This research attempts to explore digital divide among tribes in Taiwan and to reveal relevant important tribal contextual features. Firstly, we collect residents’ connection event logs from the free wireless Internet, i-Tribe. Those connect behaviors are quantified as average of daily connection frequency, daily connection time, daily data transmission, inter-login time, Internet speed, and connection quality which are corresponding to evaluate how often, how long, how much, how active, how fast and the effectiveness of the information and communication technology (ICT) accessibility of tribes. Then the tribes are grouped into tribal clusters optimally by a rank aggregation algorithm and are asked if there are significant differences among them by three stages of hypothesis testing. After the phenomenon of digital divide is confirmed, some important contextual features, longitude, female dependency ratio, network environment level, economic service level, and policy cooperation level, are identified using the ordinal regression model (ORM).
The advancement of information technology has contributed to the booming development of the telecommunication industry. The public can easily access information via computers, handsets or tablets at an increasingly lower cost. This allows the utilization of information in day-to-day life and at work, and results in tremendous influences on people’s lives and businesses (Chipeva et al. 2018; Goncalves, Oliveira, and Cruz-Jesus 2018; Gray, Gainous, and Wagner 2017). Many researchers have indicated that as the world is increasingly digitalized, the continued creation, distribution, and application of information has become the lynchpin of modern society and has evolved into innovative paradigms. It has redefined the way we live and work, promoted social, economic and cultural developments, and improved life quality (Cruz-Jesus et al. 2016; Ferro, Helbig, and Gil-Garcia 2011).
The public can obtain information they require in an open and transparent way, given the current rapid distribution of information. This should help to eliminate the effects of information asymmetry on decisions in daily life. It should also encourage fairness and justice in society and mobility across social classes (Cruz-Jesus, Oliveira, and Bacao 2012). However, the reality is that accessibility and dissemination of information is unevenly distributed. Not all people can benefit from the wave of digitalization. There are gaps of various kinds. Many scholars believe that the digital divide, i.e., significant variances in information access, searches, ownership and utilization between countries, social groups and individuals, is caused by genders, social classes, ethnicity, occupations, education, income levels, and countries of residence (Lucendo-Monedero, Ruiz-Rodríguez, and González-Relaño 2019; Szeles 2018; Van Dijk 2006). The digital divide prevents disadvantaged groups from enjoying the convenience brought about by the popularity of information technology. The inability to access and use information also deprives them of the opportunities to compete with others in a fair manner. This causes social class rigidity and adverse effects on the development of the economy, society, and culture (Ferro, Helbig, and Gil-Garcia 2011).
An increasing number of researchers are seeing the influence of the digital divide on social developments. They have embarked on relevant studies into the causes of the digital divide in order to develop responsive measures (Cawkell 2001; Goncalves, Oliveira, and Cruz-Jesus 2018). In terms of research subjects, literature on the digital divide can be categorized into cross-country comparisons and a focused country discussion. Many studies have examined the digital divide in different nations and researchers have argued that the differences from one country to another in politics, economy, society, and culture are the culprits of the digital divide. It is hoped that a comparison of institutional environments across different countries can shed light on the reasons for the differences in the levels of digitalization (Chinn and Fairlie 2007; Lindblom and Räsänen 2017). Some researchers have attempted to focus on the differences in incomes, genders, ethnicity, education, and occupations of the people in specific countries, in order to explore the correlation between the digital divide and various attributes. Their findings can serve as a template for governments in regard to the design of information policies (Bornman 2016; Chaudhuri, Flamm, and Horrigan 2005; Colley and Maltby 2008; Puspitasari and Ishii 2016; Silva, Badasyan, and Busby 2018).
All of the discussions on the reasons for the digital divide help us to better understand the problems and provide a diversity of perspectives. This allows us to think deeper about the issues of the digital divide (Goncalves, Oliveira, and Cruz-Jesu 2018; van Dijk and Hacker 2003). The literature provides different perspectives and analytical angles. There are attempts to collect data via primary research (e.g., questionnaires) or secondary research (e.g., social/economic indicators) in order to measure the digital divide. However, this type of research design ignores the context of the problems that lead to a digital divide. Numerous scholars have indicated that the digital divide is not caused by a single factor. Rather, it is directly affected by the interrelationships of industrial structures, gaps between urban and rural areas, economic status, social support, cultural developments and education systems (Silva, Badasyan, and Busby 2018; Szeles 2018; Van Dijk 2006).
To delve into the phenomenon of the digital divide and clarify the environment factors in this context, this paper seeks to examine remote tribes in Taiwan by looking at how the local individuals use wireless fidelity (Wi-Fi). The collation of data concerning the average frequency of daily connections, daily connection time, daily data transmission, inter-login time, Internet speed, and connection quality aims to establish an overall picture of data access and utilization. The data collection period was from 2017. A total of eight million wireless connection event logs were collected. This paper applied the clustering method and multivariate analysis of variances (MANOVA) in the examination of the digital divide among 70 remote tribes in Taiwan. Meanwhile, the ordinal regression model (ORM) is deployed to explore the contextual features in these tribes in order to identify the features relevant to the digital divide (these issues include geographic features of tribes, demographic features of tribes, wireless access point [WAP] infrastructure of tribes, public service level, network environment level, educational service level, medicare and social benefit level, economic service level, and policy cooperation level). Finally, this paper presents the research findings to the Taiwan government as a reference for information policymaking. Hopefully these efforts can help to eliminate the digital divide so that more people can share the benefits brought about by information technology. This will enhance the quality of life of the general public. In the meantime, this paper seeks to derive theoretical or practical implications surrounding the research findings, so that the results can serve as a reference for both academics and practitioners and provide direction for future studies.
The rest of this paper is structured as follows. First, the digital divide is widely reviewed. Second, the detailed proposed research flowchart is described. Then the experimental results of 70 tribes are presented. Finally, the concluding remarks and further research suggestions are discussed.
2 Literature Review
2.1 Brief Review of the Digital Divide
Digitalization has significantly improved the productivity of different industries and tremendously affected the existing paradigms in society. It has enormous influences on politics, cultures, economies, and laws, and has promoted governments, corporates, nonprofit organizations, and individuals to render services with new approaches and methods (Bartikowski et al. 2018; Nishijima, Ivanauskas, and Sarti 2017). To summarize, digitalization has redefined how society operates. Many scholars have explored the impact of digitalization on social functioning. Their purpose is to identify how nations or individuals behave and respond in the era of an information society. Many scholars argue that the ubiquity of information technology helps to reduce the cost of using information so that we can more efficiently create, disseminate, and apply information in life and at work (Cruz-Jesus et al. 2016). This helps to improve life quality and work efficiency. The ubiquity of the Internet serves as the best platform for information dissemination. Information can be widely distributed in an open manner. It is available to all users in an equal way. This should make knowledge travel fast and wide, contribute to social mobility, and realize social justice and fairness (van Dijk 2006).
