In order to improve Large Vocabulary Continuous Speech Recognition (LVCSR) systems, it is essential to discover exactly how our current systems are underperforming. The major intellectual tool for solving this problem is error analysis: careful investigation of just which factors are contributing to errors in the recognizers. This paper presents our observations of the effects that discourse (i.e., dialog) modeling has on LVCSR system performance. As our title indicates, we emphasize the recognition error analysis methodology we developed and what it showed us as opposed to emphasizing development of the discourse model itself. In the first analysis of our output data, we focussed on errors that could be eliminated by Dialog Act discourse tagging using Dialog Act-specific language models. In a second analysis, we manipulated the parameterization of the Dialog Act-specific language models, enabling us to acquire evidence of the constraints these models introduced. The word error rate did not significantly decrease since the error rate in the largest category of Dialog Acts, namely Statements, did not significantly decrease. We did, however, observe significant error reduction in the less frequently occurring Dialog Acts and we report on the characteristic of the error corrections. We discovered that discourse models can introduce simple syntactic constraints and that they are most sensitive to parts of speech.