We present a method of evaluating the immediate performance impact of user state misclassifications in spoken dialogue systems. We illustrate the method with a tutoring system that adapts to student uncertainty over and above correctness. First we define a ranking of user states representing local performance. Second, we compare user state trajectories when the first state is accurately classified versus misclassified. Trajectories are quantified using a previously proposed metric representing the likelihood of transitioning from one user state to another. Comparison of the two sets of trajectories shows whether user state misclassifications change the likelihood of subsequent higher or lower ranked states, relative to accurate classification. Our tutoring system results illustrate the case where user state misclassification increases the likelihood of negative performance trajectories as compared to accurate classification.