I want to receive advise on interpreting PCoA plot and statistical results. Thank you.

[yhuang816]

Hi QIIMErs, 

I have long time not posting question here and hope you all do well. 

I have difficulties of interpreting PCoA plot and statistical results. In the PCoA plot (please see attached file), PC1 shows 89.70% of variation between non-chlorinated and chlorinated samples. In both non-chlorinated and chlorinated samples, PC1 also shows 89.70% of variation between high and middle chloride levels and low chloride level. If I interpret the plot in a wrong way, please correct me. 

I thought 89.70% of variation in PC1 would give statistical significant result. However, based on p-test and anosim test, there are no significant results (p>0.05). I am confused here. Why 89.70% of variation in PCoA plot and statistical results do not match with each other? 

I also filter out non-chlorianated samples in the otu table and focus on chlorinated samples. I created a PCoA plot by using weighted unfirac distance metric. In that plot, the three samples are separated from each other on PC1 (63.36% variation). However, the adonis test tells me no significant result. 

Please see the attached slides showing each analysis step and output. 

I am looking forward to receiving your advise. 

Thank you 

Yannan 

[yhuang816]

Here is the attached file. Sorry, I forget to attach it to the previous email. 

[Colin Brislawn]

Hello Yannan,

Thanks for attaching your presentation. That really helps.

In your ordination, I notice you have six points. Do you only have six samples? That's a very small number of samples, and that could dramatically limit your statical power. It also limits the meaningfulness of the Principle Coordinates in your ordination; with so few samples, it's no wonder that PC1 explains 80% of the variation. 

If you have a small number of samples, perhaps ordination is not a good path forward. Have you considered performing a differential abundance test and focusing your narrative on the OTUs which are different between samples? This may be a better way forward.

http://qiime.org/scripts/differential_abundance.html

Good luck!

Colin

[yhuang816]

Hi Colin,

Thank you for your quick reply.

Yes, I only have six samples. In the presentation, I also attach a mapping file. The mapping file includes sample information.

I do not understand why PCoA is not good for small number of samples. Why "with few samples, it is no wonder that PC1 explains 89.70% of the variation"? Could you please explain more detail?

I will take a look at the differential abundance test and then give it a try. I will let you know how it goes. 

Thank you

Yannan

[Colin Brislawn]

Hello Yannan,

PCoA is a really elegant math technique which attempts to explain variation between samples. Here is the best explanation I can find. 

http://occamstypewriter.org/boboh/2012/01/17/pca_and_pcoa_explained/

Essentially, with fewer samples, you have fewer sources of variation for PCoA to explain. I think differential abundance is your best bet!

Colin

[yhuang816]

Hi Colin, 

I think I understand why the PCoA ordination technique is not good to explain 6 samples' variation. The disadvantage of the 6 samples that there is no duplicate sample for each of them. Therefore, no standard deviation, no statistical power. 

When I read differential abundance description, I realize two points: 1) We would recommend having at least 5 samples in each category. 2) With these techniques, we would still recommend removing low depth samples (e.g. below 1000 sequences per sample)".  In my case, the category (chlorinated treatment), I have 3 samples. Also, three of six samples have numbers of sequences that are below 1000 sequences (Please see attached otu smmary file if you are interested). My samples do not match those two points. 

Prior to spend time on computing the differential abundance test and interpreting the result, I want to get your advice whether it is worthwhile to try it? 

Thank you 

Yannan 

[Colin Brislawn]

Hello Yannan,

I think it's worth trying, just to make sure the script works and to see what the results look like.

You have a small number of samples and few reads in those samples. If at all possible, get more of both! 

Colin