Although many college courses adopt online tools such as Q&A online discussion boards, there is no easy way to measure or evaluate their effect on learning. As a part of supporting instructional assessment of online discussions, we investigate a predictive relation between characteristics of discussion contributions and student performance. Inspired by existing work on dialogue acts, project-based learning, and instructional analysis of student-generated text in generating predictive models, we make use of dialogue roles, linguistic features, and work patterns. In particular, we model the Q&A dialog roles that participants play, emotional features covered by LIWC (Linguistic Inquiry and Word Count), cohesiveness of the dialogue, the coherence captured by Coh-Metrix, and temporal patterns of participation. We use a discussion corpus from eight semesters of a computer science course, covering conversations of 173 student groups (370 students). We first remove various noises in student discussion data and normalize the discussion data. We then apply machine learning techniques and text analysis tools for classifying dialogue features efficiently. The extracted dialogue and participation features are used as predictive variables for project grades. The correlation and regression analyses indicate that the number of answers provided to others, the number of positive emotion expressions, and how early students communicate their problems before the deadline correlate with project grades. This finding confirms the argument that in assessing student online activities, we need to capture how they interact, not just how often they participate.