A comparison of confidence region estimators for multivariate simulation output

Three previously proposed methods for constructing joint confidence regions on the mean of multivariate simulation output are described and tested using data generated by Gaussian vector autoregressive moving average models. The methods that use an estimate of the variance-covariance matrix of the data are found to yield regions with lower volumes than the method that does not use an estimate of the variance-covariance matrix. The experimental design included the factors run length, dimension, autocorrelation and cross correlation.