We consider the state and disturbance estimation problem for a linear system with uncertainties as additive deterministic disturbances. Instead of attempting to minimize the effects of the disturbance as in the robust filters or to decouple the disturbance as in the unknown input observers, it is proposed to estimate the disturbance and to use it to reduce the model error and thus to improve the state estimation. This technique is denoted as model error compensator (MEG). Sufficient conditions for achieving bounded disturbance estimation error are presented. These conditions are different from those required in the unknown input observers and the robust filter for state estimation only. In addition, no differentiation of the measured signals is required. This paper also discusses how existing state observers can be used with the MEC as a two-step observer design strategy.