While the benefits of digitalization are well-documented in the academic literature, many scholars are concerned about the adverse effects of the digital divide (Ferro, Helbig, and Gil-Garcia 2011; Gray, Gainous, and Wagner 2017). There are significant variances in the opportunity, capability or literacy of using information carriers between countries, social/ethnic groups and individuals as a result of differences in gender, social class, ethnicity, occupation, education, and income level (Colley and Maltby 2008; Lindblom and Räsänen 2017). First, such differences (e.g., economic difficulty) may not allow equal opportunities for everybody to use digital carriers or technology to access the required information. Second, even if the public can obtain the digital carriers or technology required, they may not have sufficient capability or education to complete information searches and applications (Bruce 2000; Cawkell 2001; List 2019). The specific groups unable to stay on top of information will not be able to enjoy the benefits of digitalization. These people will only be able to maintain existing jobs and lifestyles. This will result in uneven opportunities in education, a reduction in available jobs, and difficulty in achieving income increases.
Based on the abovementioned observations, there is a long list of features contributing to the digital divide. This can be explored from the macro or micro perspectives. Parts of scholars start with macro views, by explaining the digital divide between countries in terms of political, policy, economic, social, cultural, educational, and infrastructure elements. They have also conducted large-scale cross-country comparisons in order to identify the causes of the digital divide (Cruz-Jesus, Oliveira, and Bacao 2012; Park, Choi, and Hong 2015; van Dijk and Hacker 2003). Scholars have explored the issues from micro perspectives, by focusing on how the differences in gender, social class, ethnicity, occupation, education, and income level of the populations in different countries affect the digital divide (Bornman 2016; Colley and Maltby 2008; Puspitasari and Ishii 2016).
Researchers may have different views regarding the causes of the digital divide, but they are all concerned about how the adverse effects of the digital divide on society can be eliminated. From the micro perspectives, the digital divide deprives an individual of the opportunity to access information in an equal manner. This affects life quality and work efficiency and prevents the possibility of competing with others in a fair manner (Bornman 2016). From the macro perspective, a significant gap in the creation, access, and application of information between countries has direct influences on international competitiveness. This will lead to a vicious circle of the poor becoming even poorer (Cruz-Jesus et al. 2016; Lindblom and Räsänen 2017; Szeles 2018). Some researchers have attempted to delve into the digital divide by looking at digital opportunities, digital inclusion, technology apartheid, and knowledge gaps to develop methods to eradicate the digital divide (Hsieh, Rai, and Keil 2008; van Dijk 2006).
What the digital divide entails evolves over time as a result of continued development of telecommunication technology (Attewell 2001; Puspitasari and Ishii 2016; van Dijk 2006). The examination of the digital divide in the early days focused on the ownership of information carriers such as phones, modems, personal computers, cable TV, and network devices. These measures were referred to as the main basis for the valuation of the digital divide (Attewell 2001; Cuervo and Menéndez 2006). With the gradual development of telecommunication technology and the increasingly comprehensiveness of internet-related infrastructure, connection bandwidth, speed, and reliability has been improving. A plethora of information carriers (e.g., handsets, tablets, wearables, and smart home devices) has been invented, and the cost of network access has continued to drop. At this juncture, connectivity (i.e., the cost of network access) is no longer a heavy burden for most countries or people.
With the continued reduction of network access costs, a lack of information literacy becomes the main cause of the digital divide (Bruce 2016; Goncalves, Oliveira, and Cruz-Jesus 2018; Yu, Lin, and Liao 2017). The continued drop of the costs associated with information creation, access, and dissimilation has contributed to the widespread use of information. Therefore, it is necessary to master the skills required to quickly and accurately identify the information needed from a massive volume of information. Many researchers have described the capability to search, identify, and apply information as information literacy. They believe a lack of information literature is a key reason for the digital divide (Nishijima, Ivanauskas, and Sarti 2017; Van Dijk 2006; Yu, Lin, and Liao 2017).
Past studies provide a diversity of perspectives regarding the contributing factors to the digital divide. However, we believe there is room for improvement. First, prior research is based on cross-country comparisons as the analytical level, with the intention of exploring the causes of the digital divide from economic, cultural, and political perspectives in different countries (Lindblom and Räsänen 2017; Park, Choi, and Hong 2015; Szeles 2018). Some studies focus on specific countries and investigate demographic attributes such as incomes, genders, education backgrounds, and occupations (Chaudhuri, Flamm, and Horrigan 2005; Lindblom and Räsänen 2017; Puspitasari and Ishii 2016). Their purpose is to articulate the variances in people’s capability in accessing, gathering or applying information. To offer a fresh outlook, this paper refers to remote tribes in Taiwan as the unit of analysis. The objective is to delve into the digital divide from the perspective of communities in order to explain the reasons for the digital divide in a comprehensive way. Second, some scholars contend that the digital divide is not caused by a single or even a small number of features. In other words, the level of digitalization can be considered the outcome of interactions among political, economic, social, technological, educational, and cultural elements (Chinn and Fairlie 2007; Szeles 2018). If the context is taken into account, and examination is conducted as to how the institutional environment affects the way people access, search, and apply information, the study will derive richer practical implications. It is hoped that these research findings can help to better respond to the digital divide and serve as a guide for the government when making digitalization policies. Finally, it is also about research design. Past studies typically conducted questionnaire surveys or referred to quantitative indicators (e.g., the number of personal computers, connection bandwidth, and expenses) as the measurement for the digital divide (Attewell 2001; Puspitasari and Ishii 2016). While these metrics can, in part, reflect the status of information access for different countries or populations, they cannot accurately evaluate users’ behavior on the Internet. Moreover, digital divide is the study of the access to the Internet (Van Dijk 2017). Our work seeks to profile how residents in remote tribes in Taiwan use Wi-Fi. This approach to evaluate information access status marks a contrast with the methods previously used in the literature. The goal is to establish more accurate and real-time measurements of the digital divide.
2.2 Digital Divide in Taiwan
Although the purchase of information devices (e.g., handsets and tablets) is no longer a heavy burden for people in Taiwan, there is a digital divide in remote tribes due to terrain restrictions. Taiwan boasts a total area of 36,193 square kilometers, but over 75% of it is covered with hills and maintains. There are 269 mountains over 3000 m in altitude. In fact, Taiwan is one of the islands with the highest density of high mountains in the world. The topological complexity has caused inconveniences in transportation in remote regions and has a direct impact on the life and livelihood of the local people. The small number of residents and the inherent limitation of the terrain makes it difficult to establish wire and wireless network infrastructure in remote tribes. Given the consideration of costs and benefits, private network operators are less willing to deploy networks in these far-off areas. This has caused a digital divide for remote tribes in Taiwan and has affected the opportunity of locals in the access and use of information.
To eliminate the digital divide in remote regions, the Taiwan government has formulated and initiated many projects (Tsai et al. 2021), such as “Accelerated Mobile Broadband Service and Industrial Development Plan,” “Wireless Broadband Environment Construction Plan for Aboriginal Regions,” and the “Special Spectrum Policy for Enhancing the Infrastructure of Wireless Broadband Networks in the Remote tribes.” The idea is to lease the equipment required from private telecom operators for wireless broadband connections in order to set up i-Tribe, a network system in remote villages in Taiwan. The establishment of WAPs provides free Wi-Fi services to locals. This combined with public services such as education, employment, medical care, leisure and social welfare can enhance life quality and economic development of these communities. It promises to offset the adverse effects of the digital divide and brings the benefits of information to residents in remote tribes.
