The objectives of the study were (1) to explore differences in the course for patients treated in a day unit specializing in personality disorders (PDs), and (2) to determine characteristics of patients with different courses and predictors of various courses. K-mean cluster analysis was applied to partition a sample of 128 patients, 101 with various PDs and 27 with axis I disorders only, into four groups representing different courses. The course was defined on the basis of global functioning (Health Sickness Rating Scale [HSRS]) at admission, discharge, and 3-years follow-up evaluation. The four courses were labeled good, fair, late improvement, and poor, demonstrating great variation in the short-term course among patients with PDs. Predictors were studied by means of polychotomous logistic regression using the patients with a fair course as the reference group. The dichotomy no PD/cluster C versus cluster A/B predicted a good versus a fair course. A poor work status the last year before admission and irregular discharge predicted a poor or late improvement course versus a fair course, also when controlling for PD clusters. None of the included variables discriminated between patients with a poor versus late improvement course.