 
C1, C2, . . . , CN Dummy variables (0 or N = 136 for our data set U(cut3  fct). Finally, the expected value of ICT given GNI is
1) for country fixed EðICT ct jGNIct Þ ¼ Uðcut 1  fct Þ þ 2ðUðcut 2  fct Þ  Uðcut1  fct ÞÞ
effects þ 3ðUðcut 3  fct Þ  Uðcut 2  fct ÞÞ
T1, T2, . . . , TM Dummy variables (0 or M = 9 for our data set
þ 4ð1  Uðcut3  fct ÞÞ ¼ 4  Uðcut1  fct Þ
1) for time (year) fixed
effects  Uðcut2  fct Þ  Uðcut 3  fct Þ ¼ Fðfct Þ
E(ICTct|GNIct) The expected value of See Appendix B for ¼ Fðb  GNIct þ a2 C 2 þ    þ aN C N þ c2 T 2 þ    þ cM TÞ
ICT given GNI formula derivation
uct Idiosyncratic errors Used in panel data
References
models
fct A linear combination of The total contribution Agarwal, R., Animesh, A., and Prasad, K. Social interactions and the ‘‘digital divide’’:
all regressors including from all regressors in explaining variations in Internet use. Information Systems Research, 20, 2, 2009,
GNI and dummy panel data models 277–294.
Andrianaivo, M. and Kpodar, K. ICT, financial inclusion, and growth: evidence from
variables of various fix- African countries, International Monetary Fund, 2011. Working paper. Available
effects at www.imf.org/external/pubs/ft/wp/2011/wp1173.pdf.
 S.H. Doong, S.-C. Ho / Electronic Commerce Research and Applications 11 (2012) 518–533 533

Banker, R., Mitra, S., and Sambamurthy, V. The effects of digital trading platforms on Johnson, R. A., and Wichern, D. W. Applied Multivariate Statistical Analysis, 6th
commodity prices in agricultural supply chains. MIS Quarterly, 35, 3, 2011, edition. Pearson Prentice Hall, Englewood Cliffs, NJ, 2007.
599–611. Kauffman, R. J., and Kumar, A. The role of MAR and Jacob externalities in the growth
Barro, R. J., and Sala-i-Martin, X. Economic Growth, 2nd edition. MIT Press, of IT industry clusters. Working paper, MIS Research Center, Carlson School of
Cambridge, MA, 2004. Management, University of Minnesota, Minneapolis, MN, 2005.
Bélanger, F., and Carter, L. The impact of the digital divide on e-government use. Kauffman, R. J., and Techatassanasoontorn, A. A. Is there a global digital divide for
Communications of the ACM, 52, 4, 2009, 132–135. digital wireless phone technologies? Journal of the Association of Information
Bresnahanm, T. F., and Trajtenberg, M. General purpose technologies: engines of Systems, 6, 12, 2005, 338–382.
growth? Journal of Econometrics, 65, 1995, 83–108. Kauffman, R. J., and Techatassanasoontorn, A. A. Does one standard promote faster
Chinn, M. D., and Fairlie, R. W. The determinants of the global digital divide: a growth? In R. Sprague (ed.), Proceedings of the 37th Hawaii International
cross-country analysis of computer and Internet penetration. Oxford Economic Conference on Systems Science, Kona, HI, January 2004, IEEE Computing Society
Papers, 59, 1, 2007, 16–44. Press, Los Alamitos, CA, 2004.
