I would recommend using dissimilarity measures derived from multiplicative beta diversity, as these measures are independent of mean species richness (see Jost, L. 2007. Partitioning diversity into independent alpha and beta components. Ecology 88:2427-2439). Then, as you correctly point, differences in composition may be derived from differences in species richness or from the substitution of some species by others. You may be interested in my 2010 paper (Baselga, A. 2010. Partitioning the turnover and nestedness components of beta diversity. Global Ecology and Biogeography 19:134-143), which deals with the separation of total dissimilarity in two components accounting for the two aforementioned patterns. You can use these indices to measure both the spatial and temporal patterns of variation in species composition. Hope this helps.

Thank you very much for your answer. I read your papers and seriously consider applying the method you suggest. How would you deal with sites containing a very small number of species? Can it bias the results?

I would further be interested in the dominance patterns of my species, which won't be possible with your method that is, if I understood it correctly, only for presence-absence data.

To get started, take a look at the options available in R, mostly in vegan (e.g. analysis of similarity) or mvabund(http://cran.rproject.org/web/packages/mvabund/mvabund.pdf)

There are also some nice ways of looking at species turnover, using shifts in mean rank abundance and 'rank clocks', which are explained in:

Collins, S. L., K. N. Suding, E. E. Cleland, M. Batty, S. C. Pennings, K. L. Gross, J. B. Grace, L. Gough, J. E. Fargione, and C. M. Clark. 2008. Rank Clocks and Plant Community Dynamics. Ecology 89:3534-3541.

Thanks for your interest in my work. Sites containing very few species are not a problem, unless their low richness is the result of incomplete sampling. If sampling is incomplete, beta diversity patterns would be biased, of course. If richness is low because of “real” patterns, then dissimilarity indices can perfectly deal with them. So it is not a methodological problem. It is true, however, that when richness is extremely low, then beta diversity patterns tend to be affected by stochastic events, as the addition or substitution of just one species can completely modify the proportions of shared/unshared species. The partitioning method is not yet available for abundance-based measures (although it will come soon). Finally, Dylan’s suggestions on R packages are really useful. If you are using the partitioning method, you may also be interested in betapart (http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210X.2012.00224.x/abstract ). Hope this helps.

I have a quick question for you, Andres... can betapart be adapted to other measures of biodiversity, including functional traits and the like?

Phylogenetic and functional beta diversity can also be partioned into two fractions accounting for true turnover or diversity gradients. For phylogenetic beta diversity, there is a paper already out by Leprieur et al (http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0042760 ), and for functional diversity the method has also been implemented by Villeger et al (a paper will appear soon in GEB). These methods have not yet been included in a package but the R code is available as supplementary material.

Great, i'll keep my eye out for it. Thanks!

An alternative/complementary method could be to test the similarity between the sampled communities by computing a similarity matrix on the taxonomical data base, and performing a two-ways Permanova (with site and year as factors). This method take in account both species richness and composition and the transformation of your data would permit to down-weight (or not) the effect of dominant species. You can do this analysis with Primer v.6, but I think you can also use the R software. Complementary tools such as nMDS (nonmetric multidimentional scaling) and SIMPER (Similaritry Percentage) would be very usefull to complete your analysis and study variations between communities.

Hoping that could help, Sincerely.

Many thanks for all your comments. I didn't know about 'rank clocks', I find this approach really interesting, although it should be difficult with only three sampling dates.

I thought about performing Permanova and displaying the community similarities using nMDS, but I have been very concerned about the recent paper published by Warton et al. "Distance-based multivariate analyses confound location and dispersion effects". They propose to look at the mvabund package, so I will look if it helps going further.

