This paper investigates the phenomenon of covariance matrix underestimation leading to possibly misleading inference which can arise from a potential pitfall in approximate GLS estimation of the regression model with ARMA disturbances.

Koreisha and Pukkila (1990a) have recently proposed a fast and efficient GLS estimator for the univariate ARMA time series model which appears to be far more robust than maximum likelihood methods and of comparable accuracy. The one drawback to this new estimator is that it requires use of the...