I have 16S data to look at the microbial communities of two types of hot springs; I've been working with this in R. Some of the data is from 2017-2019, but the bulk of the data is from last year. There's 18 samples from 11 individual hot springs of type 1, and 31 samples from 6 individual hot springs of type 2. 24 of the 31 samples of type 2 are triplicates of 8 sampling sites. No other sample has a replicate. So, this data is very skewed. I tried doing a PCoA to see whether the two types of hot springs distinctly cluster from each other, but, although they seemed to, the axes numbers were only about 5% and 8%. I tried a CAP plot and it was similar. Is there a better way to visualize clusters? Should I be transforming the data (log2 transform or relative abundance transformation) before doing ordination plots? There are definitely differences in the microbial communities of the two types, I can see it in the bar plots of what organisms are present. The data was transformed after I did ordination and before I did bar plots.
My PI says that I need to account for the n=3 samples for the 8 sites that were sampled in triplicate, as they are skewing the data. I suggested either just using the first sample from each site, random subsampling, and merging each triplicate into one samples and all suggestions were veto-ed. What other method can I use to account for this?
I suppose all what you did in analysis and as suggestions are the way to go and generally used in analysis. I can't understand what your PI want you to do in this case. And if all of your suggestion are rejected, he/she must come with any suggestion.
Further, it does not make much logic of difference to use replicates for some samples and not for other samples. Statistics is sensitive to input data and if user is using it without thinking about the input data, one should also not be so sure about the output. Blunt force and crap input data might give desired output but would it be logical?
Dear Teresa,
What do you mean by saying that the data are 'skewed'? In statistics skewness is one of distribution properties, but I can't relate this to your case.
But to the point, as you have no replicates of type 2 samples, you can compare only type 1 with type 2. I think that the most appropriate approach would be averaging type 1 samples and comparing them with type 2. As to the ordination technique, if PCA/PCoA doesn't work well, you could try NMDS (e.g. metaMDS function of the 'vegan' R package). Significance of grouping should be checked with PERMANOVA (adonis function in vegan) and homogeneity of variance with betadisper (multidimentional analogue of the Levene's test).
Best regards,
MG
There are a several issues here. The first is unbalanced sampling. If each of the triplicate samples is the same size as the individual samples then just taking one of them for the combined analysis is a sensible thing to do. If you pool them then the combined triplicates will have more taxa in them and appear different, which is not what you want. If the combined triplicate sample is the same size as the individual samples then you can pool them.
The second issue is what the nature of the data is. What are your counts - read number? Presence/absence? Whatever the answer you need to select an appropriate resemblance measure. I would normally choose Bray-Curtis and depending on the range of numbers in each sample I would generally recommend a transformation. If you don't transform the pattern you get just reflects variation in the most abundant taxa. A strong (log, fourth root) transform brings more taxa into play. I would use nMDS as my ordination method, but that is my choice.
The third is how to answer your question, which is about differences (not clustering, per se) between spring types. For this you should use some appropriate testing framework working on the resemblances, such as ANOSIM or PERMANOVA.
- Badges
- Science method