When analyzing percentage or proportional datasets where the components sum to a fixed constant (such as 1 or 100%), standard linear statistics fail due to inherent geometric constraints and negative bias. This challenge is further complicated when sample sizes between comparative groups are unequal (unbalanced designs), which breaks traditional parametric MANOVA assumptions under variance heterogeneity.
The optimal modern workflow solves this by shifting the analysis from raw proportions to log-ratio coordinates. First, absolute zero values are resolved using a Bayesian multiplicative replacement method. Next, an Isometric Log-Ratio (ilr) transformation is applied to map the data into an unconstrained coordinate space, successfully preserving standard Euclidean distances without mathematical singularity.
Once transformed, a global Permutational MANOVA (PERMANOVA) using Type III Sum of Squares is deployed to test overall group differences without distribution restrictions. For post-hoc identification of specific components driving the variation, frameworks like ANCOM-BC or ALDEx2 provide robust, bias-corrected linear regression adjustments that natively handle unbalanced group dynamics.