Well, if I had your problem I would probably use the method suggested by Nicolas, simply because that's what I'm used to. I would use something like a Bray-Curtis dissimilarity matrix using species abundances (if you have samples with very low numbers of individuals you can correct the behaviour of the Bray-Curtis index by incorporating a dummy species that has the same abundance in all samples). Then I would run PERMANOVA based on these matrices to test for differences and estimate components of variation due to year and site. Because PERMANOVA is sensitive to differences in multivariate dispersion, you can check if dispersions are significantly different using PERMDISP. However, I would use Principal coordinates analysis (PCO, also known as MDS) to look for patterns in unconstrained multivariate space instead of nMDS, simply because nMDS uses ranks and PCO looks at the dissimilarity values themselves (just for consistency with the PERMANOVA's philosophy). SIMPER gives cool results, listing all species by their contribution to the Bray-curtis similarity among samples within a group and to the dissimilarity among samples from different groups, but I think it works best as a complement to ANOSIM and nMDS instead of PERMANOVA. An alternative (more PERMANOVA-like) would be to use a canonical analysis of principal coordinates (CAP) to find axes that best discriminate among groups of interest (like a distance-based discriminant analysis) and then overlay Spearman rank correlation vectors of species abundances with the 2 canonical axes. This is also a nice way of graphically looking at patterns.

Hello Miguel, thank you for your answer. I have indeed some samples with a low number of individuals and I do not really understand what you mean with a dummy species. Could you explain a bit more as it sound very interesting. I should look at the differences between MDS and nMDS, I did not get it well yet and I will also look at SIMPER.

The original paper about adjusting the Bray Curtis similarity for denuded samples is here:

Clarke, K.R., Somerfield, P.J., Chapman, M.G., 2006. On resemblance measures for ecological studies, including taxonomic dissimilarities and a zero-adjusted Bray–Curtis coefficient for denuded assemblages. Journal of Experimental Marine Biology and Ecology 330, 55–80.

But the concept is quite simple. You simply need to add a species (you can call it species X or Bob or anything you want) that is present in all samples (like 1 individual in each sample). Then calculate Bray Curtis dissimilarities with that "dummy" species there. This simply corrects the erratic behavior of the Bray Curtis dissimilarity when you have very few individuals. The extreme case is where you have two samples with one individual each. Without adding the dummy species that is present in both samples, you would get a dissimilarity of 100 if both individuals are from different species or 0 if they happen to be both from the same species.

All methods I told you about are in PRIMER 6 with PERMANOVA extension (where you can add a dummy species automatically before calculating BC dissimilarities), which are both a bit expensive... however I think you can find most of them in R as well. Methods from the permanova package can also be run from specific free software from Marti Anderson's website. She created little FORTRAN programs to run things like PERMANOVA and PERMDISP and I think CAP and PCO too.

If you prefer menus to code, I know that the free software PAST also has some of these methods.

Hi all,

I would like to point that the discussion is now dealing with two different questions. One is the selection of the dissimilarity matrix, which is what I addressed before. Different indices provide different results because they account for different facets of dissimilarity. For example, Sorensen and Jaccard are monotonic transformations of multiplicative beta diversity, and are thus affected by turnover (species replacement) but also by richness differences. These indices can be partitioned to separate both components (as explained above). Bray-Curtis is an extension of Sorensen to account for variation in abundance and it is therefore also affected by richness and abundance gradients.

The second question is how to analyse the dissimilarity matrix, and multiple procedures are available (i.e. mantel test, canonical ordination, cluster analysis...). Here, the selection of one or another method depends on the question to be addressed. Anyway, it is a completely different question of how to measure dissimilarity. Moreover, any dissimilarity measure can be combined with any of the analyses (mantel, CAP, etc.), but the definition of dissimilarity is a previous (more basic) decision.

Hope this helps.

You can use Principal Component Analysis to graphically display the relation among sites on the number of species and their abundances per year. To do this you need to named each sample as the site name plus the year, and then request the biplot in the analysis not the points that will be the species. Whit this you will have a graph where you can see how the sites are distributed in term of richness and abundance, solset sites are more similar en both characteristics. An example of this approach an other for analyse including richness and species abundance are foun here: Barrantes, G. & L. Sandoval. 2009. Conceptual and statistical problems associated with the use of diversity index in ecology. Revista de Biología Tropical 57:451-460. Hope this helps.

