Pearson correlation

Hi All,

I would like to make sure that my interpretation in the pearson correlation is correct or not.
I used corr.axes() with NMDS.axes and metadata (Temperature and pH) to calculate in pearson correlation, following MiSeq_SOP analysis.

The result showed in below :

Feature axis1 p-value axis2 p-value axis3 p-value length
Temp 0.446652 0.062743 0.706606 0.000650 -0.172789 0.499046 0.853608
pH -0.438805 0.068272 -0.292427 0.243367 0.566083 0.012939 0.773636

I understand that Temperature move in a positive direction with p-value < 0.05 (0.000650) and also pH move in a positive direction with p-value < 0.05 (0.012939). However, if I visualize it in biplot and consider in p-value, pH does not correlate significantly (because axis1 and axis2 showed p-value > 0.05). Is it correct?

Thanks in advance.

I think your interpretation is pretty good. However… it’s muddled a bit because you’re using NMDS where the axes don’t really mean anything. In fact, if you ran nmds again, you’d get a similar scatter, but different rotation about the axes and thus different correlations. For the type of interpretation you’re after I think it would be better to do the biplot with the output from pcoa. There the first axis represents the most variation in the data, the second the second most, third, etc. PCoA output is stable between executions and always gives the same output.


Thank you very much for your suggestion. I will try to do pCoA for this analysis. :slight_smile:

Pat, I have to disagree. PCoA is only appropriate for data that are responding linearily to an underlying gradient. Community data basically never is. The artifact of this violation of the assumptions of PCoA is the points forming a horseshoe (that shows up in nearly every microbiome paper using PCoA. AARRRRGGGGGG). You can certainly do a joint plot on an NMS-the axes don’t mean anything but the relationship between the samples and the OTUs should be fairly constant. Here’s my code (bottom of this script) for making a joint plot.

Agreed - and if you’re going to use NMDS then you should probably ignore the p-values and focus on the overall length of the vector.