Here’s an issue I haven’t seen discussed. I have a number of samples from patients that are paired. Each patient gave me two sputum samples separated by time. I’d like to analyze the microbiome and compare the first to the second set. Since the samples are paired, the statistics will be different. I am concerned that I need to structure my design files, etc., differently to account for the pairing.
I can pull out raw data – for example, alpha diversity measures – and do that analysis separately in STATA, and account for the pairing. But I am worried that for some of the other analyses done within Mothur, the assumption is that the samples are independent, and these are not.
What analyses are you talking about being done in mothur? I take the beta div. matricies into R and to my analyses there but paired samples are a difficult nut to crack. You may want to focus on just the shared OTUs (or just the unshared but that dangerous because 0 means not detected, not absent)
As one example: distance matrices and the calculation of PCOA or NMDS. Let’s say I have 7 subjects, each with sample A and B. My major comparison is 1A to 1B, 2A to 2B … 7A to 7B. I’m not sure I care about the distances between 1A and 4B, for example.
This may be beyond my ability to express (and I’m not a computational person); but I’m concerned that just running the PCOA command in Mothur, with all samples just represented without consideration of pairing, is going to lead me astray.
First don’t run PCoA unless you know that your underlying gradient is linear (for natural communities it basically never is), use NMS. But that isn’t a statistical test. You can have what ever you want in an NMS and you still aren’t testing anything-it’s just a nice visualization technique to see if there are patterns between your samples. Permanova or manova are tests of differances. I don’t know manova well, but in permanova (in R) you can use strata to block by patient for repeated measures.