This paper considers non-cooperative and fully-distributed power-allocation for selfish transmitter-receiver pairs in shared unlicensed spectrum when normalized-interference to each receiver is uncertain. We model each uncertain parameter by the sum of its nominal (estimated) value and a bounded additive error in a convex set, and show that the allocated power always converges to its equilibrium, called robust Nash equilibrium (RNE). In the case of a bounded and symmetric uncertainty region, we show that the power allocation problem for each user is simplified, and can be solved in a distributed manner. We derive the conditions for RNE's uniqueness and for convergence of the distributed algorithm; and show that the total throughput (social utility) is less than that at NE when RNE is unique. We also show that for multiple RNEs, the social utility may be higher at a RNE as compared to that at the corresponding NE, and demonstrate that this is caused by users' orthogonal utilization of bandwidth at RNE. Simulations confirm our analysis.