This paper considers the problem of determining which set of 2^p leaf nodes on a binary multiscale tree model of depth N (N>p) gives the best linear minimum mean-squared estimator of the tree root. We find that the best-case and worst-case sampling choices depend on the correlation structure of the tree. This problem arises in Internet traffic estimation, where the goal is to estimate the average traffic rate on a network path based on a limited number of traffic samples.