A method which determines the number of replications in the simulation is proposed, particularly for small-sample comparison of estimators. This method takes the smallest number of replications that makes the difference of mean square errors be statistically significant and provides an efficient algorithm for calculating the standard error of the mean square error. Two examples are illustrated, the first one is on comparison of mean and median ; the second, the Kaplan-Meier type and Buckley-James type estimators of a quantile function with censored data.