This paper introduces an approach to analyzing temporal patterns of knowledge construction (KC) in online discussions, including consequences of role assignments. The paper illustrates the power of this approach for illuminating collaborative processes using data from a semester-long series of discussions in which 21 university students were assigned weekly roles. The KC contributions of all 252 posts in the discussion were coded using a five phase scheme (Gunawardena et al. 1997). Then, statistical discourse analysis was applied to identify segments of discussion characterized by particular aspects of KC, and "pivotal posts"--those posts which initiated new segments of discussion. Finally, the influences of assigned student roles on pivotal posts and KC were modeled. The results indicate that most online discussions had a single pivotal post separating the discussion into two distinct segments: the first dominated by a lower KC phase; the second dominated by a higher KC phase. This provides empirical evidence supporting the progressive nature of the KC process, but not the necessity of the full five-phase sequence. The pivotal posts that initiated later segments were often contributed mid-discussion by students playing one of two summarizing roles (Synthesizer and Wrapper). This suggests that assigning a summarizing role mid-discussion can aid group progress to more advanced phases of KC. Finally, in some discussion segments, the KC phase of a post was related to characteristics of the two preceding posts. Collectively, the results demonstrate the power of this temporal approach for investigating interdependencies in collaborative KC in online discussions.