The effects on 36 h predictions of cyclogenesis of varying the initial upper air data while keeping the initial surface pressures unchanged have been studied. The numerical models employed use the primitive equations, have eight levels, utilize sigma coordinates, and include orography, surface and internal friction, large-scale release of latent heat, long-wave radiation from clouds and water vapor, and input of heat and water vapor from water surfaces. Perturbations to initial upper air data were made both randomly and in such a way as to simulate deleting half the radiosondes. Three cases were investigated. After 36 h the root-mean-square temperature differences between perturbed and unperturbed prognoses were the same order of magnitude as the initial difference. The error in the two cases which attempted to simulate the initial stages of cyclogenesis was not much affected by reducing the density of initial upper air temperature data. However, the case which started after cyclogenesis had begun was well forecast and its prognosis was significantly degraded by deletion of data.