This paper presents a comprehensive model-building procedure for the building of STARMA (Space-Time Autoregressive Moving Average) models.
This procedure explicitly considers the possibility that 'G,' the covariance matrix of the innovations, might not be spherical. (G equals O-squared times I. The cases of diagonal 'G,' wherein the innovations are independent between sites but are allowed to have unequal variances, and general 'G,' wherein the innovations may in fact be contemporaneously correlated, are examined. Model identification, estimation, and diagnostic checking are described, and tests of the assumptions concerning the form of 'G' are incorporated into the modeling procedure. The procedure is illustrated through two examples. Figures, tables, and 18 references are included. (Author abstract modified)