Per definition, CSCL research deals with the data of individuals nested in groups, and the influence of a specific learning setting on the collaborative process of learning. Most well-established statistical methods are not able to analyze such nested data adequately. This article describes the problems which arise when standard methods are applied and introduces multilevel modelling (MLM) as an alternative and adequate statistical approach in CSCL research. MLM enables testing interactional effects of predictor variables varying within groups (for example, the activity of group members in a chat) and predictors varying between groups (for example, the group homogeneity created by group members' prior knowledge). So it allows taking into account that an instruction, tool or learning environment has different but systematic effects on the members within the groups on the one hand and on the groups on the other hand. The underlying statistical model of MLM is described using an example from CSCL. Attention is drawn to the fact that MLM requires large sample sizes which are not provided in most CSCL research. A proposal is made for the use of some analyses which are useful.