On the efficiency of statistical assimilation techniques in the presence of model and data error