PERMANOVA tests whether groups differ in multivariate space without assuming multivariate normality. Magic Stat runs it from a dissimilarity matrix — Euclidean in the mean-comparison dialog, any of the common ecological measures in the ecology module — and reports pseudo-F, R² and a permutation p.
PERMANOVA — permutational multivariate analysis of variance, also called Adonis — partitions the total sum of squared distances between samples into a between-group and a within-group part, and tests the ratio with permutations. The test statistic is a pseudo-F, and the effect size is R²: the share of multivariate variation the grouping explains.
It is used so widely because it makes no normality assumption and works from any dissimilarity measure. It is misused just as widely because a significant result can mean a difference in group location, a difference in group dispersion, or both — and PERMANOVA does not distinguish them. If dispersion could differ between groups, a test of homogeneity of dispersion (such as PERMDISP) is needed alongside, and the ecology module runs that test too.
The choice of dissimilarity is not cosmetic: Bray–Curtis for abundance, Jaccard or Sorensen for presence/absence, Euclidean for standardised continuous variables. Changing the measure changes the question.
PERMANOVA is a test, not a picture. It is normally reported next to an ordination — PCoA or NMDS — computed from the same dissimilarity measure, so that the figure and the test describe the same thing.
Dissimilarity measures (ecology module): Bray–Curtis · Euclidean · Manhattan · Canberra · Jaccard · Sorensen · Horn–Morisita. The mean-comparison dialog runs PERMANOVA on Euclidean distances.
Permutations: a spinner in the ecology module (99 to 9999, default 999); the mean-comparison dialog uses its standard number. Missing values are dropped from the selected columns before the distance matrix is built.
MANOVA assumes multivariate normal responses and uses parametric test statistics. PERMANOVA makes no such assumption and works from distances. If your variables are roughly normal and on comparable scales, MANOVA has more power; otherwise PERMANOVA is the safer test. The same dialog reports both.
Yes, whenever a significant result could plausibly come from groups differing in spread rather than in position. Run a homogeneity-of-dispersion test — the ecology module has one — and report it next to the PERMANOVA. A significant PERMANOVA with unequal dispersion is a result that needs care in the wording.
Match it to the data and state it: Bray–Curtis for abundance, Jaccard or Sorensen for presence/absence, Euclidean for standardised continuous variables. The ordination you report should use the same measure.
More is better within reason; 999 is the common convention and the default here. With very few samples the minimum achievable p is bounded by the number of permutations, which is worth stating in the methods.
No — the analysis runs locally. The only automatic signal is an anonymous installation counter with no data in it.