BACKGROUND: Latent class analyses (LCA) are increasingly being used to target specialized HIV interventions, but generalizability of emergent population structures across settings has yet to be considered. We compare LCA performed on two online samples of HIV negative Chinese men who have sex with men (MSM) to detect more generalizable latent class structures and to assess the extent to which sampling considerations impact the validity of LCA results. METHODS: LCAs were performed on an 1) nationwide online survey which involved no in-person contact with study staff and a 2) sentinel surveillance survey in which participants underwent HIV and syphilis testing in the city of Guangzhou, both conducted in 2014. Models for each sample were informed by risk factors for HIV acquisition in MSM that were common to both datasets. RESULTS: An LCA of the Guangzhou sentinel surveillance data indicated the presence of two relatively similar classes, differing only by the greater tendency of one to report group sex. In contrast an LCA of the nationwide survey identified three classes, two of which shared many of the same features as those identified in the Guangzhou survey, including the fact that they were mainly distinguished by group sex behaviors. The final latent class in the nationwide survey was composed of members with notably few risk behaviors. CONCLUSIONS: Comparisons of the latent class structures of each sample lead us to conclude that the nationwide online sample captured a larger, possibly more representative group of Chinese MSM comprised of a larger, higher risk group and a small yet distinct lower group with few reported behaviors. The absence of a lower risk group in the Guangzhou sentinel surveillance dataset suggests that MSM recruited into studies involving free HIV/STI testing may oversample MSM with higher risk behaviors and therefore greater risk perception. Lastly, two types of higher risk MSM were emergent across both samples distinguished largely by their recent group sex behaviors. Higher odds not only of self-reported HIV infection but also of closeted tendencies and gender fluid identities in this highest risk group suggest that interacting factors drive individual and structural facets of HIV risk.