Predicting changes in brain connectivity between anatomical regions is essential for brain mapping and neural disorder diagnosis across different age spans from a limited data (e.g., single timepoint). Such learning tasks become more difficult when handling a single dataset with missing timepoints, let alone multiple decentralized datasets collected from different hospitals and with varying incomplete acquisitions. With the new paradigm of federated learning (FL) one can learn from decentralized datasets without data sharing. However, to the best of our knowledge, no FL method was designed to predict time-dependent graph data evolution trajectory using non-iid training longitudinal datasets with varying acquisition timepoints. In this paper, we aim to significantly boost the predictive power of data owners (e.g., local hospitals) trained with several missing timepoints while benefiting from other hospitals with available timepoints in a fully data-preserving way. Specifically, we propose a novel 4D GNN federated architecture, namely 4D-FED-GNN+, which acts as a graph self-encoder when the next timepoint is locally missing or as a graph generator when the next timepoint is locally available in the training set. We further design a mixed federation strategy that alternates (i) GNN layer-wise weight aggregation at each timepoint and (ii) pairwise GNN weight exchange between hospitals in a random order. Our comprehensive experiments on both real and simulated longitudinal datasets show that overall 4D-FED-GNN+ significantly outperform locally trained models. Our 4D-FED-GNN+ code available at https://github.com/basiralab/4D-FED-GNN.