When identifying rational polynomial transfer functions from input-output data, estimates of the coefficients and their covariances can be obtained. However, the estimated value of any given coefficient depends considerably on the chosen model order, leading to segmented confidence regions when the model order is not known. Fortunately the frequency responses of the transfer function models are little affected by using too high a model order. These are more likely to have a consistent estimate with a single surrounding confidence region, In this paper, the confidence limits established for the frequency response of single-input-single-output systems by Edmunds are extended to multivariable systems. A statistical lest of the method is then carried out using an illustrative example.