In this paper, federated learning (FL) over wireless networks is investigated. In each communication round, a subset of devices is selected to participate in the aggregation with limited time and energy. In order to minimize the convergence time, global loss and latency are jointly considered in a Stackelberg game based framework. Specifically, age of information (AoI) based device selection is considered at leader-level as a global loss minimization problem, while sub-channel assignment, computational resource allocation, and power allocation are considered at follower-level as a latency minimization problem. By dividing the follower-level problem into two sub-problems, the best response of the follower is obtained by a monotonic optimization based resource allocation algorithm and a matching based sub-channel assignment algorithm. By deriving the upper bound of convergence rate, the leader-level problem is reformulated, and then a list based device selection algorithm is proposed to achieve Stackelberg equilibrium. Simulation results indicate that the proposed device selection scheme outperforms other schemes in terms of the global loss, and the developed algorithms can significantly decrease the time consumption of computation and communication.