the analysis of biological traits with settlement approach is very interresting. There are statistical tools, especially in R (ADE4 library). This is the analysis RQL. There is a tutorial (in French): http://pbil.univ-lyon1.fr/R_svn/pdf/ter7.pdf

but I do not know if this analysis is to fit only one variable (1 biological trait). Other approaches are unlikely to be possible as discriminant analysis ...

Thank you. I tried with rql, but the results are indeed not straightforward to understand with only 1 trait and 1 explanatory factor (isolation).

Hi Valérie,

First, you could check for differences in assemblage structure between levels of site isolation. You could visualize patterns using PCoA or CAP (function vegan::capscale in R) and test significance using Permutationnal Multivariate ANOVA (function vegan::adonis in R, or PERMANOVA in PRIMER). However, you have between 3 and 4 replicates (site per isolation level), isn't it? Know that this is acceptable but really low. As regard to the dissimilarity measure, Anderson et al. 2011 stressed the need to use distinct measure, which emphasis distinct facets of assemblage structure. The most used in ecology is Bray-Curtis, which emphasis both composition and abundance patterns. I would start by using this one, with previous square root transformation for down weighting highly abundant species (if any).

Then, if you detect differences in species composition and/or abundance between isolation levels, you could check which species are specialized to a given isolation level. For doing so, you could re-use your CAP and overlay arrows representing the correlation of species with the CAP axes. These arrows (one per species) could be colored as function of your apriori specialization levels. If your arrows are aggregated by colors, it would mean that your specialization levels were good apriori. See, if I well understood, some species are apriori generalist and you don't expect differences while some others are specialist and you expect differences. Hence, these hypothesis would be supported if you get long arrows for specialists (in direction toward their specialized isolation level) and short arrows for generalists (symptomatic of non significant species/axe correlation).

Another complementary approach, which is more inferential, could be to split your data in 4, depending on your apriori specialization levels. Then you could, level by level (of specialization), check for differences between isolation levels using PerMANOVA.

I don't know if you are familiar in using R. It would be the best because it is flexible (e.g. I think to arrows colors, I don't know how to do it in Primer). If you are not familiar but have time (a lot), I advice you to use R. If you are in a hurry, PRIMER/PERMANOVA (which is not free at all) will efficient be more than enough.

I hope this help, good luck.

Pierre

Anderson et al 2011, Navigating the multiple meanings of b diversity: a roadmap for the practicing ecologist

A simpler approach could be: An Analysis Within / Between. PCA or AFC analyzing the variation between groups, or in a group. Between the analysts focus on the difference between the groups. This may be a solution.

Hi Pierre,

You exactly got my point. I have 10 sites pro isolation level. I already performed ordination of my species matrices and I have indeed differences in species abundances and composition in relation to the level of isolation. Your second part is what I need. Can you give me a little bit more information how to compute for each species the correlation with the CAP axes? This idea of visual display sounds very good. I am working in R. I think that vegan should make it, but I am quite bad in plotting beautiful graphs.

Cheers,

Valerie

I had a look at ade4 and the RQL analysis. It may be an option but I find the interpretation of the results quite difficult. Especially in regard to the relation between the trait and the factor, I don't manage to get the relevant information from the outcome.

This is applied by multivariate analysis via running the environmental factors against the species matrix by which the influential factor can be identified. This is can be dione using the routine BIO-ENV in PRIMER Program. To identify the species detrmine the spatial variations among the sites, the routine BVSTEP need to be applied.

Thank you Thamer, unfortunately I have never used PRIMER, would you know of a corresponding function in R?

Hi Valérie,

I suggested to do a PCoA with Bray-Curtis. As, Alian Ponsero said, you could also do a PCA, but you should before Hellinger transform your data in order to exclude joint absence while using euclidean distance.

For fitting a PCoA, use vegan::capscale with as RHS formula ~1. This means that you ask a CAP without constraint, so a PCoA.

Then, you can use vegan functions for plotting sites and species scores. I prefer extracting coordinates from capscale object and plotting using ggplot. Have a look below to codes, you could easily adapt.

Thamer Al-Dawood suggestion is really interesting too. In R, you could use a similar approach with indicspecies::multipatt. This detect and test which species are indicator (high occurence and/or abundance) for one or several groups. Use the argument duleg=F in order to enable some species to be indicator of more than one isolation.level (not generalist neither specialist).