3 Research Methodology
In Taiwan, mobile Internet services are charged at flat rates of around 599 NTD-2699 NTD (Chunghwa Telecom 2021), but it is still relatively high considering a monthly basic wage of 24,000 NTD (Taiwan Ministry of Labor 2021). Moreover, even using an unlimited package, there is only a limited amount of data for the max speed roaming. Hence, the users always tend to save their roaming data and are more likely to use the Wi-Fi. Based on the Wi-Fi log records, this research attempts to explore the phenomenon of digital divide and realize multiple related contextual features among remote tribes in Taiwan. In general, a single person uses only one to two smart devices (e.g., iPhone, iPad) for activities of a normal daily life, and they are likely to use them together at a single location (bring the devices with them). Only a minor number of persons use different devices at different locations for the convenience of their particular works, for example, professors, researchers, or medical specialists, etc., who need to access the network frequently (independent of the locations) to update their work for an expedited research progress or medical treatment, etc. In the tribes, probability of these particular scenarios should be very low. Moreover, the logs collected in our work are massive. Therefore, the selection bias can be avoided.
Our basic idea is to cluster these tribes based on their features of ICT accessibility and then ask if there are significant differences among clusters’ mean vectors. After the phenomenon of digital divide is confirmed, the important demographical features that potentially relate to the digital divide are identified. The above procedure can be done via a clustering procedure and then follows a regression one. The overall flowchart of the proposed methodology is shown in Figure 1. The detailed procedure is described latter.
3.1 Exploring the Digital Divide
To better understand digital divide in remote regions of Taiwan, this study collects data from the i-Tribe network system regarding how locals use Wi-Fi. I-Tribe was initiated in 2015 (Taiwan Executive Yuan 2015), and now has a contribution in significantly minimizing urban–rural disparities and making the right to digital participation a reality for all (Digi Taiwan 2018). By collecting data of the connection event logs, this paper is able to profile how local people access and apply information. Access to the i-Tribe network does not require authentication. Anybody can log on with smartphones, tablets or computers anytime. This is a very convenient service for connecting to the Internet. This paper samples the records of Wi-Fi access by locals and profiles the user behavior after data cleaning. The technique of rank aggregation is used to choose the optimal clustering method and number of clusters. The clustering is based on how residents in remote tribes access information, in order to identify digital divide among tribal clusters.
3.1.1 Extracting Features of Tribal ICT Accessibility
In order to quantify the ICT accessibility of tribes, several features are extracted. Different from most of the features described in Section 2.3 which come from large-scale investigation using either phone interview, face-to-face interview or questionnaire (Brandtzæg, Heim, and Karahasanović 2011), our ICT features are purely computed from users’ log data, i.e. the actually usage behaviors of Wi-Fi. The main features used in this study are average daily connection frequency, daily connection time, daily data transmission, inter-login time, Internet speed, and connection quality, which are evaluated in the perspective of how often, how long, how much, how active, how fast, and the effectiveness of the ICT accessibility of tribes respectively. We hence named the above features as 6Hs which are described below. As shown in Figure 2, the 6Hs are firstly aggregated at the level of wireless devices which are identified by their media access control address (MAC address) using average operation, and are then aggregated again to the level of tribes.
Average daily connection frequency counts the daily login frequencies of each network card which are then divided by number of days as well as number of network cards to be analyzed. This index evaluates how often the tribal users averagely use the wireless networks per day. The higher the index is, the more login frequency of tribal users is.
Average daily connection time accumulates the daily time difference between logout time and login time of each network card which are then divided by number of days and number of network cards. The unit of measurement for this index is second. This index reveals how long the tribal users averagely use the Internet per day. A higher value of this index means users spend more time on Internet.
Average daily data transmission accumulates the daily data volume to be transmitted of each network card including downloads and uploads. The unit of measurement for this index is gigabit. Then all of those daily data transmissions are also aggregated to the level of tribes by taking the average. This index measures the average, in gigabits, of files downloaded or uploaded by the tribal users per day. The higher the index is, the more transmission behaviors engaged in by the tribal users.
Average inter-login time calculates the time difference between two successive login times of each network card which are then divided by the number of logins minus one as well as the number of network cards. This index shows how actively the tribal users use the WAP faculties per day. A higher value of the average inter-login time means the login interval of the tribal users is denser.
Average Internet speed maps the types of IEEE 802.11 wireless standards (such as IEEE 802.11a, 802.11b, 802.11g, 802.11n, 802.11ac,802.11ad, and so on) that a network card temporally used into the corresponding fastest maximum speed is then taken as the average operation to quantify the possible Internet transmission speed. The unit of measurement for this index is megabits per second. Then all of those Internet transmission speeds of each network card are then aggregated to the tribe level by taking the average again. This index shows how fast the Internet speeds are as the tribal users are surfing the Internet using the WAP. The transmission speeds are faster with a higher average Internet speed.
Average connection quality accounts for the percentage of very short-term connections of each network card and then can be averaged up to the level of tribal. A short-term connection signifies a bad connection experience caused by the WAP failure or the transmission signal is interfered with for some reasons. This index shows how good the Internet connection qualities are as the tribal users are connecting to the WAP faculties. The higher the value of the index, the better the quality is.
3.1.2 Segmenting Tribes into Clusters using Multiple Clustering Algorithms and Clustering Validation Indexes
This research employs clustering techniques in the explorative phase of huge number logs to discover tribe clusters with similar ICT accessibility. Cluster analysis is a method of exploring internal structure of data. There are no specific procedures for data preparation. Rather, it is about gathering similar observations into clusters according to their common traits. Distance is often used as a clustering basis. The closer the distance is, the higher the similarity is. The clustered data is characterized with lower heterogeneity within groups and higher heterogeneity between groups. It remains a challenge to select the appropriate clustering method and determine the reasonable number of clusters for any given data set. No single clustering algorithms promise well-rounded results for all data structures. Meanwhile, different clustering validation measurements evaluate different qualities of clustering. Therefore, the use of one clustering method and one clustering validation index is not sufficient (Sekula, Datta, and Datta 2017). According to suggestions by Brock et al. (2011), this paper employs multiple cluster methods and clustering validation indexes and combines all of the results by the rank aggregation algorithm, which was developed by Pihur, Datta, and Datta (2007, 2009. Sekula, Datta, and Datta (2017) described this process as optimal clustering, as it allows objective, robust, and automated judgment of optimal methods and cluster numbers by alleviating the burden of human judgment.
To perform optimal clustering, this paper selects seven cluster algorithms, seven clustering validation measurements and rank aggregation algorithms based on the Monte Carlo cross-entropy (MCCE) approach, described as follows:
The set of clustering algorithms
This paper deploys partition-based, hierarchical-based and model-based method for clustering algorithms and analysis. This includes the following:
Partitioning methods construct partitions for a dataset and then evaluate them by the chosen criterion. Typical methods adapted in this research are k-means (abbreviated as kmeans), partition around medoids (abbreviated as pam), and clustering large applications (abbreviated as clara). The kmeans represents each cluster its center, while the pam and the clara represent that by one of its samples. The pam iteratively finds the medoids using the whole dataset, but the clara chooses a representative set of datasets.
Hierarchical methods build up a hierarchical decomposition of a dataset using agglomerative or divisive strategies. In hierarchical clustering (abbreviated as hierarchical), the agglomerative concept is adopted where each sample is considered as a cluster, and similar cluster pairs are merged recursively. In divisive analysis clustering (abbreviated as diana) and the self-organizing tree algorithm (abbreviated as sota), the divisive concept is adopted where all samples are considered as a single cluster, and a dissimilar sample is split recursively.
