Initializing an ensemble prediction system requires an estimate of analysis uncertainty. In the future, we expect the deployment of ensemble Kalman filters will result in the merging of ensemble prediction and data assimilation, automatically providing estimates of analysis uncertainity. In the interim, we are investigating methods of generating initial conditions for ensemble prediction which involve running multiple parallel data assimilation cycles, with the current NCEP operational spectral statistical interpolation (SSI). Currently, both NCEP and ECMWF use dynamical methods to generate perturbations to a control analysis which have grown rapidly in the recent past (NCEP), or will grow rapidly during the subsequent forecast (ECMWF). Both of these methods suffer from the requirement that an aribtrary scaling must be applied at each analysis time to define the amplitude of the initial perturbations. Therefore, these methods cannot represent day-to-day changes in the amplitude of analysis error. By running parallel perturbed data assimilation cycles to generate initial conditions for ensemble prediction, we hope to capture day-to-day variations in analysis error amplitude, and thus produce better ensemble forecast probability distributions. Currently, the Canadian Meteorological Center (CMC) uses such a method to generate ensemble initial conditions. Both the observations and the assimilating model are perturbed in each parallel assimilation cycle of the CMC system. We will present results with the NCEP system for which only the data assimilation, not the assimilating model, is perturbed. Among the issues which will be discussed are the effect of the background error statistics on the character of the analysis perturbations which result from perturbing observations in each parallel assimilation cycle.