Path-dependence supposes that divergent, predictable outcomes can be due to historical contingencies rather than intrinsic differences. It describes how institutions lock into one of several alternative stable states, and is related to canalization and regime shift in developmental biology, evolution, and ecology. However, there is a dearth of strong, mechanistic tests for path-dependence outside of experiments. Here, we use the extensive ecological and economic data available for global fisheries to test for path-dependence in harvest rates, which directly impact targeted stocks and indirectly alter ecosystems. We first show with a generic model that the basic dynamics of managed fisheries—including gradual institutional change, limited resource access, and single-stock rent maximization—predict path-dependent harvest rates that converge to one of two states (optimal rates or overfishing), but only when potential fish productivity is high and fishing is costly. We then used data from 217 fisheries worldwide to test the model, which explained a substantial portion of the variation in harvest rates (R2=0.29, p<1x10-4) and outperformed single equilibrium models and statistical models. Path-dependence implies that undesirable management policies can be hard to eliminate, but leaves hope for durable shifts toward desirable policies.