I am working on diversity analysis, NMDS has been used to discriminate ecosystems , its getting a stress value 0.02,0.03 etc how can is explain the stress value in the interpretation, i may wonder if any provides reliable literature also.
Many thanks in advance
Hi Sileesh,
A rule of thumb: stress > 0.05 provides an excellent representation in reduced dimensions, > 0.1 is great, >0.2 is good/ok, and stress > 0.3 provides a poor representation.
https://jonlefcheck.net/2012/10/24/nmds-tutorial-in-r/
Clarke (1993) suggests that > 0.20 is basically random; < 0.15 is good; < 0.10 is ideal i.e., configuration is close to actual dissimilarities
Maarten's answer is good, except the 'less than' sign is backwards. Basically, the opposite of what he said. Instead, it should read:
A rule of thumb: stress
Hej! I found something more about this question and I just add them here:
1. https://rdrr.io/cran/vegan/man/monoMDS.html "Minimum stress. If stress is nearly zero, the fit is almost perfect. Usually, this means that the data set is too small for the requested analysis, and there may be several different solutions that are almost as perfect. You should reduce the number of dimensions (k), get more data (more observations), or use some other method, such as metric scaling (cmdscale, wcmdscale)." 2. https://github.com/MadsAlbertsen/ampvis2/issues/78 "Euclidean distance is very different from Bray-Curtis or others, and it's generally not suited for community data without transformation." 3. https://r-help.stat.math.ethz.narkive.com/cogi8MVR/r-problem-with-metamds-in-vegan "The Bray coefficient is designed for analyzing community structure by comparing the counts of different species between samples. " So, it seems if some sample in your whole datasets differed too much from other samples, the NMDS1 would be oversaturated, meaning the stress is too low, i.e., the fit is almost perfect. This is good for NMDS but not what you expected in microbial community analysis. So the solution is 1) add more samples (usually not an option in real cases), or 2) use other cluster methods. In my experiences, the NMDS works well with a denoised and transformed dataset (i.e., small reads were filtered, and reads counts were transformed as relative abundance).
NMDS generally helps us to represent the species communities found across several localities in a reduced number of dimensions. If two localities are found nearer to each other that means the species composition is somewhat similar. If they are found apart with some distance that means the species composition is not similar to each other. If the stress value is less than 0.05, which may not come for all types of datasets (that means it is rare for the stress value in NMDS to reach less than 0.05), your NMDS is excellent. If it is less than 0.1 It is great, 0.2 is good and less than 0.3 then it is considered a poor representation. I hope this help
Dr Sileesh Mullasseri
- Badges
- Science method