Firstly your question is difficult to answer, as observed community composition is ALWAYS due to chance.

Secondly "differences in species abundance "  is perhaps something you want to measure, rather than dismiss? 

I think perhaps you need to further explain what you mean by "differences in species abundance ", then I will have another go at helping you.

Assuming all communities are environmentally similar and the same size. Ideally you need repeats of your treatments so that will minimize the effect of  a lot of the variation that is due to chance. Alternatively you need to be able to rank treatments, or have records over time. Then clearly identify what you want to assess, then see if your place of study/work has a statistician to help you.

Hi Gayle, thank you for pointing these unclarities. I have five treatments (actually 4 treatments and one control), each replicated four times, giving me a total of 20 sites. The 20 sites are surveyed four years. In each site I sampled the species present (abundance data), ending with a site x species matrix for each year.

I would like to know the impact of the different treatments on the species communities (if possible by assessing a difference to the species community from the control "treatment"). I am figuring that the difference between the "chance component" and the "true" treatment effect could be actually assessed by performing matrix randomization and by accounting for the variability among the replicates.

Hi Valerie

That is a good dataset, so you should be able to find a difference, hopefully.

so for each replicate you have present/absent/numbers for each species? 

and/or ages/gender/nests/seeds for each species?

or just e.g. there were 7 species in plot 3 in year 2.

How much do the species interact?  do they eat each other, and/or compete with each other?

were all the  treatments identical when you first measured them? that is did you apply the treatment, or assess one that was already there? 

Thanks for your prompt reply

Gayle.

Thank you very much for your willigness to help me.

I have numbers for each species but no further details on specific features such as gender and ages or else.

Whereas species interact is not easy to assess, such that my first investigation is more related to whether species co-occur. Then, in regard to the outcomes I should have discussion with experts on their opinion about patterns of interactions (of course I have some ideas beforehand, but not yet enough solid knowledge to integrate such information in statistical models).

There is a time zero with identical starting conditions for all sites.

Valerie

Thanks Valerie.

The easiest way is probably to analyse each species separately, at least initially. I would do a graph up for each species, with dots, with each treatment a different colour (giving 20 dots at each time, and 80 dots on the graph). Put time on the x axis and abundance on the y axis.  Then you will be able to visually see what species look interesting, and also if any species can be grouped for discussion because they behave in similar or contrasting ways. You will also see the effects of any weather variations. If numbers change markedly you may have to use a log scale.

You will also see if variation within each treatment is greater than the differences between treatments.  That is, do replicates vary more than treatments? If this is the case then you CAN NOT get significant results. One possibility is combining the data as suggested above, either across species or across treatments, which MAY overcome this. However you can't do that without a biological explanation, because if you fiddle around enough random chance will sometimes throw up something, significant, but meaningless.

Let me know how it goes, Gayle.

Dear Gayle,

Thank you very much for your suggestion. However it was not exactly what I was looking for. I cannot really look at each species independently (too much species and many not abundant enough to be treated independently), but rather I was looking for an analysis at the community level. I would not like (at least in the first time) to know exactly which species increases or decreases in abundance, but if the compositional structure of the species matrix is affected by treatment.

As you said if the variability among replicates is larger than among treatments, a treatment effect cannot be assessed. Back to my first question. Could a null model help me assessing any treatment effect and how to parameterize these null models (I guess I need at least one different for each year).

Hi Valérie,

PERMANOVA is exactly the method you need. It is a multivariate (dissimilarity-based) ANOVA, with p-values obtained via permutations.

In R, you can run it with vegan::adonis and also maybe with BiodiversityR package (I don't remember).

Post-hoc multiple comparisons (control vs treatments) could be done manually (multiple runs of adonis on subsampled data set) and then you correct p-values bu using "p.adjust" function

In PERMANOVA/adonis, for testing differences in composition, use jaccard on presence/abscence transformed data

for  testing differences in composition and/or abundances, use bray-curtis on sqrt transformed data (or another transformation, e.g. log if you need to down weight highly abundant species)

You can use both dissimilarity measures, tests are complementary.

You might also consider mvabund, which is multivariate glm

cheers,

Hi Pierre,

Thank you very much. I know PERMANOVA using adonis but I was unaware of the possibility to perform post-hoc tests. If I understand it correctly, you would perform several analyses, each of them containing matrices for control and for one of the treatments?

