To batch or not to batch?

Abstract

When designing steady-state computer simulation experiments, one may be faced with the choice of batching observations in one long run or replicating a number of smaller runs. Both methods are potentially useful in the course of undertaking simulation output analysis. The tradeoffs between the two alternatives are well known: batching ameliorates the effects of initialization bias, but produces batch means that might be correlated; replication yields independent sample means, but may suffer from initialization bias at the beginning of each of the runs. We present several new results and specific examples to lend insight as to when one method might be preferred over the other. In steady-state, batching and replication perform similarly in terms of estimating the mean and variance parameter, but replication tends to do better than batching with regard to the performance of confidence intervals for the mean. Such a victory for replication may be hollow, for in the presence of an initial transient, batching often performs better than replication when it comes to point and confidence-interval estimation of the steady-state mean. We conclude---like other classic references---that in the context of estimation of the steady-state mean, batching is typically the wiser approach.