Econometric studies of the effects of research on productivity have typically imposed arbitrary restrictions on the length and shape of the R&D lag profile. These restrictions are likely to have biased up both the measured effects of R&D on productivity and the estimated rates of return to research. This paper argues that the useful stock of public knowledge depreciates, if at all, only gradually, and we use this notion to develop a new model, which we test using data on aggregate U.S. agriculture. We reject the conventional specification in favor of a more flexible, dynamic, alternative model, in which the impact of R&D on productivity lasts much longer than in previous studies. Consequently, the real, marginal rate of return to public agricultural R&D in the United States is estimated to be less than 10 percent per annum, much smaller than the typical rates of return reported in scores of previous studies, based on conceptually flawed and inappropriately restrictive dynamic specifications. We show that conventional approaches using the same data would have resulted in a much greater (biased) estimate of the rate of return to research.