Model-based method (abbreviated as model) introduces mixtures of normal distributions to fit the dataset using the expectation-maximization (EM) algorithm.
The set of clustering validation indexes
This paper selects internal validations and stability validations as the clustering validation indexes of the quality of clustering. Internal validations are statistical measurements of clustered results based on complete data sets. They measure the following:
Connectivity: the degree of neighboring observations in the same cluster.
Dunn index: the ratio of the minimum distance between different clusters to the diameter of the largest cluster.
Silhouette width: the means of the level of confidence for the designated cluster of each observation.
Stability validations are the comparison of stability of clustering results based on datasets with features gradually removed and the result based on a complete dataset.
Average proportion of non-overlap (APN): measurement of the average proportion of the observations placed in different clusters.
Average distance (AD): measurement of the average distance between observations in the same cluster.
Average distance between means (ADM): measurement of the average distance between observations and the center point of the same cluster.
Figure of merit (FOM) measures the average intra-cluster variance.
In these cluster validation indices, except for when the value range of Silhouette width is from −1 to 1, the range of values of other indices is from 0 to infinity; and except for Dunn index and Silhouette width, where larger values mean better cluster results, with other indices STB.
The rank aggregation algorithm
Given the availability of numerous clustering algorithms and clustering validation measurements in the field of machine learning, how to select the clustering algorithm and determine number of clusters that performs the best for our application is not an easy task. A manual aggregation of multiple results is impracticable, and an automatic scheme, rank aggregation, is needed. Suppose we have m ordered lists (L i ) produced by m given evaluation methods, the rank aggregation is an optimization problem which iteratively finds out the most appropriate proposed ordered list δ* with minimized total distance between δ* and all L i
where δ is any proposed ordered list; and d(.) is the weighted Spearman’s footrule distance (Fagin, Lotem, and Naor 2003) as shown in Equation (2). The minimization operator is carried out for all possible proposed ordered lists δ.
where r δ (t) is the rank of element t in δ and N(.) represents the corresponding max–min normalization score with a range from 0 to 1 which let the weights across different measurements be comparable. The MCCE algorithm is a stochastic search method which is able to propose different ordered lists δ and then iteratively find out the most appropriate δ* via the stages of initialization, sampling, updating, and convergence (Pihur, Datta, and Datta 2007, 2009).
3.1.3 Confirming the Digital Divide using Hypothesis Testing
To confirm the existence of digital divide across tribes, this paper conducts three stages statistical tests on the significant differences of ICT accessibility for tribal clusters. Literature posits the use of (non-)parametric one-way analysis of variance (ANOVA) tests to examine statistical differences (Cruz-Jesus, Oliveira, and Bacao 2012; Mardikyan et al. 2015; Wu et al. 2014). However, this paper performs the MANOVA test to review multiple features at once, before visiting individual features with one-way ANOVA tests. This is followed with post-hoc comparisons to explore the relative strengths of respective factors across clusters.
3.2 Exploring Relevant Contextual Features for Digital Divide using a Regression Model with Ordinal Response
When it comes to the exploration of the features concerning digital divide among tribes, the contextual feature files of different tribes are the input and the information regarding the ranking tribal clusters which fall into the derived in Section 3.1 are the output for the establishment of the ORM. The purpose is to calculate and determine the important contextual features pertaining to digital divide in tribes.
3.2.1 Extracting Tribal Contextual Features
This paper gathers the data regarding the contextual features relevant to the causes for digital divide in tribes. The contextual features consist of nine aspects. The data is mostly sourced from the National Development Council.
Geographics: the tribe locations, administrative zones e.g. city/county, and latitudes/longitudes.
Demographics: tribal resident compositions, e.g. ethnicity, number of people, male to female ratio, dependency ratio.
WAP infrastructure: WAP deployments within tribes with details on the building phases.
Public service level: the level of government agencies and service stations in the tribes, such as culture and health centers, industry demo sites, after-school tutoring and assistance, immersive kindergartens, libraries and information kiosks, schools.
Network environment level: the level of investment in the network environment by government agencies, such as fibre-to-the-home (FTTH), school networks, and network infrastructure.
Educational service level: the level of learning opportunities in the tribes. Examples are after-class support, tribal languages learning, cultural centers, and educational activities.
Medicare and social benefit level: the level of medical services, social welfare institutions, e.g. mobile medicare, culture and health stations, employment assistance offices, native family support centers, and social benefits organizations.
Economic service level: the level of economic developments in the tribes; such as tourism, agriculture, hot spring, creative culture, and economic activities.
Policy cooperation level: help and assistance to tribes via development policies from local governments and policy considerations.
3.2.2 Extracting Clustering prior to Regression with the Ordinal Response
Ordinal data is one of scale of measurements in statistics where the features have ordered categories and the difference between is unmeasurable. The result of ranking the tribal clusters’ ICT accessibility described in Section 3.1 is treated rightfully as an ordinal response and can be analyzed using the ORM. The ORM is proposed by Agresti (2002) is one of models for such task because the response of samples is treated as ordered categorical and a flexible regression framework (Christensen 2015). The choice of link functions may influence complexity and performance of ORM (Javali and Pandit 2010). During the modeling process, the ORM is used for modeling the ranked tribal clusters associated contextual features with different link functions in this study such as logit, probit, log–log, and complementary log–log. The most useful link function can then be determined via selecting the one that yields the best fit. The parameters of ORM are usually estimated by maximum likelihood (ML) using Newton–Raphson algorithm. In sum, an ORM can be used to perform the following tasks in this study. Firstly, to identify the significant tribal demographical features that influences the ranking of tribal clusters. Secondly, to describe the direction of the relationship between the ranking of tribal clusters and the tribal demographical features. It is hoped that the analyzed model is interpretable, thus some modern machine learning techniques such as ordinal classification and regression tree (CART) or ordinal random forest (RF) are not taken into consideration.
4 Research Experimental Results
The analyses presented in this section are conducted using a self-programming toolkit under the R environment with the dplyr, optCluster, ggmap, mapproj, caret MASS, and AER packages (http://cran.r-project.org/).
4.1 Data Collection and Leaning Processes
This paper collects data from two sources: (1) wireless connection event logs for the information systems deployed at tribes by the Institute for Information Industry and Chunghwa Telecom Company Limited; (2) open data concerning contextual features of different tribes gathered by the National Development Council over time.
Wireless connection event logs are primarily for the analysis in Section 4.2.1 below. These are a total of 8,100,456 logs from a total of 60,267 network cards in 2017. Omissions or errors are inevitable in the process of data collection due to transmission stability issues. To avoid the inference of data deficiency on analytical algorithms, this paper cleans the data before analysis. This ensures the consistency of data formats and the lock-in of subjects for analysis. Three criteria were established for the cleaning procedures of data.
Removal of logs with temporal connections: connections are occasionally lost at the system end or at the terminal end, due to poor weather, topological issues, network overloading and device problems. Some connections logged were very brief. If users cannot maintain stable connections, they will create a series of very short logs due to constant disconnections. This type of log inflates the number of connections and reduces the average log-on time, upload/download traffic, and hence should be removed. This paper deletes the logs with connections shorter than 10 s.
Removal of MAC addresses with short-term usages: this paper sets out to analyze the usage pattern of native peoples. Visitors (e.g. tourists or infrequent users) are not the research subjects and, in fact, their logs inflate the number of connections. This paper thus deletes the MAC addresses with less than eight connections and activities less than five days from the sample pool.
