||At the heart of traditional Chinese medicine (TCM) is the concept of syndrome type, which refers to the pathological condition of a patient at a given stage of a disease. Unlike diagnosis in Western medicine, diagnosis in TCM means, in most cases, to determine a patient’s syndrome type instead of his disease type. A patient’s syndrome type is not directly observed. Rather, it is indirectly assessed on the basis of symptoms. A major weakness of TCM is that syndrome types are not rigorously defined, leading to subjectiveness and arbitrariness in diagnosis. To rectify the situation, researchers in China have been seeking for gold standards for TCM diagnosis for forty years. Unfortunately, no such standards have been found. Latent class analysis (LCA) is a popular method for model-based clustering. It has been used in western medicine, psychology, and social science to provide statistical justification and qualification for latent concepts. In this thesis, we first apply LCA to two TCM data sets. We find that the latent class (LC) models resulting from LCA fit data poorly. We hence study a generalization of LC models called 3-level latent structure (3-LS) models. With the generalization, we are able to find models that fit the data well. Those models might provide TCM researchers insights about TCM theories and be used to guide diagnosis in practice1. 1These are for future work.