The relationship between diagnoses generated by the Structured Clinical Interview for DSM-III-R Personality Disorders (SCID) and by nonstructured psychiatric interviews was examined. The purposes were to evaluate which DSM-III-R diagnoses were most reliably chosen, and to compare diagnostic practices between two clinical sites. Diagnoses generated by researchers using the patient version of the SCID and by psychiatric interviews were compared for 100 patients. The participants had been randomly assigned to one of two acute treatment sites within the same institution, as part of a larger study of an alternative to inpatient hospitalization. Overall reliability between the SCID and the clinicians, as determined by weighted Kappa, was poor. There was considerable variability among the major diagnostic categories, with higher agreement for schizophrenia and bipolar disorder than for others. The agreement for schizoaffective disorder was extremely low. There were also significant differences in the patterns of diagnoses between the two sites. The patient version of the SCID appears to produce results that are very different from clinical practice, which, in turn, may be influenced strongly by location.