Conversational agents are more and more investigated by the community but their ability to keep the user committed in the interaction is limited. Predicting the behavior of children in a human-machine interaction setting is a key issue for the success of narrative conversational agents. In this paper, we investigate solutions to evaluate the child's commitment in the story and to detect when the child is likely to react during the story. We show that the conversational agent cannot solely count on questions and requests for attention to stimulate the child. We assess how (1) psychological features allow to improve the prediction of children interjections and how (2) exploiting these features with Pattern Mining techniques offers better results. Experiments show that psychological features improves the predictions and furthermore help to produce robust dialog models.