there may be the answer in Numerical Ecology with R

http://www.ievbras.ru/ecostat/Kiril/R/Biblio/R_eng/Numerical%20Ecology%20with%20R%20%28use%20R%29.pdf

Chapter 4 Cluster Analysis

I would use Ordination analysis altogether instead of cluster analyses. NMDS or a PCA (if your data are linear) for benthic communities (Also can be done on environmental variables) to differentiate sites or communities from one another. RDA or a CCA for correlating environmental variables to species. All can be done in PAST a free software available online.

You can use of CCA, DCA, PCA and NMDS, CCA using to relationship between assemblage and physico-chemical parameters, NMDS using for study of similarities and dissimilarities within assemblage. These analyses conducted using many softwares such as Primer; PC-ORD, R and S plus ...

Regards 

I wounlt recommend any of those clustering method, as previously suggested i'd recommend other ordination methods (e.g. nMDS or CCA, RDA, etc if you have other environmental variables)

I would prefer the ordination analysis in particular PCA for benthic communities and it can be used with environmental variables. However CCA, DCA and NMDS can still be used, as suggested a number of softwares can perform such analysis, i use PAST, a freely downloadable with a manual

I think you can use single linkage clustering to know the groups. Other ordination techniques (PCA, CCA,DCA, NMDS) are useful and they will tell you if the groups are similar to or different from the single linkage method in ordination space. Thereafter, if you have distinct groups, you test if the groups are well separated using ANOSIM (Analysis of similarity, (Clarke, 1993)). I have used this method in one of my papers on land snail diversity in Omo Forest Reserve.

I think you can use Cluster Analyses (CA), PCA or NMDS. 

nMDS in PRIMER or R package 'vegan'; PERMANOVA to establish significance of between-groups differences, CCA, PERMDISP or covariate PERMANOVA to examine environmental drivers of assemblage composition or difference; see Howell 'A benthic classification system to aid in the implementation of marine protected area networks in the deep/high seas of the NE Atlantic' Biological Conservation 05/2010 if you are identifying biocoenoses.

Thanks each and everyone of you for your answers. It's now clear for me that clustering is not enough. Cheers

You can use a powerful ordination method: The Self-Organizing Map algorithm (SOM Toolbox version 2 for Matlab), (see Vesanto et al. (1999) for practical instructions). Briefly, combining ordination and gradient analysis functions, the SOM is convenient to visualize high-dimensional data in a readily interpretable manner without prior transformation. Multivariate ordination and classification techniques may be inefficient at revealing patterns of community organization especially for macroinvertebrate community analysis due to many taxa with skewed distributions. The SOM algorithm is an unsupervised learning procedure that transforms multi- dimensional input data into a two-dimensional map subject to a topological (neighborhood preserving) constraint. The SOM, thus, plots the similarities of the data by grouping similar data items together onto a 2D-space (visualized as a grid) using an iterative learning process. The SOM algorithm is specifically relevant for analyzing sets of variables that vary and co-vary in non-linear fashions, and/or that have skewed distributions. Additionally, the SOM algorithm averages the input dataset using weight vectors through the learning process and thus removes noise.

The beauty of techniques toady is that whichever you use you can always validate your dataset. There are many software programmes that are easy to use and provide 2D and 3D visualisations. Average linkage, Ward or Single linkage are all good if you can validate with other ordination methods such as Principal Components. It is better to start with looking at the distributions to know how uniformly or skewed they are. That would provide guidance as to the type of ordination methods you can use to reveal the patterns in the macrobenthic community you are evaluating. 

Hi Bernabé,

just a word here why ordination and nMDS methods should be applied instead of cluster methods in your case.

One reason is that during the process of clustering encaptic groups are built out of the samples which are then treated as entities in the further process. This is not valid. Secondly, samples from nature are not inherently of hierarchical nature, but instead should be treated as being on the same level. Clustering, however, requires a hierarchical structure within the data. But this is rarely considered in biological studies.

Ordination and nMDS treat ecological samples in a more adequate way. Better use these.

Best regards

Armin

When you use cluster analysis, I recommend to standardize your variables (for example z-transformation, which results in mean=0 and SD=1 for the z.transformed variables). Ward´s clustering method (together with squared Euclidian distances) gives the best results - at least in my experience. 

If you have extreme values in your variables, you may need to do a log-transformation before you start.

good luck, Irene

Seems to me everyone here is recommending methods without addressing the question, which is about linkage options. 

Single linkage (nearest neighbour clustering) is related to ranks (advantage) but tends to produce 'chained' dendrograms with each sample joined to the next without clear clusters (disadvantage).

Complete linkage (furthest neighbour) tends to produce very stark widely separated clusters. 

Experience tends to show that group average linkage (UPGMA) or wards or something intermediate (flexible beta) tends to give the most ecologically interpretable structure.

For these methods a plot of cophenetic distance against similarity allows one to assess the extent to which the observed structure reflects the true underlying structure.  SIMPROF would also give an idea of the interpretability of observed structre.

Flat clustering (KR clustering) does not give a hierarchy, and could be worth trying.

Then, for reasons outlined above, you could try ordination/ANOSIM etc.

What is appropriate really depends on your data and what you want to know, which you don't tell us.

Hi Paul,

I think the question was in fact addressed by me and others. It was not about how these methods work in detail, but which of these would be most adequate for macrobenthic datasets. And here I gave a clear recommendation not to use these at all since other methods, which were outlined, would be more adequate. Maybe it was only my interpretation of the question, but I think the essence of it meant by Bernabé was on how to analyse these data adequately based on our own experience. But maybe he can clarify this.

Cheers

Armin

Hola Bernabe,

I don't know what are your research questions regarding your communities. However, I can assume that you have samples from different areas, depths, etc, taken at different times with samples or data with abiotic factors. So likely, you probably are searching for differences related to those factors. Cluster methods do not have an underlying hypothesis, however it is better suited to questions about how similar, close or related are samples using any give distance measure. Since you are most likely interested on differences (rathern than similarities), I would recommend you to use ordination method such as nMDS, MDS, PCA, PCO first. You can use those to explore dissimilarity patterns in your biotic and abiotic data, and them you can test hypothesis using PERMANOVA and relate your biotic with abiotic data with BIOENV, RDA etc.

Hope this helps.