An earlier post by @CarlyRae first introduced me to the fact that when examining beta diversity I need to consider not only the results of a multi or univariate anova, but also examining if the within-group distances to a centroid were equivalent among my groups. I'm wondering how to best interpret the situation where everything turns out to be significant. That is, if the results of both PERMANOVA and PERMDISP suggest significant differences, where do most people go next? I've run these tests both in QIIME and separately with vegan in an R environment; while QIIME's visualizations show barplots, Carly's suggested ordinating these data.
In that spirit, I was futzing around in R this morning and came up with this:
The points represent samples, the colors represents the groups of samples they belong to. This led me to a few questions:
Pardon my complete ignorance, but could I have generated this ordination in QIIME anyway? I wasn't sure if the input in running the vegan
betadisp()function generates an output that is similar to what is created with
qiime diversity pcoa+
qiime emperor. My confusion here is that the input for calculating dispersions in Vegan is a distance matrix, just like the input for the PCoA plot in QIIME. However, the input for the plot I made was the output of the
betadisp()function, whereas the input for QIIME's
PERMDISPis a distance matrix. Perhaps I need to deconstruct the resulting
.qzvartifact following PERMDISP and feed that into the
Regarding the above plot, I was worried that the proportion of variance explained by the two eigenvectors was quite low, and wondered how researchers use this information in their interpretations of the figure itself. If the values were something along the lines of 55% and 25% respectively, how would that change your interpretation compared to my existing values of 6% and 5%?
Given the above plot, it strikes me that my samples are more associated by the month they were collected than the site they were collected, though both appear to matter. This was indicated as much in the Adonis test. What would be extremely helpful to me is gaining some insights about how researchers pool these collective bits of information into a final assessment. Given that I have heterogeneous dispersions across groups and am violating an assumption of the Adonis test, how do the resulting barplots (from QIIME's viz output) or this ordination assist in making an interpretation of how grouping factors explain diversity? Is it sufficient to inform readers that these data provide evidence for differences between groups, but that those differences can be explained by both inter-group variation in distances to centroids, and by differences between groups? I'm guessing there are always limitations to what one can infer - I'd greatly appreciate others experiences with what they were confident in affirming in their work when both PERMANOVA and PERMDIST were significant.