Abstract:
|
Microbiome data frequently feature the analyses of matched sets, such as when comparing the microbiomes of the same individual before and after treatment. Ignoring the clustering in matched data can lead to invalid inference. Moreover, power gains are also possible when the analysis accounts for matching, as a high-dimensional nuisance parameter (i.e., the individual- or set-specific microbiome for each cluster) could be eliminated. We show how this can be accomplished using the Linear Decomposition Model (LDM), a linear model based on simultaneous projections on observations (as in standard linear regression) and features (as in Principal Components Analysis). Using both simulated data and real matched microbiome data, we show that an analysis that removes the between-cluster heterogeneity can increase the power to detect both global effects and individual OTUs.
|