This paper presents a method of identifying a basis set of error parameters in robot calibration using the Singular Value Decomposition (SVD) method. With the method, the error parameter space can be separated into two: observable subspace and unobservable one. As a result, for a defined position error model, one can determine the dimension of the observable subspace, which is vital to the estimation of error parameters. The second objective of this paper is to study, when unmodeled error exists, the implications of measurement configurations in robot calibration. For selecting measurement configurations in calibration, and index is defined to measure the observability of the error parameters with respect to a set of robot configurations. As the observability index increases, the attribution of the position errors to the parameters becomes dominant and the effects of the measurement and unmodeled errors become less significant; consequently better estimation of the parameter errors can be obtained.