In the dynamic average consensus problem, agents in a communication network use information from their immediate neighbors to track the average of the group's time-varying inputs. Estimators based on the internal model principle solve this decentralized averaging problem with zero steady-state tracking error while providing robustness to network topology changes, agent failures, and communication faults. We develop a systematic process for designing these estimators. By formulating estimator synthesis as a robust control problem, we decouple the design process from specific networks. This formulation allows us to use an existing robust pole placement method to design estimators that meet performance specifications for a set of networks.