Thank you Andrés for this important point. I admit I go a bit lost with the many possibilities and began to confound selection and analysis of dissimilarity measures. Both approaches require carefully thinking about how meaningful it is with the data at hand. I know that my data are limited by the amount of species and especially the large variance in species number and individual abundances between sites, which require decision about what to pool, if some transformation to reduce influence of abundant species is needed and consideration of what the dissimilarity measure puts emphasis on.

Regarding the proposition of Luis, what is the difference between using PCA or nMDS for displaying the relation among sites and species?

I think that PCA relies on the Euclidian distance. Is that the main difference?

About your question on the differences between PCA and nMDS. PCA is used to analyze similarities between the data. In the other hand nMDS is used to analyze dissimilarities. If you use nMDS with Euclidian distance the results will be pretty similar to calculate a PCA, therefore if you want to see differences between both methods, do you need to use Grower measures. Hope this help.

As previously suggested, we have been using similarity matrices based on presence absence and Bray-Curtis based cluster analysis, MDS based on time since fire, rainfall, etc. for 12 years of plant community analysis, reference site to restoration site. Also you might find it instructive to run a correlation matrix with your species assemblages to show which species tend to co-occur and which are exclusive. This was helpful for us discovering new indicator species of our reference community.

Thank you very much. Correlation matrix is a good idea, I may use it depending on my interest for specific species.

I've just been reading a paper in which community similarity was compared over time and space using "Multiple Regression on Dissimilarity Matrix" (MRM), carried out with the package ecodist. Does anybody know about this? Is it a good way?

The first basic procedure is to arrange/organize your data ordinally according to sampling sites per month for each year. Having done all that you would have your data linearly arranged per sites per year, making it easy to make a visual assessment of the data before subjecting it to suggested statistical analysis.

Thank you Azubuike. I do not have data for each month; I just sampled the sites once a year, which do not allow me to have clear time series. But you're right that well arranged data and good visualization are very helpful!

Eduard is right, but in addition Ozgul, I just wanted to say that for the importance of a factor (criterion), you should check the ECV (Estimated components of variation) in PERMANOVA. I PERMANOVA, ECV is a more accurate measure than the (pseudo-)F value, whilst the p-value is a measure of the strength of evidence to test the hypothesis. In contrast to PERMANOVA, the R in ANOSIM can be interpreted directly and is an absolute measure of the strength of the difference between groups; it is also scaled between -1 and +1 and can be compared between different studies. the PERMANOVA pseudo-F (or pseudo-t), on the other hand, is dependent on the degrees of freedom and is hence not necessarily comparable between studies. The higher the df, the stronger the evidence. Also, the pseudo-F value is dependent on the within variability of the groups since it is scaled against the between-group variability. So if you have a highly variable group, the differences with another group may be attenuated because of the large within-group variability. So for PERMANOVA it is the ECV that tells you somethingabout the effect size and it is the p-value that should be used as a measure for strength of evidence.

Furthermore, unlike ANOSIM, PERMANOVA partitions the variability and gives you a better idea of which criterion(or interaction) is causing the variability.

I would go with PERMANOVA, particularly since ranking of dissimilarities (ANOSIM) results in loss of information and may result in less power, and ANOSIM doesn't allow partitioning of variability.

Hope this may be of use, Best wishes, Jeroen

and what about adonis in R package vegan? I think it is quite similar to PERMANOVA and superior to ANOSIM

Yes, vegan package in R is maybe the most powerfull tool to study the biodiversity along space and time.

Thank you very much. I know vegan well and it's a really nice package that I am using a lot, but could not find a really suitable method to analyse between year changes in species composition, especially if it would be possible to separate spatial from temporal changes.

I've read some time ago about Partial Triadic Analysis, and a method called STATICO.

It uses co-inertia matrix, instead of correlation, to do PCA, or CCA. The examples are exactly like what you have in hands- 3 years of species counts in several sites.

The analysis is performed with ADE4 R package.

I'm doing something similar myself. Look into this method; http://onlinelibrary.wiley.com/doi/10.1111/2041-210X.12137/full

The differences in species composition of communities among time and space can be assessed using Multi-response Permutation Procedures (MRPP) (McCune et al., 2002; Van Sickle, 1997). MRPP is a nonparametric statistical method that can be used to test the hypothesis that there are no differences in species composition between both or more seasons or places. Values of relative frequency of species are often transformed prior to MRPP analyses in order to improve the linearity of relationships and reduce the range in values (Peck, 2010).