However, abundance of the different species varies annually because of differences in environmental factors and I would like to control for this (thus my idea of normalizing abundance data).The idea behind is that I want to be able to compare the results for each year independently of underlying differences in species abundance between years (I am not primary interested in knowing whereas species abundances change with time). But maybe this is not such an issue.

this is a very interesting question, as i am working on similar data too.

so far, I gather, in term of null model idea, it seems the abundance could be exchanged as well as species names in simulation, however i do not know how to do it, maybe need write your own R code?

on the second thought, if you calculate the residual of annual abundance data in some way (means cut off the effect of time), then use them to do next analysis. still, just a idea.

I would like to follow your progress. good luck!

Hi Jingming Zheng,

I find your idea of taking residuals really interesting. Still I see the effect of year as a kind of latent variable that will have a lot of factors correlated to it and I might loose some information by working only on the residuals (difficult to explain my thoughts). Also there certainly is some kind of interactions among year and other variable, such that a variable might be highly significant in one year and not in another. Not sure this can be recognized by working on the residuals.

But these are very early thoughts and I would be really interested to have further opinions about this.

Valerie

Valérie,

A simple design close to your own experimental design is the 2 ways crossed ANOVA design Year X Treatment, which allow to assess both between-years variations and between-treatments variations, as well the interaction term year x treatment. Hence, this allows to look at treatments effect even if there is also year effect.

However, it seems that your design is a repeated measurment. The replicates are dependent from year to year, since they were sampled in the same sites. The simple  2 ways design is consequently not appropriate, especially the permutation scheme. Maybe you could add a strata argument in adonis. Ask to Vegan developpers Jari Oksanen or Gavin Simpson, directly or through the R-SIG-ECO.

For graphical visualization, I would suggest partial dbRDA, have a look to vegan::cmdscale, with the arguments "distance.matrix~Treatments,partial=Years"

HTH

Hi Pierre, thank you very much for this suggestion. Yes I will need to control for the dependence among the replicates for correct inference.

The problem of directly including a Year X Treatment effect is that I am mostly interested to know the differences among treatments for each year and the temporal effect within a given treatment. This is why I did not mention the temporal aspect of the study in my question. All of the answers are highly interested but do not exactly address the original question (which I admit is not extremely clear).

Hello Valeria and others, 

I am also working on a similar data set- I have several years of species abundance data in  treated and non-treated coastal dunes 

(Sampled design: 4 treatment types + 1 control type, 3 replicate sites per dune type, and three plots of 12 samples (traps) per site)

I am also struggling with how to normalise across years. And I have been looking at four different taxa so I would like to also standardise between taxa too!

I have begun to look at RDA just taking each year and each taxa separately. I also am looking at principal response curves for the temporal factor but im not sure it means anything since there is so much inter-annual environmental noise.

I am about to attend an intensive multivariate stats course on CANOCO and will hopefully be able to address some of these questions and I will let you know. I shall ask them about Jingming's residuals idea. 

I look forward to working out some ideas with you and Im so glad to find other people working on similar issues!

Tania

Hi Tania,

Great that you have the opportunity to attend this course. I would be very happy if you could give some feedback about.

Cheers

I recommend a book with R script on muti-variates analysis method, "Numberical Ecology with R" by Daniel Borcard et al. 2012.  if you want work with CANOCO, certainly it will be very helpful as all methods by  CANOCO now can be done with R packages by other peers.

good luck. 

Valérie,

Legendre & Gauthier (2014) Statistical methods for temporal and space-time analysis of community composition data.

this paper presents a method  that might be suitable for your data. R scripts are in appendix.

http://rspb.royalsocietypublishing.org/content/281/1778/20132728#sec-10

Hi Pierre,

Thank you very much for this reference.

Ok, yes but one "normalizes" as a gesture towards meeting the requirements imposed by formal analytic methods in relation to a "normal" (or other kind of) probability curve. In point of fact one looses, or, at the least warps information from the reference data-set in the name of meeting such "requirements". The actual distributions that one records already tell a lot about what is going on at the population level. If one "normalizes" then something is lost or distorted. Do as much as you can with the raw data as you find it. In any event the result of so doing will be to emphasize the variance component making it harder to say one group might be significantly different from another

The more you massage and transform your data by normalizing it the further you will depart from what the ecosystems are trying to tell you directly. In brief you may be driven to more readily identify "isolates" or differences where no such differences exist as far as the interacting populations themselves are concerned.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7409812/