The 'natural refuge strategy" for delaying insect resistance to transgenic cotton that produces insecticidal proteins from Bacillus thuringiensis (Bt) relies on refuges of host plants other than cotton that do not make Bt toxins. We tested this widely adopted strategy by comparing predictions from modeling with data from a four-year field study of cotton bollworm (Helicoverpa armigera) resistance to transgenic cotton producing Bt toxin Cry1Ac in six provinces of northern China. Bioassay data revealed that the percentage of resistant insects increased from 0.93% in 2010 to 5.5% in 2013. Modeling predicted that the percentage of resistant insects would exceed 98% in 2013 without natural refuges, but would increase to only 1.1% if natural refuges were as effective as non-Bt cotton refuges. Therefore, the results imply that natural refuges delayed resistance, but were not as effective as an equivalent area of non-Bt cotton refuges. The percentage of resistant insects with nonrecessive inheritance of resistance increased from 37% in 2010 to 84% in 2013. Switching to Bt cotton producing two or more toxins and integrating other control tactics could slow further increases in resistance.