Statistical user simulation is a promising methodology to train and evaluate the performance of (spoken) dialog systems. We work with a modular architecture for data-driven simulation where the “intentional” component of user simulation includes a User Model representing userspecific features. We train a dialog simulator that combines traits of human behavior such as cooperativeness and context with domain-related aspects via the Expectation-Maximization algorithm. We show that cooperativeness provides a finer representation of the dialog context which directly affects task completion rate.