Removal of invalid logs: few logs may report abnormal values due to transmission errors. This paper leaves out these unidentifiable or missing data. Examples are MAC addresses not identifiable (e.g. 00-00-00-00-00-00) and IP addresses strange to this project.
After the abovementioned data cleaning, a total of 1.11 million logs were removed and approximately 6.99 million logs were still in the sample pool. About 40,000 MAC addresses were deleted, and only 20,575 remained. This number was close to the 26,000 non-dependent native residents in the government’s record.
The data files of contextual features for individual tribes are used for the analysis in Section 4.2.2. The process of data pre-treatment is a lot simpler, and only focused on a few tribes such as Tjuabal, Pucunug, and Gadu by filling in the data holes of demographic characteristics with missing values. All the other data sets were comprehensive.
4.2 Experimental Processes and Results
4.2.1 Confirming Digital Divide Phenomenon among Taiwan’s Tribes
After data cleansing, this paper measures the tribal residents’ ICT accessibility for information access with 70 tribes as the analyzed units based on their log data measured by MAC activities. The 6Hs for 70 tribes are measured as the ICT accessibility. In order to avoid dimension domination in clustering algorithms, this paper standardizes the data for these features. To ensure the clarity in the assessment of ICT accessibility across tribes, this paper deploys the larger the better (LTB) methodology for the ideal function of the standardized ICT accessibility. In other words, the standardized value of the average inter log-in time is multiplied with −1. This completes the processing of the 6Hs for all the tribes.
The next step is to identify the clusters of information access intensity for different tribes. This paper performs optimal clustering with MCCE-based rank aggregations. Hyper-parameter setup is in reference to the default value of the package, optCluster, in the R language. For example, the maximum number of iterations allowed is 1000. The MCCE process stops once the best solution does not change after seven iterations. This approach enables the objective determination of the optimal methods and the optimal number of cluster numbers for a given dataset. As shown in Figure 3, this paper selects the clustering algorithms and cluster validation indicators described in Section 3.1.2 for analysis. The line graph records the results of the seven clustering algorithms run on the seven cluster validation indicators for three to eight clusters. The clustering of three to eight is because (1) the analysis on two clusters yields little meaning; (2) there are often empty clusters once the number of clusters exceeds nine.
When it comes to ranking, the rank aggregation approach treats all the seven methods in six clusters as different clustering techniques (hence, δ * and δ in Eq. (1) and the word vector of 7 × 6 = 42 of the L i length). This is followed with a comparison of the performance of all the seven cluster validation indicators (thus m = 7). To avoid the tediousness of the ranking results of all the 42 combinations, this paper only presents the top five results in Table 1. The tests and numbers in the table represent clustering algorithms and the number of clusters, respectively. According to Table 1, hierarchical clustering performs better for all the cluster validation indicators among the Top 5. It also appears more frequently than others. As hierarchical-3 is most often in the Top 1, it should be the suggested method. The analysis on the optimal rankings produced by rank aggregation suggests the optimal proposed list should be δ* = [hierarchical-3, kmeans-4, hierarchical-6, hierarchical-7, hierarchical-4, model-3]. Finally, the cluster analysis of ICT accessibility across tribes is based on hierarchical clustering as the optimal technique and three clusters as the optimal number of clusters. The clustering results are shown in the Google Map in Figure 4 in which green, red, blue, and red are corresponding to cluster 1, 2 and 3.
|Clustering validation indexes||1||2||3||4||5|
The means and sample sizes of ICT accessibility of different tribes are summarized in Table 2. To confirm the existence of digital divide across tribes, this paper conducts statistical tests on the significance of differences of ICT accessibility for three clusters. The first stage is to conduct the MANOVA test on the mean of 6Hs of the three clusters. The test result suggests that the value of Wilkes lambda is 0.103, with a p-value of 0.000. This implies significant differences in the ICT accessibility and indicates the presence of digital divide across these three clusters of tribes. The second stage is to perform the one-way ANOVA on each ICT accessibility to find out which ones have significant differences across tribes. The results are shown in Column 5 of Table 2. The test result indicates no statistically significant difference in the average connection quality for the three tribes. We can also see that the Internet access is not necessarily conditional upon the connection quality. It should depend on users’ specific purposes or habits, which are not similar in respective tribes. The p-values of other ICT accessibilities are all below 0.05. Finally, Tukey’s honestly significant difference (HSD) test is conducted for a post-hoc comparison on the ICT accessibilities with statistical significance in order to rank the strengths of these features across tribes. The results are summarized in Column 6 of Table 2. The Cluster 3 reports better performance in the average daily connection frequency, connection time, data transmission, and inter-login time. The Cluster 2 follows closely behind, with relative strong performance in the average Internet speed.
|Features of tribal ICT accessibility||Cluster 1||Cluster 2||Cluster 3||p-Value of one way ANOVA||Post-hoc comparisons via TukeyHSD|
|Average daily connection frequency||−0.711||−0.635||0.814||0.000***||Cluster 3 > Cluster 2 = Cluster 1|
|Average daily connection time||−0.506||−0.245||0.359||0.007**||Cluster 3 > Cluster 2|
|Average daily data transmission||−0.318||−0.517||0.611||0.000***||Cluster 3 > Cluster 2 = Cluster 1|
|Average inter-login time||−2.552||−0.072||0.571||0.000***||Cluster 3 > Cluster 2 > Cluster 1|
|Average Internet speed||−0.207||0.427||−0.414||0.027*||Cluster 2 > Cluster 3|
|Average connection quality||−0.841||0.408||−0.272||0.527||–|
|Number of tribes||6||31||33|
4.2.2 Revealing Digital Divide Contextual Features among Taiwan’s Tribes
The ORM is used to achieve the goal in which the contextual features are the input and the ranked clustering labels will be the ordered response. Especially, some categorical features such as city, county or ethnicity of tribe are not taken into consideration because ORM needs a huge degree of freedom to model those dummy codes that our number of tribes in dataset does not support. Another feature, the number of people in the tribe, is discarded because it is highly correlated with other features. The analysis process is divided into two stages. The first stage is to select an appropriate link function for the model of ORM. To do so, we adopt the residual deviance and the Akaike information criterion (AIC) to compare the model performance of different link functions. A lower residual deviance or AIC value indicates a better fit. According to order of suitability of residual deviance in Table 3, the ORM with complementary log–log function is a better fit (113.32) as compared to logit (115.19), log-log function (115.84), and, probit (116.30). This phenomenon is also supported by another measurement, AIC. It finds that both measurements have lowest value when the complementary log-log is used, with link function thus adopted in following analysis. Thus, we conclude that the ORM with complementary log–log function has a better fit to tribal clusters ordinal labels.
|Link functions||Residual deviance||AIC|
After the best link function is select, the second stage is to identify significance and directionality of tribal contextual features and their interpretations. The regression coefficients with their values, standard errors, T values, and p-values are shown in Table 4. Initially, there is no significance test on each feature in ORM because Agresti (2002) and Venables and Ripley (2002) believed the Wald tests or other approximate methods are inappropriate. In this research, we use the quasi-t Wald tests to estimate the significance of coefficients in which it calculates p-values by comparing the T values against a T distribution with a residual degrees-of-freedom. The result in Table 4 shows that tribe’s female dependency ratio, longitude, network environment level, and policy cooperation level are the significant contextual features relating to the digital divide among tribes. Here, the meaning is, for example, the number of the female users in each cluster significantly affecting the Internet access in the tribes. Similar explanations can be provided for the other features with significant p-values observed. In addition, it also finds that the economic service level can be said to be significant to some extent due to its p-value being very close to 0.1. The meaning of significance and the corresponding implication of above contextual features will be discussed in depth in the next section.
