-1

Having produced a Bray-Curtis dissimilarity with my Hellinger-transformed data (26 samples, 3000+ species/OTUs), I went on to build a MDS plot. I got the following metrics:

Dimensions: 2 
Stress:     0.111155 
Stress type 1, weak ties
Two convergent solutions found after 2 tries
Scaling: centring, PC rotation, halfchange scaling 
Species: expanded scores based on ‘ALG_Hellinger’

However, the corresponding Shepard's plot looked as follows:enter image description here

Which, although achieving good fits seems as if the BC dissimilarity has not enough resolution to differentiate across samples. Is this right?

Testing it through ANOSIM, I got the following,

ANOSIM statistic R:     1 
Significance: 0.001 

Permutation: free
Number of permutations: 999

Upper quantiles of permutations (null model):
 90%   95% 97.5%   99% 
 0.123 0.166 0.203 0.249 

 Dissimilarity ranks between and within classes:
                  0%   25%   50%    75% 100%   N
 Between               97 154.0 212.0 266.50  325 229
 Cliona celata complex 19  32.0  46.0  59.00   66  21
 Cliona viridis         3  26.5  37.5  48.50   60   6
 Dysidea fragilis      56  56.5  57.0  59.50   62   3
 Phorbas fictitius      1  18.5  48.5  79.75   96  66

And ADONIS told me the same:

 Permutation: free
 Number of permutations: 999

 Terms added sequentially (first to last)

      Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)    
 SCIE_NAME  3    7.8738 2.62461  43.049 0.85445  0.001 ***
 Residuals 22    1.3413 0.06097         0.14555           
 Total     25    9.2151                 1.00000           
 ---
 Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

This is, the differences among the samples are significant, but the MDS ordination seems somewhat misleading.

How can I test another aspect of the MDS or change anything about this analysis, if even needed?

Thank you in advance!

André

André Soares
  • 309
  • 1
  • 13
  • Why would you want to Hellinger transform your data before computing Bray Curtis dissimilarities? The purpose of the transformation is to subsequently compute Euclidean distances. It makes no sense to me to do what you are doing here. – Gavin Simpson Aug 02 '15 at 22:44

1 Answers1

1

I don't think that the Shepard plot is poor. Rather, it shows that your data are strongly clustered. This is consistent with adonis which says that most (85%) of variation is between clusters. It is also consistent with anosim which shows that within-cluster distances are much shorter than between-cluster distances.

Jari Oksanen
  • 3,287
  • 1
  • 11
  • 15