Environmental effects on species interactions can be studied by comparative analyses of network structure. For example, comparison of interaction networks among study sites can provide clues to geographic variation of host breadth. Obligate plant–ant interactions are ideal systems to explore these phenomena because they are long term and can be accurately sampled in the field. We tested two hypotheses: (1) network structure and host specialization do not vary among communities, and (2) the effects of plant extinction do not vary among communities. We sampled 10 or more plants for each of the 30 ant–plant species found in three Neotropical locations. We found that network specialization, H2′, was significantly higher than expected in random networks. The ant or plant specialization index, d′, distribution did not vary among localities, neither varied in link or asymmetry distribution. Plant extinction simulations showed that these interactions are vulnerable to plant loss, and the null model was more robust than the observed networks. This study provides a foundation on which plant and ant phylogenies can be added to explore compartment evolution.