This paper compares digital and analog transmission schemes for wireless federated learning (FL) in resource-constrained networks. The study highlights the fundamental differences between the two approaches, focusing on their suitability for different scenarios. Digital transmission decouples communication from computation, leading to communication-limited performance, while analog transmission enables over-the-air computation (AirComp), improving spectrum utilization. However, analog transmission is less power-efficient and sensitive to channel state information (CSI) imperfections. The paper establishes a unified framework for evaluating FL performance under imbalanced device sampling, strict latency targets, and transmit power constraints. It analyzes the convergence behavior of FL under both schemes, showing that analog transmission achieves a better optimality gap with more devices, but is affected by computation errors due to imperfect CSI. The study also optimizes device sampling for both schemes and validates the results through numerical simulations. The findings indicate that digital transmission has better power utilization, while analog transmission is more spectrum-efficient. The paper concludes that both schemes are effective for wireless FL, but their performance depends on the specific network conditions and requirements.This paper compares digital and analog transmission schemes for wireless federated learning (FL) in resource-constrained networks. The study highlights the fundamental differences between the two approaches, focusing on their suitability for different scenarios. Digital transmission decouples communication from computation, leading to communication-limited performance, while analog transmission enables over-the-air computation (AirComp), improving spectrum utilization. However, analog transmission is less power-efficient and sensitive to channel state information (CSI) imperfections. The paper establishes a unified framework for evaluating FL performance under imbalanced device sampling, strict latency targets, and transmit power constraints. It analyzes the convergence behavior of FL under both schemes, showing that analog transmission achieves a better optimality gap with more devices, but is affected by computation errors due to imperfect CSI. The study also optimizes device sampling for both schemes and validates the results through numerical simulations. The findings indicate that digital transmission has better power utilization, while analog transmission is more spectrum-efficient. The paper concludes that both schemes are effective for wireless FL, but their performance depends on the specific network conditions and requirements.