![[ Visit Client Website ]](images/banner.gif)
Schedule:
Sunday, January 24, 2027
Continental 1-4 (Hilton San Francisco Union Square)
* noted as presenting author
Background and Purpose: Broadband access to high-speed internet is a critical determinant of health, education, and social connectedness in the U.S., with household subsidy programs serving as an important tool for community wellbeing. As part of the Infrastructure Investment and Jobs Act of 2021, the Affordable Connectivity Program (ACP) provided monthly income-based discounts on home internet bills and was adopted by approximately half of all qualifying households in the U.S. from 2021 until its termination in 2024. Gaps remain in how to improve rates of adoption and to consider how federally administered benefits can be tailored to county-, state-, and region-specific implementation for broadband uptake. Using a geographically weighted random forest model, this study asks the following question: What predictors are most important in broadband benefit adoption of the ACP, and how do global and local patterns compare to each other?
Methods: Cross-sectional data on 22 predictors of ACP adoption were compiled within geospatial centroids of most zip codes in the U.S. (N = 20547). We chose neighborhood-level predictors from previous evaluations of the ACP and Census, administrative, and public survey datasets on sociodemographic, wealth, and referral (e.g., free and reduced school lunch provider, phone company provider) predictors, all aggregated by zip code. We built a random forest model that was calibrated with bandwidth and local weighting to generate feature importance scores for every local (i.e., zip code centroid) submodel. Finally, we generated partial dependence plots to visualize marginal effects of the top three predictors on ACP adoption.
Results: The global model explained 77% of variance in ACP adoption. Government phone company referral, rent burden, and poverty were the top three predictors and explained 47% of variance. Local phone company referral (VI = 0.178) was the most important feature in the rural Northeast, rural Southeast, Appalachian, and Ozark regions. Local rent burden (VI = 0.162) was the most important feature in the Great Plains and Heartland regions. Local poverty (VI = 0.125) was the most important feature in metropolitan areas (e.g., Dallas, Chicago, Northeast coast) and non-metropolitan high-poverty areas (e.g., West Virginia). Positive relationships existed between each top three predictor and the outcome. Our model explained more variance (R² = 0.739) than that of previous geographically weighted regressions (R² = 0.289) using the same data.
Methods: Cross-sectional data on 22 predictors of ACP adoption were compiled within geospatial centroids of most zip codes in the U.S. (N = 20547). We chose neighborhood-level predictors from previous evaluations of the ACP and Census, administrative, and public survey datasets on sociodemographic, wealth, and referral (e.g., free and reduced school lunch provider, phone company provider) predictors, all aggregated by zip code. We built a random forest model that was calibrated with bandwidth and local weighting to generate feature importance scores for every local (i.e., zip code centroid) submodel. Finally, we generated partial dependence plots to visualize marginal effects of the top three predictors on ACP adoption.
Results: The global model explained 77% of variance in ACP adoption. Government phone company referral, rent burden, and poverty were the top three predictors and explained 47% of variance. Local phone company referral (VI = 0.178) was the most important feature in the rural Northeast, rural Southeast, Appalachian, and Ozark regions. Local rent burden (VI = 0.162) was the most important feature in the Great Plains and Heartland regions. Local poverty (VI = 0.125) was the most important feature in metropolitan areas (e.g., Dallas, Chicago, Northeast coast) and non-metropolitan high-poverty areas (e.g., West Virginia). Positive relationships existed between each top three predictor and the outcome. Our model explained more variance (R² = 0.739) than that of previous geographically weighted regressions (R² = 0.289) using the same data.
Conclusion and Implications: Broadband benefit adoption shows strong regional patterns. The non-parametric approach of random forest modeling is well-suited for regressing non-linear relationships between predictor and outcome variables in ACP enrollment data. Future programs for internet as a common good may consider partnerships and data sharing with existing service providers as an effective tool for improving adoption in hard-to-reach communities. Not only is broadband access a priority during times of public health emergency (e.g., COVID-19 pandemic), but early and sustained access over time also improves community access to health, education, and social connectedness.