Multimodal learning analytics (MMLA) research has made significant progress in modelling collaboration quality for the purpose of understanding collaboration behaviour and building automated collaboration estimation models. Deploying these automated models in authentic classroom scenarios, however, remains a challenge. This paper presents findings from an evaluation of collaboration quality estimation models. We collected audio, video and log data from two different Estonian schools. These data were used in different combinations to build collaboration estimation models and then assessed across different subjects, different types of activities (collaborative-writing, group-discussion) and different schools. Our results suggest that the automated collaboration model can generalize to the context of different schools but with a 25% degradation in balanced accuracy (from 82% to 57%). Moreover, the results also indicate that multimodality brings more performance improvement in the case of group-discussion-based activities than collaborative-writing-based activities. Further, our results suggest that the video data could be an alternative for understanding collaboration in authentic settings where higher-quality audio data cannot be collected due to contextual factors. The findings have implications for building automated collaboration estimation systems to assist teachers with monitoring their collaborative classrooms.