Networks Connecting the Digital Divide

Kateryna Savchyn, William Frankenstein


Broadband infrastructure availability across the United States is a complex (and political) issue. There are an estimated 19 million people in the US that without reliable internet access, and Congress along with USDA has allocated funds in upwards of $550 million in 2019 alone toward loan programs such as ReConnect that bring infrastructure to underserved communities. Congress is currently considering a $2 trillion infrastructure plan for achieving universal rural broadband. Past work in this area has focused on traditional econometric approaches to identify gaps in broadband access and design programs to address these shortcomings; however, these approaches do not consider the embedded relationship that providers typically have with areas of high economic activity. To address this challenge, this work combines network analysis methods with machine learning approaches to identify potential drivers of persistent digital divides. The primary data utilized to answer these questions is a bipartite provider by US county network, which is derived from FCC Form 477. This data details provider services by Census FIPS Code (block group level) which we aggregate to the county level. It is important to note that our outcomes are unavoidably optimistic due to the well-established over-estimation of the FCC data which identifies locations with a potential for access to service (not actual provision) as being serviced. We use this data along with a comprehensive list of US county codes in order to build a provider to community network that helps us identify key network providers and determine which counties lack access broadband services. The identification of isolate counties serves as a feature generation technique for classification where isolate and non-isolate counties becomes our response variable. To augment this data, we gather urban and rural classifications from USDA Rural-Urban Community RUC codes and build a provider to RUC code network to visualize the relationships between levels of urbanity and connectivity. To supplement this we utilize physical land features from the American Community Survey (ACS), as well as economic variables (e.g. Income, Unemployment, and Commuting Type). These county-descriptive features allow us to test hypotheses that levels of urbanity and county economic metrics influence broadband connectivity levels. Our findings show that the broadband coverage network has a substantially larger number of isolate counties when utilizing FCC-defined broadband speeds. In analyzing these isolate counties, we found that contrary to conversations regarding the digital divide, that degrees of urbanity as defined by USDA Rural-Urban Continuum Codes do not significantly impact the ability to identify isolates. Finally, although we find some economic variations between clusters of counties with high and low proportions of isolates, our results show that overall economic metrics regarding Income, Unemployment, Poverty, Industry, Occupation, and Commuting Type do not have predictive power in identifying isolates, inversely to previous research on adequacy of economic demographics’ ability to predict degrees of connectivity. These results combined with previous research imply that county land features, urbanity and economic indicators are not linked with connectivity as defined by FCC availability data.

← Schedule