Cheers,

Pierre

# codes never tested, let me know if it doesn't work.

# let's say bio is your sites X species matrix with abundance data

# env is your matrix sites X environmental factor, contain isolation.level

#ref_sp is a matrix with 2 colunms, species.names and specialization.level

species names of bio must match species names in ref_sp

tbio=sqrt(bio) # for down weighting highly abundant species

mcap=capscale(tbio~1,distance="bray")

#sites coordinates

dims=c(1,2) # for the 2 first CAP axes (which are PCo axes because of formula ~1)

site=scores(mcap,display="wa",choices=dims)

site.env=cbind(site,env)

#species axes correlations

cor.min=0.3 # threshold under which species are considered weakly correlated to the 2 first axes and consequently will be not plotted

cor_sp=as.data.frame(scores(mcap, dis="sp", scaling=-2, const = sqrt(nrow(tbio)-1),choices=dims))

names(cor_sp)[1:2]=c("CAP1","CAP2")

cor_sp$cor=with(cor_sp,sqrt(CAP1^2+CAP2^2))

cor_sp$sup=FALSE;cor_sp$sup[cor_sp$cor>=cor.min]

Hi Valérie,

Have a look to this chapter book:

http://www.planta.cn/forum/files_planta/chapter_10_analysis_of_ecological_distance_by_ordination_162.pdf

Pay attention to the scaling (1 ou 2) when computing species scores. Interpratation will differ. Maybe scaling=1 is a better option in your context because arrow shows the direction from the origin for which sites have larger abundances for this species. While scaling=2, is used to reflect the correlations among species. Species that have a small angle between their arrows are expected to be strongly positively correlated. (see p8 of the pdf file at the URL above)

Cheers,

Pierre

Thank you very much. I will try and let you know how it is working. I did not know about the package indicspecies, I should definitively have a look at it, I may use it for further analyses too.

Good point to make me aware of the scaling issue, I usually do not care much about it, I'll be more attentive in the future :)

Cheers

What do you mean by R? there are correlation coefficient (r) and the R2. In case you need a help on Primer, happy to do the analysis with the explanation.

Sorry Thamer, I spoke about the statistical software "R".

Hi Pierre, After some trials and failures I'll manage to get the plot. It looks really nice, thank you very much!

I will have a closer look at the ggplot2 package, but maybe you could suggest me a way to implement a hull around the sites belonging to the same isolation levels (I guess it should be feasible).

Cheers,

Valérie

Happy to help you. For hull, try this

######################### compute hull coordinates (ignore warnings)

hull=with(env,ordihull(mcap,isolation.level,lty=2,draw="none"))

coor_hull=data.frame(do.call("rbind",hull),env[row.names(do.call("rbind",hull)),])

names(coor_hull)[1:2]=c("CAP1","CAP2")

#plot

mon.plot2=mon.plot1+geom_path(data=coor_hull,aes(x=CAP1,y=CAP2,group=isolation.level),colour="black",lty=2)

Thank you very much, I really appreciate your help! The formula does produce an error, I had to specify a name for the third column of (coor_hull):

names(coor_hull)[1:3]=c("CAP1","CAP2","isolation.level")

However, the hull doesn't match the points...there should be some problem.

Soory, I forgot to adapt my codes. Here is something that should work

hull=with(env,ordihull(mcap,isolation.level,lty=2,draw="none"))

coor_hull=data.frame(do.call("rbind",hull),env[row.names(do.call("rbind",isolation.level)),])

names(coor_hull)[1:2]=c("CAP1","CAP2")

should work, not so sure. If it doesn't, you could send me your whole code and I could homogenize object names between the different code parts I sent you.

Mean while, you could also try:

env$hull=env$isolation.level

hull=with(env,ordihull(mcap,hull,lty=2,draw="none"))

coor_hull=data.frame(do.call("rbind",hull),env[row.names(do.call("rbind",hull)),])

names(coor_hull)[1:2]=c("CAP1","CAP2")

Hi,

You raised me doubts so this morning I posted a question about scaling method for species scores.

http://stats.stackexchange.com/questions/49487/species-scaling-methods-in-pcoa-or-cap-using-vegancapscale

Look at the codes therein, everything works. You will have spider instead of hull. For getting colored arrows, do as I mentionned above: you have to merge cor_sp and ref_sp by "specialization.level" and add color= specialization.level in the aes() of geom_segment()

cheers

Pierre

Great thank you!

Hi Hillary,

I previously tried to investigate my data with regression trees, but it was really difficult to interpret. However I am interested in the mvpart package, thank you for making me aware of it.

I found a good way to analyze this kind of data. It is not much used, but merits attention: the fourth corner analysis implemented in the package ade4. Reference: Dray & Legendre, Ecology, 89, 2008, pp. 3400–3412

Thank you Alex, actually 4th corner analysis is more suited to my data, because I want to investigate a single trait at a time. I think that RLQ is very useful to investigate a combination of traits (traits are combined in a distance matrix). Correct me if I am wrong.

Cheers

I did not know this variation analysis RLQ, ...very interesting.

Hi Valérie, another approach that you could use (maybe more simple but also more easy to interpretate) is with path model and structural equations. Since you are interested to test a single trait, you could calculate the community weighted mean and use the species abundance as a weighting factor. You can also calculated some other metrics that measure trait variation such as kurtosis or coeffcient of variation. You can have a index of isolation per plot, this will explain others variables that you did not measure. The path analysis can explain realtionships between your environmeantal factors, isolation, and x metric of community functional property. Cheers

Hi Geovana, I made some analyses with structural equation modelling before, but I don't see how to integrate multivariate data and categorical explaining variables within. Maybe you could give some additional information or an example from a paper?Thank you very much

Hi, Valerí. I have the same situation; I want to identify what environmental factors influence the distribution of a species studied. I made two anyleses; first, I modeled the potential distribuction (any method that you selected) and made a correlation analysis of the occurrence and environmental variables used to model (i.e. precipitation, slope). Second, I made a varpar analysis grouping the variables in factors (climate, vegetation, terreain and human effect). This last analysis permited to identify what factor shows higher influence in the area modeled of the potantial distribution.

