Hi! I am building a multivariate PERMANOVA model: sequencing plate+habitat+sex. I am primarily interested in the variation that sex has on microbiome composition and wanted to control for sequencing plate and habitat. I wanted to build this multivariate model because I am fairly sure that plate and sex are confounded.
However, after using permdisp (beta-group-significance plugin), the first two variables (plate and habitat) do not exhibit homogeneity of dispersion but sex, does. Plate and habitat do not have balanced designs but sex does.
This leads me to my two questions: 1) Can I not draw conclusions regarding mean differences using my PERMANOVA multivariate model since I broke its assumptions regarding homogeneity of dispersion, even if it is my covariates and not my variable of interest and therefore 2) would I only be able to utilize a univariate model for sex?
I think questions about confounding are best discussed with your advisor or a local statistican. There's a limit to waht we can do interms of your models/systems. Sometimes it can be helpful to check your assumptions about the distributions and assocations of non-microbiome covariates.
With regard to dispersion, what do your boxplots and PCoA tell you about the dispersion? I find visualizing the data is often a reliable and helpful way to evaluate differences in dispersion vs differences in centroids. Is there a potential your beta diversity is confounded with alpha diversity? In unweighted metrics, differences in alpha diversity between groups can sometimes lead to greater dispersion in the lower diversity group, even if the centroids are differnet. You'll often see this in PC 1 of a jaccard or unweighted UniFrac plot.