|Items||Coefficients||Standard error||T value||p-Value|
|WAP building phases||0.03||0.41||0.07||0.95|
|Male dependency ratio||−3.79||5.60||−0.68||0.50|
|Female dependency ratio||−9.89||4.12||−2.40||0.02*|
|Public service level||−0.21||0.22||−0.97||0.33|
|Network environment level||−0.66||0.37||−1.78||0.07*|
|Education service level||0.06||0.20||0.31||0.75|
|Medical social benefit level||−0.05||0.19||−0.25||0.81|
|Economic service level||0.25||0.16||1.63||0.10|
|Policy cooperation level||0.62||0.27||2.31||0.02*|
|Cluster 1| Cluster 2||−109.81||0.04||−2660.68||0.00***|
|Cluster 2| Cluster 3||−107.41||0.42||−255.70||0.00***|
5 Discussion and Conclusion
In this study, the contribution of the results can be divided into the exploration of the cause of digital divide and the design of innovative research. First, this study attempts to explore the determinants of the digital divide, hoping to clarify the causes of the digital divide, and then conceive various measures in response (Cawkell 2001; Goncalves, Oliveira, and Cruz-Jesus 2018; van Dijk 2006). To this end, we attempt to construct a framework for evaluating digital divide, using the records of Wi-Fi Internet access by residents in remote tribes as the analysis content, taking into account regional contextual features, and applying clustering analysis, multivariate analysis of variance, and ordinal logistic regression to provide an integrated and operative analytical framework as a reference for future researchers engaged in digital research.
Secondly, we also echo the views of previous researchers that the causes of digital divide are quite complex, and cannot be explained with a single point of view (Ferro, Helbig, and Gil-Garcia 2011; van Dijk and Hacker 2003). To this end, this study attempts to focus on the contextual features of the area, that is, when considering the digital divide issue, in addition to confirming that if the local information infrastructure is complete, or if the residents have the ability to access information, contextual features should also be included in the consideration at the same time. According to the empirical results, it shows that (1) if the locations of Taiwan’s remote tribes are examined, the impact of the south and the north (longitude) is not significant, and the impact of the east and the west (latitude) is significant, that is, in terms of the status of information access, the residents in the western part of Taiwan’s remote tribes have a clear advantage over those in the eastern part. The possible cause is that Taiwan’s economic activities are still dominated by the west, and most of the major cities are also located in the west, where the manufacturing and service industries are relatively more developed, providing many opportunities for information application, such as employment, education, medical care, and entertainment. Therefore, the residents in the western part of Taiwan’s remote tribe have stronger willingness to access information; (2)when the female dependency ratio is lower, it shows that the dependent population (less than 15 years old or older than 65 years of age) is lower than the working population (between 15 and 65 years of age), which indicates the burden of raising the female dependent population in the remote tribes is relatively small, and the composition of the population is mainly the young and the middle-aged. At this point, there is a higher demand for information access; (3) when the network infrastructure of the remote tribes is more incomplete, it shows that the construction of the Internet facilities is not perfect, and the scope covered is not comprehensive enough. Under the condition that good Internet quality cannot be provided, the residents of the remote tribes are naturally prone to apply i-Tribe system to access the network; (4) when the level of economic development of the remote tribes is higher, with the rise of local characteristic industries such as sightseeing, agricultural products, hot springs or cultural creativity, it will enhance the residents’ willingness to introduce e-commerce, and increase the popularity of products or sights through flexible marketing methods, which will help to improve the residents’ access of information; (5) if the residents of the remote tribes have a higher level of cooperation with the government’s policies, and reduce the obstacles of the government to promote digitalization, it will obviously accelerate the promotion of various network infrastructures, so that people can enjoy the benefits brought by the digitalization, and naturally improve their willingness to access information.
Thirdly, the results obtained by this study can also serve as an important reference in supporting the government’s role in formulating a digital policy. The government has successively promoted several projects to try to increase the level of digitalization in the remote tribes, so that the network infrastructure will gradually spread in the remote tribes and enhance the residents’ access for information, and achieve the goals of information circulation, job creation, economic development, and better quality of life. Physical access to the Internet is preceded by the motivation, attitude, and expectation of getting physical access. It is not a single decision to adopt and purchase a particular technology but a continuing process of getting access to new versions of hardware and software, peripheral equipment, and subscriptions. Obtaining physical access only makes sense when people are provided with an education of technology. Thus, certain technology skills and competencies would result in the access determination. It is clear that the purpose of access and the final goal of appropriating the technology will be looked for the actual usage, which is significantly relevant to the digital divide. According to the research results, there are indeed some digitalization in the remotes areas that have been improved, showing the effectiveness of the implementation of the policy. However, we have also found that some of the remote tribes still need continuous efforts, and are still waiting the government for a more effective digitalization policy. We hope to provide the results of the research to the government as a reference to develop a digitalization policy, take into consideration the contextual features, development conditions and digitalization levels of each remote tribe, and select the most appropriate policy tools. Thus, make the limited budget more effective and respond to the needs of the public and the legislature for accountability.
Fourth, in the design of research methods, different from the previous discussion on digital divide issues, the researcher tried to use questionnaires (Bornman 2016; Hsieh, Rai, and Kei 2008), interviews or secondary materials (Park, Choi, and Hong 2015; Silva, Badasyan, and Busby 2018) as a measure of digital divide. Considering the terrain restrictions (mainly mountainous) in various remote tribes of Taiwan, the traffic is inconvenient and affects the local economic development. Due to the high cost of constructing network facilities, it directly affects the local residents’ access to information. In order to accurately evaluate the digitization of Taiwan’s remote tribes, this study chose to use the WAP set up in each remote tribe to obtain the record of residents using Wi-Fi Internet access, and tried to use 6Hs to evaluate the level of digital divide in the area under the consideration of reliability and validity. Applying wireless networks as a basis for assessing the level of digitization not only provides an innovative method of data collection, but also more accurately reflects the public’s access to information. The government authorities can collect the network usage status of each remote tribe and promote different types of digitalization policies for different situations in each region, such as strengthening the construction of basic network infrastructures or improving the information literacy of residents, and effectively eliminating the digital divide faced by each remote tribe.
Finally, in the choice of analysis units, different from the previous researchers’ approach, the digital divide is discussed at the level of country, enterprise, household and individual (Bartikowski et al. 2018; Chinn and Fairlie 2007; Chipeva et al. 2018). This study chose regional as the unit of analysis to explore the digital divide problem faced by Taiwan’s remote tribes, and joins the discussion of contextual features to explore the impact of digitalization from different perspectives. In previous studies, most researchers cut through macro perspectives, conducted cross-country comparisons, explored through different institutional environments in each country, and explained the reasons for the difference in the level of digitization in various countries using other factors such as politic, economic, society, culture, or basic infrastructure (Szeles 2018; van Dijk and Hacker 2003). There were also researchers who studied the digital divide of a single country from a micro perspective, trying to explain the ability of different people to access information by gender, age, class, race, occupation, and education (Chaudhuri, Flamm, and Horrigan 2005; Colley and Maltby 2008; Lindblom and Räsänen 2017). This study chose the region as the research object, using Meso as the starting point of thinking, considering the contextual features, development conditions and the level of digitization of each remote tribe, and expected that the most appropriate policy tools can be selected according to the characteristics of each region.
