Multivariate Multinomial Logit Models for Dyadic Sequential Interaction Data

The analysis of discrete dyadic sequential behavior and, in particular, the problem of forecasting future behavior from current and past behavior in such data is the main theme of the present article. We propose to use multivariate multinomial logit models and the potential of which will be demonstrated with data on Imagery play therapy. In such a therapy, the therapist tries to draw a child into Imagery play, so that it can act out its emotions and feelings which gives the therapist the opportunity to communicate with the child. As the therapist wants to interact clinically with the child, it is important to draw it into Imagery play as soon as possible. Our aim is to find out how the therapist achieves this by examining and forecasting new behavior from past behavior. New behavior can be modeled by special weighting schemes, but this procedure can be technically problematic. In this paper the nature of these problems will be explained and methods to solve them will be proposed. Moreover, we will explicitly attempt to answer the substantive questions for which the data were collected through a detailed interpretation of the parameters of the models.