Subspace identification methods, such as N4SID, MOESP, CVA etc have proven to be very successful for identification of multivariable, linear dynamic systems. These methods are associated with a number of design variables, or user choices. These include prediction horizons, weighting matrices and ways to perform the estimation. It is known that these choices may have a substantial influence on the model quality, at the same time as there is no comprehensive theory for rational decision making. In this contribution we study certain aspects of the choices, in particular their influence on both bias and variance. We also illustrate some aspects using a larger example, the Tennessee-Eastman identification challenge.