The personal human microbiome.
a. Do samples differ in composition by subject-id (i.e., across individual)?
My solution is to first run filter-samples to remove the fmt material (because single samples cause error on beta diversity calculation), then run beta-group-significance analysis on the unweighted_unifrac_distance_matrix.qza . My result is attached here: unweighted-unifrac-subjectID-beta-significance.qzv (545.8 KB). Scrolling down to the pairwise PERMANOVA results, all subjects are significantly different from each other. I also notice that if the subjects were to compare with themselves, they could possibly be significant too, because looking at the box plots, the distance of each subject to itself seems highly variable (e.g. B101's whiskers range from 0.2 to 0.7).
Is this an acceptable answer to this question? I would like to see an example answer to check against my own.
Hi @jjmmii, I think you’re on the right track! From looking at your plots/stats, it looks like the “global” PERMANOVA test (i.e. the results at the top of the page, applied to all individuals) says there are at least two individuals who have significantly different community compositions. The pairwise results confirm this and suggest that all individuals have significantly different compositions from the others.
However, this interpretation of the results isn’t necessarily correct, which I’ll explain below in response to your other question.
Great question! What you’re seeing here is that individuals have differences in the variability of their community composition across different sampling events. In other words, an individual’s community composition may vary more (or less) than another individual’s. This paper describes this “personalized variability” of the human microbiome in more detail.
The PERMANOVA test is not all that robust to differentiating between differences in community composition (i.e. differences in centroids) vs differences in community variability (i.e. differences in dispersion). From the PERMANOVA test alone, we can’t actually say whether these significant results indicate difference in composition or difference in variability. There is another test, PERMDISP, that specifically tests for differences in variability, so PERMDISP and PERMANOVA are typically used together to determine whether the effect is centroid or dispersion differences.
Unfortunately, qiime2 does not yet support PERMDISP. @yoshiki is actively working with a student in his lab to implement PERMDISP in qiime2, and it should be available by the end of 2017 at the latest. We’ll follow up here when it’s available! In the meantime you can use qiime1’s compare_categories.py script to execute PERMDISP, or you can do it in R with the vegan package.