Conference Agenda

Federal broadband initiatives such as the Broadband Equity, Access, and Deployment (BEAD) program prioritize infrastructure expansion to close the digital divide. However, substantial adoption gaps persist even in areas where broadband service is technically available. A central policy challenge is distinguishing between infrastructure-driven non adoption and household level constraints such as affordability, device access, and socioeconomic disadvantage. This distinction is particularly difficult when infrastructure data are incomplete, lagged, or subject to dispute, as has been the case with recent broadband mapping efforts.

This paper asks the following research question: Can nationwide demographic data alone reveal structurally distinct pathways to broadband non adoption that are relevant for targeting communications policy interventions?

To address this question, the study uses ZIP code level data from the American Community Survey 2019 to 2023, covering 14,685 U.S. ZIP codes. Fifteen demographic features are used as inputs, excluding broadband adoption rates from the representation stage to avoid leakage. The methodological approach applies a supervised representation learning framework in which an autoencoder compresses demographic features into a lower dimensional latent space, while a prediction head regularizes the latent representation toward adoption-relevant structure.

Unlike conventional regression models that estimate marginal effects of individual predictors, the latent representation approach captures bundled structural patterns within the demographic data. This allows the model to separate composite disadvantage from more proximate access constraints without imposing strong interaction specifications or functional form assumptions.

The analysis identifies two latent dimensions that exhibit strictly monotonic relationships with broadband adoption and together explain 53 percent of cross ZIP variation. The first dimension predominantly reflects rural ZIP codes. Although total population is its strongest correlate, population alone shows no meaningful univariate relationship with adoption rates. This suggests that rural status functions as a proxy for bundled structural disadvantages rather than as a direct determinant of non adoption. The latent structure captures the combination of lower population density, lower income, lower educational attainment, and higher rates of device scarcity that commonly characterize rural areas. Rurality therefore operates as a composite structural condition rather than as a single predictive variable.

The second latent dimension captures device poverty. In this dimension, lack of computing devices and income constraints are strongly associated with lower subscription rates, consistent with a more proximate household level access pathway.

Importantly, the latent structure also identifies ZIP codes that demographic indicators would predict to have relatively high adoption but that underperform in practice. These cases suggest the presence of infrastructure, affordability, or market structure barriers not directly observable in demographic data alone. Conversely, some ZIP codes with high composite disadvantage exhibit higher than expected adoption, potentially reflecting successful local interventions or competitive market conditions.

The disciplinary foundations of this work draw from communications policy, applied economics, and computational social science. The contribution is methodological and policy oriented. Methodologically, it demonstrates how representation learning can be used to uncover structural bundles of disadvantage that are difficult to isolate using standard regression approaches without extensive interaction modeling. Substantively, it provides a nationwide ZIP level analysis that distinguishes between intervention pathways relevant to broadband policy.

The findings suggest that broadband non adoption operates through at least two structurally distinct mechanisms requiring different policy responses. High composite disadvantage areas may require infrastructure investment combined with wraparound support, while device poor areas may benefit more immediately from affordability subsidies and digital literacy programs. By separating these pathways using demographic information alone, the framework offers a practical triage tool for prioritizing broadband resources when infrastructure availability information is unavailable, lagged, or disputed.