CINIC. Statistical survey report on Internet development in China. China Internet Kauffman, R. J., and Techatassanasoontorn, A. A. Is there a global digital divide for
Network Information Center, Beijing, China, January, 2010. digital wireless technologies? Journal of the Association for Information Systems,
Corrocher, N., and Ordanini, A. Measuring the digital divide: a framework for the 6, 12, 2005, 338–382.
analysis of cross-country differences. Journal of Information Technology, 17, 1, Kauffman, R. J., and Techatassanasoontorn, A. A. International diffusion of digital
2002, 9–19. mobile technology: a coupled-hazard state-based approach. Information
Cuervo, M. R. V., and Menendez, A. J. L. A multivariate framework for the analysis of Technology and Management, 6, 2, 2005, 253–292.
the digital divide: evidence for the European Union. Information & Management, Gibbs, J. L., Kraemer, K. L., and Dedrick, J. Environment and policy factors shaping
43, 6, 2006, 756–766. global e-commerce diffusion: a cross-country comparison. The Information
Crenshaw, E. M., and Robinson, K. K. Jump-starting the internet revolution: how Society, 19, 1, 2003, 5–18.
structural conduciveness and global connections help diffuse the internet. Mahajan, M., Nimbhorkar, P., and Varadarajan, K. The planar k-means problem is
Journal of the Association of Information Systems, 7, 1, 2006, 4–18. NP-hard. Lecture Notes in Computer Science, 5431, 2009, 274–285.
Dasgupta, S., Lall, S., and Wheeler, D. Policy reform, economic growth, and the Meso, P., Musa, P., Straub, D., and Mbarika, V. Information infrastructure,
digital divide: an econometric analysis. Working paper 2567, World Bank, governance, and socio-economic development in developing countries.
Washington, DC, 2001. European Journal of Information Systems, 18, 2009, 52–65.
Davis, D., and Bouldin, D. A cluster separation measure. IEEE Transactions in Pattern Niles, S., and Hanson, S. A. New era of accessibility. Journal of the Urban and Regional
Analysis and Machine Intelligence, 1, 1979, 224–227. Information Systems Association, 15, 1, 2003, 35–40.
Deb, K. Multi-Objective Optimization Using Evolutionary Algorithms. John Wiley and OECD. Guide for measuring the information society. Paris, France, 2005.
Sons, New York, 2001. OECD. Understanding the digital divide. Paris, France, 2001.
Dewan, S., Ganley, D., and Kraemer, K. L. Across the digital divide: a cross-country Okoli, C., Mbarika, V. W. A., and McCoy, S. The effects of infrastructure and policy on
analysis of the determinants of IT penetration. Journal of the Association of e-business in Latin America and Sub-Saharan Africa. European Journal of
Information Systems, 6, 12, 2005, 409–432. Information Systems, 19, 2010, 5–20.
Dewan, S., Ganley, D., and Kraemer, K. L. Complementarities in the diffusion of Papaioannou, S. K., and Dimelis, S. P. Information technology as a factor of economic
personal computers and the internet: implications for the global digital divide. development: evidence from developed and developing countries. Economics of
Information Systems Research, 21, 4, 2010, 925–941. Innovation and New Technology, 16, 3, 2007, 179–194.
Dewan, S., and Kraemer, K. Information technology and productivity: evidence from Rice, R. E., and Katz, J. E. Comparing internet and mobile phone usage: digital divides
country-level data. Management Science, 46, 4, 2000, 548–562. of usage, adoption, and dropouts. Telecommunications Policy, 27, 8–9, 2003,
Emrouznejad, A., Cabanda, E., and Gholami, R. An alternative measure of the ICT- 597–623.
opportunity index. Information and Management, 47, 2010, 246–254. Sacchi, A., Giannini, E., Bochic, R., Reinhard, N., and Lopes, A. B. Digital inclusion with
Fathian, M., Akhavan, P., and Hoorali, M. E-readiness assessment of non-profit ICT the Internet: would you like fries with that? Communications of the ACM, 52, 3,
SMEs in a developing country: the case of Iran. Technovation, 28, 9, 2008, 2009, 113–116.
578–590. Sciadis, G. (Ed.) From the Digital Divide to Digital Opportunities: Measuring
Fazio, M., Simone, A., Gregori, E., and Riccardini, F. Measuring the Digital Divide, the Infostates for Development. ITU, Geneva, Switzerland, 2005.
