Frequency downshift (FD) in wave trains on deep water occurs when a measure of the frequency, typically the spectral peak or the spectral mean, decreases as the waves travel down a tank or across the ocean. Many FD models rely on wind or wave breaking. We consider seven models that do not include these effects and compare their predictions with four sets of experiments that also do not include these effects. The models are the (i) nonlinear Schrodinger equation (NLS), (ii) dissipative NLS equation (dNLS), (iii) Dysthe equation, (iv) viscous Dysthe equation (vDysthe), (v) Gordon equation (Gordon) (which has a free parameter), (vi) Islas-Schober equation (IS) (which has a free parameter), and (vii) a new model, the dissipative Gramstad-Trulsen (dGT) equation. The dGT equation has no free parameters and addresses some of the difficulties associated with the Dysthe and vDysthe equations. We compare a measure of overall error and the evolution of the spectral amplitudes, mean, and peak. We find: (i) The NLS and Dysthe equations do not accurately predict the measured spectral amplitudes. (ii) The Gordon equation, which is a successful model of FD in optics, does not accurately model FD in water waves, regardless of the choice of free parameter. (iii) The dNLS, vDysthe, dGT, and IS (with optimized free parameter) models all do a reasonable job predicting the measured spectral amplitudes, but none captures all spectral evolutions. (iv) The vDysthe, dGT, and IS (with optimized free parameter) models do the best at predicting the observed evolution of the spectral peak and the spectral mean. (v) The IS model, optimized over its free parameter, has the smallest overall error for three of the four experiments. The vDysthe equation has the smallest overall error in the other experiment.