Skip to content

metaMDS: unrealistic good stress

2 messages · Daniel, Tyler Smith

#
Hello,

i'm using metaMDS to ordinate trait mean data of plants in different 
ecosystems.
I used different calibrations but the stress differs only  between 0,001 
an 0,1e-19.
I have never seen such a good stress, so I wonder about possible 
mistakes, but i cannot find one.

Data looks like that:

plots      trait1      trait2      traits3
1            2,4           0,2         20
2            0,5           0,3         21 
.                .               .            .
300          .               .            .

My standard calibration is:
metaMDS(test
,distance = "euclidean"   # because metric data
, k = 3                                 # because I have 3 traits, but I 
couldn't plot a scree plot
, trymax = 3-5                     # not more, because too large dataset
, autotransform =TRUE
, noshare =1    # Ordination may be very difficult if a large proportion 
of sites have no shared species. In this case, the results may be 
improved with stepacross dissimilarities, or flexible shortest paths 
among all sites. No manipilation: noshare=1.0
,trace = TRUE         # will monitor the final stresses
,plot = TRUE       
,zerodist="add"         
)

thanks for helping me out

Daniel

-----------------------------------------------------------------
Daniel Hornstein  
University of Bayreuth, Germany
Department of Disturbance Ecology

phone: 0921//55-2173
#
Daniel <daniel_ho at web.de> writes:
You've got three traits, and you're using MDS to project them into three
dimensions (k = 3). There's no stress, because you've got as many
dimensions as variables. You only get stress when you try and squeeze
the variation from three variables into less than three dimensions. Try
again with k = 2 and you should get some stress.

Put another way, why are you trying to project your data into three
dimensions? If you want to plot it on paper, you need to project it into
two dimensions. If you actually are using the three ordination axes in
some subsequent analysis, you could just use your variables (possibly
transformed) instead.

HTH,

Tyler