Studies in tropical streams are relatively few, and one of the still-unresolved methodological issues is sample size. Adequate sample size for temperate streams cannot be extrapolated for tropical sites, because of the differences in species richness and the proportions of rare species. We evaluated reliable sample size for estimation of resemblance among samples of macroinvertebrate assemblages inhabiting riffles of tropical streams, using the autosimilarity approach. Sample sizes were much larger than those currently employed in tropical studies. Sampling units consisted of individuals associated with single stones (15–20 cm). Evaluations employed the Bray-Curtis index for abundance data and the equivalent Sorensen index for presence-absence data. Autosimilarity curves were constructed using both sampling units and individuals. The estimation of resemblance among samples was strongly dependent upon sample size at reduced sampling effort, particularly for the Bray-Curtis index. For the same sampling effort, fixed counts of individuals obtained randomly from sampling units gave better estimations of resemblance, and their similarity curves tended to stabilize earlier than those using sampling units. A minimum of 9–15 sampling units (stones) or 150–850 individuals is necessary for adequate estimations of resemblance using presence-absence data, and 13–18 sampling units or 750–1550 individuals are required for relative abundance data in tropical streams.