This paper provides an explanation for the puzzling phenomenon in Tieslau et al. (1996, Journal of Econometrics 71, 249}264) that a substantial e$ciency loss occurs if low-order autocorrelations are omitted when estimating the di!erencing parameter, d. This is because for all n strictly bigger than 1, the nth-order autocorrelation function does not depend uniquely on the di!erencing parameter. ...