Broadband is a foundational infrastructure, yet in the United States, it remains unaffordable for millions of households and uncompetitive in many markets, prompting historic public investment. Two flagship initiatives, the Affordable Connectivity Program (ACP) and the Broadband Equity, Access, and Deployment (BEAD) program, form the core of the federal broadband policy. However, the welfare and market structure effects remain poorly understood. Using rich data on markets, providers, and broadband products, I combine a difference-in-differences design with continuous treatment intensity and a structural model of demand, firm entry, and product offerings to quantify the impact of these programs. The ACP, a $30 monthly subsidy to households, raises social surplus by 7.6-8.8% and yields up to $1.90 in welfare per dollar spent. The BEAD program, which subsidizes providers’ fixed costs, expands entry (13.9–19.5%) and p roduct variety (12.6–16.7%), but also increases marginal costs and prices by up to 18%. Welfare gains diminish at higher subsidy levels, with the largest benefits in the 25–50% range. The findings indicate a key trade-off: consumer-side subsidies deliver immediate and cost-effective affordability gains, while infrastructure subsidies reshape market competition to generate long-run efficiency, albeit at a higher fiscal cost.

This paper introduces a new method that integrates machine learning techniques with traditional econometric tools to correct bias in two-step estimation for dynamic discrete choice models under incomplete information. Unlike most existing approaches that primarily address bias from conditional choice probabilities while taking the transition matrix as given, this method simultaneously addresses both sources of bias. I propose a two-step estimator that jointly includes the conditional choice probabilities and transition matrix as first-step estimates, providing a robust, non-iterative solution. The estimator achieves square root N consistency and asymptotic efficiency. In addition, I establish a simple connection between Neyman orthogonality and the Zero-Jacobian property, a key result that guarantees consistent and efficient estimators in single-agent models. Monte Carlo simulations using a common benchmark data-generating process demonstrate the good performance of the proposed method.