T-RFLP data sets are characteristically both high-dimensional and noisy (high, uncontrolled variation). Although analysis of T-RFLP data has developed considerably over the last decade, there remains a lack of consensus about which statistical analyses offer the best means for finding trends in these data. Our objective was to determine the most robust and informative multivariate statistical tools for analyzing T-RFLP data. We empirically tested and theoretically compared three T-RFLP data sets using the most common methods in the literature: principal component analysis (PCA), non-metric multidimensional scaling (MDS), correspondence analysis (CA/RA), detrended correspondence analysis (DCA) and a technique new to T-RFLP data analysis, the additive main effects multiplicative interaction (AMMI) model. Also known as a doubly-centered PCA, AMMI uses analysis of variance (AOV) to partition the variation into treatment and terminal-restriction fragment (T-RF) main effects and the treatment by T-RF interactions, and then applies PCA to the interactions. Therefore, instead of examining overall variability of the T-RFs in relation to treatments (i.e. how common T-RFs occur over all the treatments), AMMI examines the differential responses of T-RFs to the treatments, a more appropriate approach for microbial community analysis. Our findings indicate that the main effects (rareness or commonness of T-RFs) commonly account for the majority of variation in T-RFLP data sets. This becomes the dominant signal of multivariate ordination techniques that do not first partition out the main effects, and consequently obscures the pattern of interest (treatment x T-RF interactions). This fundamental difference makes AMMI a valuable tool in detecting significant patterns in T-RFLP data sets.