Multiple correlated outcomes in a randomized experiment with two groups and two time points: How to analyze? In a randomized trial, subjects are randomized into two groups, a control group and an intervention group. There are multiple outcomes (15 in total), namely miRNA (microRNA) which are measured before (baseline) and after the intervention period (follow-up). These miRNA are highly correlated. We don't have specific hypotheses concerning the change of individual miRNAs, i.e. we don't have hypotheses of the sort "miRNA1 is expected to change in the intervention group but not miRNA2".
Usually, I would analyze such a design using an ANCOVA, with the measurement at follow-up as dependent variable and the baseline measurements and a group indicator as predictors (see Vickers et al. (2001)). We could do separate ANCOVAs for each miRNA, but I suspect it would be beneficial to take the correlation among the miRNAs into account. Also, with so many tests we'd need to make some multiplicity adjustments.
Specifically, I can think of three possible analyses that seem reasonable to me:

*

*Indiviual ANCOVAs for each miRNA.

*Summarize the miRNA using a PCA (principal component analysis) and use the components (e.g. the first one) in the ANCOVA as described above.

*A MANVOA on the change scores of the miRNA.

Which analysis would you recommend?
 A: I'll go with number 1. 15 outcomes is not much, so results of your analysis won't be very broad. Probably you can show them in a single clever table.
Plus, I'd show correlation matrix (or heatmap) of outcomes to help explaining why ANCOVA results are similar for several miRNAs.
Number 2 has, in my opinion, a large disadvantage that it does not deal with actual miRNAs. I think it would be quite frustrating for a reader to see a conclusion that some strange combination(s) of miRNAs change in intervention group, while he/she probably expects to see if specific miRNAs changes in intervention group.
But that's only my (probably narrow) point of view.
One more method that comes to my mind is to cluster miRNAs with some method that allows choosing "representatives" of clusters (partitioning around medoids, maybe?), run ANCOVAs only on those "representatives", and say that results for the rest of miRNAs would be similar to those obtained for "representatives".
With 15 miRNAs you may explicitly state which "representative" represents which miRNAs clustered together.