5.2 Limitations and Future Research
There is also something to be desired in this study, and for this reason we tried to propose research limitations and future research directions. First, we attempt to respond to the researchers’ suggestions and choose regional as an analysis unit for discussing the digital divide issue (Gray, Gainous, and Wagner 2017; Lucendo-Monedero, Ruiz-Rodríguez, and González-Relañ 2019). In order to seize the relevance between each remote tribe’s contextual features and digital divide, we tried to combine different data sources, such as geographic location obtained, data period, male/female dependency ratio, network environment, education services, medical care, social welfare, economic development, and policy coordination. The empirical research shows that some of the contextual features are indeed related to digital divide, and can be used as a reference for the government to formulate a digitalization policy. However, limited by the current data type and analysis tools, we can only test the relevance (positive or negative) of each contextual feature to digital divide. However, whether there is a causal relationship between them is still not validated. It is recommended that future researchers conduct more in-depth exploration of contextual features and understand whether there is a causal relationship between each contextual feature and digital divide, or if contextual features have an interactive effect of interaction.
Secondly, as many researchers have suggested, the discussion of digital divide should be based on inter-disciplinary, multi-level or multi-dimensional perspectives (Ferro, Helbig, and Gil-Garcia 2011; Lindblom and Räsänen 2017; Szeles 2018). This is because the causes of the digital divide problem are quite complex, and the impact is not only huge but also covers many levels, so researchers need to think in more diverse perspectives. It can easily be observed that the causes of digital divide concern at least information development, communication technologies, computer science, economics, psychology, sociology, and so on. Therefore, how to develop an integrated framework to explore the cause and influence of digital divide from a systematic perspective will be an important research topic in future study.
Finally, in the observation mechanism for constructing the digital divide situation, it is recommended that follow-up researchers can refer to the research design used in this study to regularly collect various factors related to the digital divide phenomenon, and use long-term longitudinal research to replace short-term cross-sectional studies (Bornman 2016; Nishijima, Ivanauskas, and Sarti 2017; van Dijk 2006), presenting evolution process in each region at different points in time, to provide more practical information to the relevant authorities, improve the decision-making quality of the digitalization policy, and indeed reduce the negative impact of the digital divide problem.
Today, with the continuous advancement of information technology, people can easily obtain information through various vehicles and at the same time give higher value to information through the acquisition, dissemination, reorganization, creation, and application of information. When people are benefiting from the digitalization, the problem of digital divide is also caused, which leads to the difference in the ability of information access between countries, regions or individuals, producing another form of inequality, and causes a negative impact on economic and social development. In order to explore the digital divide problem faced by remote tribes in Taiwan, we select Taiwan’s remote tribes as the research case, trying to use the region as the analysis unit to study the contextual features in depth, and constructed an integrated analysis framework. Data-driven, evidence-based, and bottom-up measures are the core spirits to seize a deeper understanding of the determinants that lead to digital divide, while thinking about what to do in response. We look forward to using operational experience as an example to provide experience to countries with digital divide issues, supporting countries in formulating better digitalization policies, and reducing the problems that may arise from information inequality.
Funding source: Ministry of Science and Technology in Taiwan
Award Identifier / Grant number: MOST-110-2221-E-182-049
Award Identifier / Grant number: MOST-110-2218-E-218-001-MBK
This work was supported by the Ministry of Science and Technology in Taiwan under Grant MOST-110-2221-E-182-049 and Grant MOST-110-2218-E-218-001-MBK.
Agresti, A. 2002. Categorical Data Analysis. Hoboken New Jersey, US: John Wiley & Sons. Inc., Publication. https://mybiostats.files.wordpress.com/2015/03/3rd-ed-alan_agresti_categorical_data_analysis.pdf (accessed October 3, 2021).Search in Google Scholar
Bartikowski, B., M. Laroche, A. Jamal, and Z. Yang. 2018. “The Type-of-Internet-Access Digital Divide and the Well-being of Ethnic Minority and Majority Consumers: A Multi-Country Investigation.” Journal of Business Research 82: 373–80, https://doi.org/10.1016/j.jbusres.2017.05.033.Search in Google Scholar
Bornman, E. 2016. “Information Society and Digital Divide in South Africa: Results of Longitudinal Surveys.” Information, Communication & Society 19 (2): 264–78, https://doi.org/10.1080/1369118X.2015.1065285.Search in Google Scholar
Brandtzæg, P. B., J. Heim, and A. Karahasanović. 2011. “Understanding the New Digital Divide—A Typology of Internet Users in Europe.” International Journal of Human-Computer Studies 69, no. 3: 123–38, https://doi.org/10.1016/j.ijhcs.2010.11.004.Search in Google Scholar
Brock, G., V. Pihur, S. Datta, and S. Datta. 2011. “clValid: an R Package for Cluster Validation.” Journal of Statistical Software 25 (4): 1–22, https://doi.org/10.18637/jss.v025.i04.Search in Google Scholar
Bruce, C. S. 2000. “Information Literacy Research: Dimensions of the Emerging Collective Consciousness.” Australian Academic & Research Libraries 31 (2): 91–109, https://doi.org/10.1080/00048623.2000.10755119.Search in Google Scholar
Bruce, C. S. 2016. “Information Literacy Research: Dimensions of the Emerging Collective Consciousness. A Reflection.” Australian Academic & Research Libraries 47 (4): 239–44, https://doi.org/10.1080/00048623.2016.1248236.Search in Google Scholar
Chaudhuri, A., K. S. Flamm, and J. Horrigan. 2005. “An Analysis of the Determinants of Internet Access.” Telecommunications Policy 29 (9–10): 731–55, https://doi.org/10.1016/j.telpol.2005.07.001.Search in Google Scholar
Chinn, M. D., and R. W. Fairlie. 2007. “The Determinants of the Global Digital Divide: A Cross-Country Analysis of Computer and Internet Penetration.” Oxford Economic Papers 59 (1): 16–44, https://doi.org/10.1093/oep/gpl024.Search in Google Scholar
Chipeva, P., F. Cruz-Jesus, T. Oliveira, and Z. Irani. 2018. “Digital Divide at Individual Level: Evidence for Eastern and Western European Countries.” Government Information Quarterly 35 (3): 460–79, https://doi.org/10.1016/j.giq.2018.06.003.Search in Google Scholar
