This paper addresses the problem of computing the sample variance of datasets scattered across a network of interconnected agents. A general procedure is outlined to allow the agents to reach consensus on the variance of their local data, which involves two cascaded (dynamic) average consensus protocols. Our implementation of the procedure exploits the distributed ADMM, yielding a distributed protocol that does not involve the sharing of any local, private data nor any coordination of a central authority; the algorithm is proved to be convergent with linear rate and null steady-state error. The proposed distributed variance estimation scheme is then leveraged to tune personalization in "personalized learning" where agents aim at training a local model tailored to their own data, while still benefiting from the cooperation with other agents to enhance the models’ generalization power. The degree to which an agent tailors its local model depends on the diversity of the local datasets, and we propose to use the variance to tune personalization. Numerical simulations test the proposed approach in a classification task of handwritten digits, drawn from the EMNIST dataset, showing the better performance of variance-tuned personalization over non-personalized training.