In closed-loop simulations of the Amoco catalytic cracker under MPC, it was observed that the step response model identified at a low throughput operating point provided a better closed-loop performance when used on the high throughput operating point than a model identified at the latter operating point (Kalra and Georgakis, Ind. Engng Chem. Res. 33, 3063, 1994). The present paper reports on further simulation experiments and analysis that was performed to determine if it is the nonlinear nature of the process or deficiencies in model identification that caused the unexpected closed-loop behavior. When output error models identified from PRBS signals were used in the MPC algorithm this behavior was not observed, ruling out that nonlinearities might have been the cause. A closed-loop stability analysis of the process under MPC, with truth models as those derived via PRBS tests, confirms what is seen in simulations. A comparison of the models obtained from the two types of tests indicates that the step tests contained inadequate power to excite frequencies of interest to closed-loop control. Furthermore, the step test data from the high throughput operating point was adversely affected by the presence of noise cause by the numerical calculations in this complex model. The conclusion was that the use of step tests to directly determine the step response coefficients, as is commonly done, may not in some cases be a safe approach.