Digital Divide: Enhancing Access to ICTs. OECD, Paris, France, 2000. Seo, H.-J., Lee, Y. S., and Oh, J. H. Does ICT investment widen the growth gap?
Frechette, G. sg158: Random-effects ordered probit. Stata Technical Bulletin, 59, Telecommunications Policy, 33, 8, 2009, 422–431.
2001, 23–27. Shirazi, F., Gholami, R., and Higo’n, A. The impact of information and
Friedkin, N. A Structural Theory of Social Influence. Cambridge University Press, communication technology (ICT), education and regulation on economic
Cambridge, UK, 1998. freedom in Islamic Middle Eastern countries. Information and Management, 46,
Guerrieri, P., and Padoan, P. C. Modeling ICT as a general purpose technology: 8, 2009, 426–433.
evaluation models and tools for assessment of innovation and sustainable Talukdar, D., and Gauri, D. K. Home Internet access and usage in the USA: the trends
development at the EU level. 35, 2007. in the socio-economic digital divide. Communications of the Association for
Han, J., and Kamber, M. Data Mining: Concepts and Techniques, 2nd edition. Elsevier Information Systems, 28, 1, 2011, 85–98.
Morgan Kaufmann, San Francisco, 2007. Tcheng, H., Huet, J. M., Viennois, I., and Romdhane, M. Telecoms and development in
Hanafizadeh, M., Saghaei, A., and Hanafizadeh, P. An index for cross-country Africa: the chicken or the egg? Convergence Letter, 8, 16, 2007.
analysis of ICT infrastructure and access. Telecommunications Policy, 33, 7, 2009, UNCTAD. E-commerce and development report 2001. Publication No. E.01.II.D.30,
385–405. United Nations, New York, NY, 2001.
Hausman, J. Specification tests in econometrics. Econometrica, 46, 6, 1978, UNCTAD. Information economy report 2009: trends and outlook in turbulent
1251–1271. times. Publication sales No. E.09.II.D.18, United Nations, New York, NY,
Heeks, R. ICT4D 2.0: the next phase of applying ICT for international development. 2009.
IEEE Computer, 41, 6, 2008, 26–33. UNCTAD. Information economy report 2010: ICTs, enterprises and poverty
Ho, S. C., Kauffman, R. J., and Liang, T. P. A growth theory perspective on B2C e- alleviation. Publication sales No. E.10.II.D.17, United Nations, New York, NY,
commerce growth in Europe: an exploratory study. Electronic Commerce 2010.
Research and Applications, 6, 3, 2007, 237–259. UNESCO. Measuring and monitoring the information and knowledge societies: a
Ho, S. C., Kauffman, R. J., and Liang, T. P. Internet-based selling technology and e- statistical challenge. UNESCO, Paris, France, 2003.
commerce growth: a hybrid growth theory approach with cross-model Vicente, M. R., and Lopez, A. J. Accessing the regional digital divide across the
inference. Information and Technology Management, 12, 3, 2011, 409–429. European Union. Telecommunications Policy, 35, 3, 2011, 220–237.
Hsiao, C. Analysis of Panel Data, 2nd edition. Cambridge University Press, Cambridge, Witten, I., and Frank, E.. Data Mining, Practical Machine Learning Tools and
UK, 2003. Techniques. Morgan Kauffman, San Francisco, CA, 2005.
IDC. Information society index. Framingham, MA, 1995. Available at www.idc.com/ World Bank. Country and lending groups. Washington, DC, 2009. Available at
groups/isi/main.html. data.worldbank.org/about/country-classifications/country-and-lending-groups.
ITU. World telecommunication development report: access indicators for the Zorn, M. A remark on method in transfinite algebra. Bulletin of the American
information society. Geneva, Switzerland, 2003. Mathematical Society, 41, 10, 1935, 667–670.