Temporal variation in microbiome measurements can reduce statistical power in research studies. Quantification of this variation is essential for designing studies of chronic disease. We analyzed 16S ribosomal RNA profiles in paired biological specimens separated by 6 months from 3 studies conducted during 1985-2013 (a National Cancer Institute colorectal cancer study, a Costa Rica study, and the Human Microbiome Project). We evaluated temporal stability by calculating intraclass correlation coefficients (ICCs). Sample sizes needed in order to detect microbiome differences between equal numbers of cases and controls for a nested case-control design were calculated on the basis of estimated ICCs. Across body sites, 12 phylum-level ICCs were greater than 0.5. Similarly, 11 alpha-diversity ICCs were greater than 0.5. Fecal beta-diversity estimates had ICCs over 0.5. For a single collection with most microbiome metrics, detecting an odds ratio of 2.0 would require 300-500 cases when matching 1 case to 1 control at P = 0.05. Use of 2 or 3 sequential specimens reduces the number of required subjects by 40%-50% for low-ICC metrics. Relative abundances of major phyla and alpha-diversity metrics have low temporal stability. Thus, detecting associations of moderate effect size with these metrics will require large sample sizes. Because beta diversity for feces is reasonably stable over time, smaller sample sizes can detect associations with community composition. Sequential prediagnostic specimens from thousands of prospectively ascertained cases are required to detect modest disease associations with particular microbiome metrics.