Even a well-designed and carefully collected reference must be employed with caution in order to minimize spurious variation and contain necessary computational effort. The first of these needs arises since reference-based analyses assume that any systematic compositional differences inherent in the data outweigh any technical variation, which is particularly problematic when combining data generated from different protocols or platforms [31]. In fact, biological conclusions can be driven by technical variation even if the researchers are careful (as in [40], where samples were found to cluster by the extraction kit used), which underscores the need for accepted community standards for sample handling, sequencing, and data analysis in order to minimize the potential for introducing such variation. Bioinformatic strategies to mitigate any remaining variation, such as trimming sequences to a common length between studies, have shown to help normalize platform bias [29]. Sometimes, stronger measures are necessary: for example, the American Gut Project received samples from self-reported healthy individuals that contained levels of gammaproteobacteria beyond anything previously observed in healthy populations (although similar to those observed in samples from ICU patients [manuscript