Christensen, R. H. B. 2015. Analysis of Ordinal Data with Cumulative Link Models—Estimation with the R-Package Ordinal. R-package version, 1–31. http://people.vcu.edu/∼dbandyop/BIOS625/CLM_R.pdf (accessed October 1, 2021).Search in Google Scholar
Chunghwa Telecom. 2021. 5G Rate Plans. https://english.taiwanmobile.com/product/5Grateplan.html (accessed October 1, 2021).Search in Google Scholar
Colley, A., and J. Maltby. 2008. “Impact of the Internet on Our Lives: Male and Female Personal Perspectives.” Computers in Human Behavior 24 (5): 2005–13, https://doi.org/10.1016/j.chb.2007.09.002.Search in Google Scholar
Cruz-Jesus, F., M. R. Vicente, F. Bacao, and T. Oliveira. 2016. “The Education-related Digital Divide: An Analysis for the EU-28.” Computers in Human Behavior 56: 72–82, https://doi.org/10.1016/j.chb.2015.11.027.Search in Google Scholar
Cuervo, M. R. V., and A. J. L. Menéndez. 2006. “A Multivariate Framework for the Analysis of the Digital Divide: Evidence for the European Union-15.” Information & Management 43 (6): 756–66, https://doi.org/10.1016/j.im.2006.05.001.Search in Google Scholar
Digi Taiwan. 2018. Expanding Telecommunications Infrastructure Helps Shrink Urban-Rural Gap. https://digi.taiwan.gov.tw/news/expanding-telecommunications-infrastructure-helps-shrink-urban-rural-gap/# (accessed October 2, 2021).Search in Google Scholar
Fagin, R., A. Lotem, and M. Naor. 2003. “Optimal Aggregation Algorithms for Middleware.” Journal of Computer and System Sciences 66 (4): 614–56, https://doi.org/10.1016/S0022-0000(03)00026-6.Search in Google Scholar
Ferro, E., N. C. Helbig, and J. R. Gil-Garcia. 2011. “The Role of IT Literacy in Defining Digital Divide Policy Needs.” Government Information Quarterly 28 (1): 3–10, https://doi.org/10.1016/j.giq.2010.05.007.Search in Google Scholar
Goncalves, G., T. Oliveira, and F. Cruz-Jesus. 2018. “Understanding Individual-Level Digital Divide: Evidence of an African Country.” Computers in Human Behavior 87: 276–91, https://doi.org/10.1016/j.chb.2018.05.039.Search in Google Scholar
Hsieh, J. J. P. A., A. Rai, and M. Keil. 2008. “Understanding Digital Inequality: Comparing Continued Use Behavioral Models of the Socio-Economically Advantaged and Disadvantaged.” MIS Quarterly 32 (1): 97–126, https://doi.org/10.2307/25148830.Search in Google Scholar
Javali, S. B., and P. V. Pandit. 2010. “A Comparison of Ordinal Regression Models in an Analysis of Factors Associated with Periodontal Disease.” Journal of Indian Society of Periodontology 14 (3): 155–9, https://doi.org/10.4103/0972-124X.75909.Search in Google Scholar
Lindblom, T., and P. Räsänen. 2017. “Between Class and Status? Examining the Digital Divide in Finland, the United Kingdom, and Greece.” The Information Society 33 (3): 147–58, https://doi.org/10.1080/01972243.2017.1294124.Search in Google Scholar
List, A. 2019. “Defining Digital Literacy Development: An Examination of Pre-service Teachers’ Beliefs.” Computers & Education 138: 146–58, https://doi.org/10.1016/j.compedu.2019.03.009.Search in Google Scholar
Lucendo-Monedero, A. L., F. Ruiz-Rodríguez, and R. González-Relaño. 2019. “Measuring the Digital Divide at Regional Level. A Spatial Analysis of the Inequalities in Digital Development of Households and Individuals in Europe.” Telematics and Informatics 41: 197–217, https://doi.org/10.1016/j.tele.2019.05.002.Search in Google Scholar
Mardikyan, S., E. A. Yildiz, M. D. Ordu, and B. Simsek. 2015. “Examining the Global Digital Divide: a Cross-Country Analysis.” Communications of the IBIMA: 592253, https://doi.org/10.5171/2015.592253.Search in Google Scholar
Nishijima, M., T. M. Ivanauskas, and F. M. Sarti. 2017. “Evolution and Determinants of Digital Divide in Brazil (2005–2013).” Telecommunications Policy 41 (1): 12–24, https://doi.org/10.1016/j.telpol.2016.10.004.Search in Google Scholar
Park, S. R., D. Y. Choi, and P. Hong. 2015. “Club Convergence and Factors of Digital Divide across Countries.” Technological Forecasting and Social Change 96: 92–100, https://doi.org/10.1016/j.techfore.2015.02.011.Search in Google Scholar
Pihur, V., S. Datta, and S. Datta. 2007. “Weighted Rank Aggregation of Cluster Validation Measures: A Monte Carlo Cross-Entropy Approach.” Bioinformatics 23 (13): 1607–15, https://doi.org/10.1093/bioinformatics/btm158.Search in Google Scholar
Puspitasari, L., and K. Ishii. 2016. “Digital Divides and Mobile Internet in Indonesia: Impact of Smartphones.” Telematics and Informatics 33 (2): 472–83, https://doi.org/10.1016/j.tele.2015.11.001.Search in Google Scholar
Sekula, M., S. Datta, and S. Datta. 2017. “optCluster: An R Package for Determining the Optimal Clustering Algorithm.” Bioinformation 13 (3): 101–3, https://doi.org/10.6026/97320630013101.Search in Google Scholar
Silva, S., N. Badasyan, and M. Busby. 2018. “Diversity and Digital Divide: Using the National Broadband Map to Identify the Non-adopters of Broadband.” Telecommunications Policy 42 (5): 361–73, https://doi.org/10.1016/j.telpol.2018.02.008.Search in Google Scholar
Szeles, M. R. 2018. “New Insights from a Multilevel Approach to the Regional Digital Divide in the European Union.” Telecommunications Policy 42 (6): 452–63, https://doi.org/10.1016/j.telpol.2018.03.007.Search in Google Scholar
Taiwan Executive Yuan. 2015. Taiwan Providing Free Wi-Fi in Indigenous Communities. https://english.ey.gov.tw/Page/61BF20C3E89B856/e8320767-808b-4ba4-96a9-2b4108d398a2 (accessed October 1, 2021).Search in Google Scholar
Taiwan Ministry of Labor. 2021. Taiwan Labor e-News Letter No. 55. https://english.mol.gov.tw/homeinfo/33506/47769/normalForSeason (accessed October 2, 2021).Search in Google Scholar
Tsai, W. Y., T. C. Chou, Y. H. Chen, and P. T. Jan. 2021. “Understanding Aboriginal Tribe Wireless Broadband Construction Trajectories through Actor-Network Theory Views.” Journal of Internet Technology 22 (1): 41–52, https://doi.org/10.3966/160792642021012201004.Search in Google Scholar
Van Dijk, J. A. G. M. 2017. “Digital Divide: Impact of Access.” In The International Encyclopedia of Media Effects. https://onlinelibrary.wiley.com/doi/10.1002/9781118783764.wbieme0043.Search in Google Scholar
Wu, T. F., M. C. Chen, Y. M. Yeh, H. P. Wang, and S. C. H. Chang. 2014. “Is Digital Divide an Issue for Students with Learning Disabilities?” Computers in Human Behavior 39: 112–7, https://doi.org/10.1016/j.chb.2014.06.024.Search in Google Scholar
Yu, T. K., M. L. Lin, and Y. K. Liao. 2017. “Understanding Factors Influencing Information Communication Technology Adoption Behavior: The Moderators of Information Literacy and Digital Skills.” Computers in Human Behavior 71: 196–208, https://doi.org/10.1016/j.chb.2017.02.005.Search in Google Scholar
© 2022 Ssu-Han Chen et al., published by De Gruyter, Berlin/Boston
This work is licensed under the Creative Commons Attribution 4.0 International License.