Objective: The purpose of this investigation was to derive an empirical classification of eating disorder symptoms in a heterogeneous eating disorder sample using latent class analysis (LCA) and to examine the longitudinal stability of these latent classes (LCs) and the stability of DSM-IV eating disorder (ED) diagnoses. Method: A total of 429 females with ED symptoms were assessed using the Eating Disorder Examination every 6 months for 2 years. LCA was used to derive empirical classification at baseline. Latent transition analysis (LTA) was used to examine the longitudinal stability of LCs, and Markov modeling procedures were used to examine DSM-IV ED diagnoses over all the time points. Results: LCA yielded a 3-class solution: binge eating and purging, binge eating only, and low body mass index. LTA indicated that these LCs showed greater stability over 2 years than DSM-IV diagnoses with the probability of remaining in the same class ranging from 0.69 to 0.91 for LCs and from 0.40 to 0.75 for DSM-IV diagnoses. Transition patterns also revealed more stability for LCs with only 21% changing classes compared with 63% of the DSM-IV diagnostic categories. Conclusion: Empirically derived classes of ED symptoms showed greater longitudinal stability than DSM-IV diagnoses over a 2-year time period, suggesting that modifying the criteria to be consistent with empirically based classification (e.g., reducing frequency requirements of binge eating and purging) may reduce ED diagnostic crossover in DSM-5 . (Contains 1 note, 4 tables, and 2 figures.)