As complement, is possible to perform a similarity percentage (SIMPER) analysis. Useful to identify the species that contribute most strongly to the dissimilarity between the two seasons. SIMPER (Clarke, 1993) is based on the decomposition of Bray-Curtis dissimilarity index. The SIMPER functions performs pairwise comparisons of groups of sampling units and finds the average contributions of each species to the average overall Bray-Curtis dissimilarity.

Both analysis are availables on R packages.

Good research!

Thank you very much. I am really interested in the new method proposed by Johannes. I downloaded the article and will try to understand the approach. I'll contact you if I need additional information. I will also consider MRPP and look for applications related to spatio-temporal analysis.

In general we use bray curtis dissimllarites to indicate species turnover between the assemblages or sample

I would agree with Miguel. These analysis you can do in PRIMER with PERMANOVA add ons.

@Sutomo Sutomo and @Valérie Coudrain

You can do PERMANOVA in R better than any other statistical software.

Yes, Permanova is good :-)

Spatial and Temporal differences (Beta Diversity) can be assessed with the PERMANOVA method. The PRIMER software package is not needed to do so, you can use the adonis function in the R 'vegan' package. 

The PERMANOVA method is relatively insensitive to heterogeneity in dispersions when compared to ANOSIM, and the betadisper function in the R vegan package can specifically address this question. Marti Anderson gives a good analysis of the methods in Ecological Monographs (Anderson and Walsh, 2013, click below for link). Using ANOSIM has many shortcomings, an important one is Type I error when faced with heterogeneous dispersion. Instead of using one metric, you can use both 'adonis' and 'betadisper' functions to address your questions.

These methods are limited by the distance-metric used, so use caution. If you are looking at insects, then you are likely to see huge 'blooms' in the numbers seasonally, and spatially. You may want to model specific taxa, the negative binomial would be a good starting point.

Best Regards, hope this helps

http://izt.ciens.ucv.ve/ecologia/Archivos/ECO_POB%202013/ECOPO2_2013/Anderson%20y%20Walsh%202013.pdf

 https://www.researchgate.net/post/What_are_the_criteria_for_selecting_variables_for_PCA_analysis_in_Primer_6_how_to_screen_variables

Use Similarity and Dissimilarity indices . but if you are ding work on same place then should have use identical quadrates as previously used.

Hello colligues I have never done Permanova analysis before but would to predict the risk of preterm birth by looking on wether difference in vaginal bacteria composition exists before and after birth? Is that correct to use permanova! Expecting my data frame contains a a column of preterm birth PB (present/absent) over 37 weeks and column of species composition (SP1, SP2 SP3 etc) is Bray-Curtis dissimilarity index appropriate for such a design? Help please

Hello, I recommend the beta diversity indexes. These indexes are very useful and currently very easy to make. I recommend that you see the following available works: https://www.researchgate.net/publication/284554411_Compositional_similarity_and_beta_diversity https://www.researchgate.net/publication/46417414_A_review_of_beta_diversity_studies https://www.researchgate.net/publication/316917854_Beta-diversity_partitioning_Methods_applications_and_perspectives If you are interested in the R program there are several packages that allow you to calculate these indices.

Hi, I also recommend these jobs, they are very valuable

Article Analyzing beta diversity: Partitioning the spatial variation...

Article Beta-diversity partitioning: Methods, applications and perspectives

Article Beta-diversity partitioning: Methods, applications and perspectives

Similarity and Dissimilarity indices can be used.

Interesting research. Is this type of research carried out to explain specific processes of species evolution and phyletic flora and fauna?

PERMANOVA with similarity and dissmilarity indices is the best option here with PRIMER or using R package vegan in R. nMDS plot will give distirbution in 2D or 3D. Bupple plots can be used to explore the data.

Very good and interesting discussion

Firstly the community structure to be analysed and then to go for variation with these values through ANOVA