The within-groups estimator is inconsistent in dynamic panels with fixed T since the sample mean used to eliminate the individual effects from the lagged dependent variable is correlated with the error term. This paper suggests to eliminate individual effects from an AR(1) panel using backward means as an alternative to sample means. Using orthogonal deviations of the lagged dependent variable from its backward mean yields an estimator that is still inconsistent for fixed T but the inconsistency is shown to be negligibly small. A Monte Carlo simulation shows that this alternative estimator has superior small sample properties compared to conventional fixed effects, bias-corrected fixed effects and GMM estimators. Interestingly, it is also consistent for fixed T in the specific cases where (i) T = 2, (ii) the AR parameter is 0 or 1, (iii) the variance of the individual effects is zero.