BACKGROUND Diagnostic changes may reflect evolution of an illness, emergence of newly disclosed information, or unreliability of assessment. This study evaluates the stability of research diagnoses in a heterogeneous first-admission sample with psychosis. METHODS A group of 547 subjects initially diagnosed with a psychosis were reassessed 6 and 24 months after enrollment. The DSM-IV consensus diagnoses were formulated by psychiatrists blind to previous research diagnoses. The analysis focuses on agreement over time and the effects of demographic, family history, and clinical variables on the shift from a nonschizophrenia diagnosis to schizophrenia. RESULTS Seventy-two percent of 6- and 24-month diagnoses were congruent. The most temporally consistent 6-month categories were schizophrenia (92%), bipolar disorder (83%), and major depression (74%); the least stable were psychosis not otherwise specified (44%), schizoaffective disorder (36%), and brief psychosis (27%). The most frequent shift in diagnosis at 24 months was to schizophrenia spectrum (n=45). These 45 subjects had a similar illness course after 6 months as the 171 subjects in this category at both assessments, but their prior clinical functioning was better. Risk factors predicting change to a schizophrenia spectrum diagnosis include facility variables (schizophrenia diagnosis, longer stays, and given antipsychotic medication on hospital discharge); prehospital features (psychotic > or =3 months before admission, poorer adolescent adjustment, lifetime substance disorder); and negative symptoms. CONCLUSIONS Changes in diagnosis, particularly to schizophrenia, are mostly attributable to the evolution of the illness. Rigid adherence to DSM-IV requirements may have led to underdiagnosis of schizophrenia. The findings support the need for a longitudinally based diagnostic process in incidence samples.