We study the impact of user association policies on flow-level performance in interference-limited wireless networks. Most research in this area has used static interference models (neighboring base stations are always active) and resorted to intuitive objectives such as load balancing. In this paper, we show that this can be counterproductive in the presence of dynamic interference that couples the transmission rates to users at various base stations. We propose a methodology to optimize the performance of a class of coupled systems and apply it to study the user association problem. We show that by properly inducing load asymmetries, substantial performance gains can be achieved relative to a load-balancing policy (e.g., 15 times reduction in mean delay). We present a practical, measurement based, interference-aware association policy that infers the degree of interference-induced coupling and adapts to it. Systematic simulations establish that both our optimized static and adaptive association policies substantially outperform various dynamic policies that can, in extreme cases, even be susceptible to Braess's paradox-like phenomena, i.e., an increase in the number of base stations can lead to worse performance under greedy association policies. Furthermore, these results are robust to changes in file-size distributions, large-scale propagation parameters, and spatial load distributions.