Abstract: A new robust structure selection method is proposed to deal with the identification problem of linear regression models with unknown but bounded (UBB) errors. The model-order determination criterion is based on maximizing the determinants' ratio of two ellipsoidal axis information matrixes obtained through repeating a recursive outer-bounding ellipsoid algorithm. Unlike the usual assumptions on noise, namely, independence, constant variance, or martingale difference properties, the assumption in this paper is of asymptotic independence. Strong consistency of the method is proved.