Hi Carlos. Thank you very much. Variation partitioning is indeed a good way to analyze the contribution of different environmental variables to the distribution of species. But how would you include a species trait, such as specialization or body size?

Dear Valeri, would you may give examples of you factors and funtional traits that you are measuring?

Geovana

Oh yes of course. I collected Hymenopteran species in trap-nests in 30 different sites.

10 sites were isolated, 10 sites connected and 10 sites adjacent to forest. Each site had a different amount of forest cover. The collected species are classified according to a specialization level. I thus have a site x species matrix and a defined trait for each species. I would like to investigate if the distribution of the species among the different sites is related to isolation level or wood gradient and if this distribution can then be related to a certain trait (f.ex. if most specialized species occur in adjacent sites). Can you better see the kind of data I'd like to analyze?

Cheers

Hello.

I think that a good and simple way to analyze this kind of information is a Chi-square test. The null hypothesis is that at all levels of isolation, the species proportion is the same of the global species proportion.

Thank you. Yes I could investigate the proportion of species with a defined trait and find out with a Chi-square test if some traits are more linked to specific isolation levels. This is a way that summarize information from the individual species, I have to think about the value of a species-based approach compared to a global trait-based approach.

HI Valérie, I understand better your data. I think you could do two ANCOVAs, one per each traits:

Factor 1: 2 levels (isolatted/non isolated)

Covariate: Forest cover or wood gradient (if I understood well)

Dependent variable: body size (in case of categorical traits you can use binomial distribution, but I am not totally sure about it).

What you will know with an ancova is if there is a difference between body size for isoleted and non isolated places and if this difference is influenced by forest cover.

Hi Geovana, thank you for this suggestion. I have indeed a categorical response variable. I am not pretty sure, but maybe your approach is quite similar to that of Fernando but for a continuous variable, since I think that categorical variables may best be tested with a Chis-square test.

Don't hesitate to correct me if this is wrong.

Hi again, I do not see clear how with Chi2 your can use your covariate, I think it is not possible, maybe I am wrong.

Hello Valerie

You are thinking to use a Canonical correspondence analysis to analysi your data:"this is correspondence analysis of a site/species matrix where each site has given values for one or more environmental variables". Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

Can any one help me to define man

A man is an adult human male. But I think this is not the topic of this thread :-)

A general comment and a suggestion: Instead of trying to figure out the best statistical treatment of data AFTER you collected the data and then looking around for possible hypotheses to be tested, why don't you try to set up a sampling design with relative methods of analysis BEFORE collecting data in the light of certain clearly stated hypotheses to be tested? That should be the correct way to proceed for scientific hypothesis testing. Often "A good definition of a problem is worth more than its solution" (A. Einstein)

Why do we always think that everything that Einstein uttered had to be taken seriously? If one has a problem, surely the most important factor is its solution. One has to define the problem before one can commence working on a solution but the definition just cannot be worth more than a solution.

You are perfectly right when you say that it is rather obvious that one has to define the problem before one can begin working on its solution, and this has to be taken very seriously in science expecially in problem definition in ecological and environmental research. My personal concern, and the arena of RESEARCHGATE provides plenty of examples, is that most of researchers do not care much of this stage that is fundamental for hypothesis testing.and inference and, instead, they look around for a "technical solution" (mathematical-statical techniques like advanced multivariate analysis, asymmetric anova, and other very sofisticated modeling techniques) once they have collected data does not matter how. In other words, they are looking for a technical solution to a problem that is not technical. Hence, in this condition, the ultimate solution (answer) to a problem is often unsatisfactory and invalid for inference. That is why a very good definition of a problem create the condition for having its solution at hand.

You are absolutely right saying that data analysis should be thought about before collecting data. However, once data have been collected, some additional questions may arise from the data at hand. Because of the difficulty of collecting field data, I think that it may be worth trying to answer these further questions even if the data had not been collected for this particular purpose. Even if there will be some error that we have to acknowledge, I find it unwise to left aside some potentially insightful data just because they were not expected or clearly statistically framed beforehand.

It goes without saying that a problem has to be defined before attempting trying the solve it. It is unscientific to embark on a project without some idea of sampling and analysis. However, I also agree that all sorts of things can come out of data, some of which can be unexpected, and these are often a bonus. A researcher always has to keep an open mind.

I was making the philosophical point that the definition of a problem cannot be considered more important that solution of the problem itself. The methods for solving the problem should be considered first in any research, but if one has got that far, one has already defined the problem.

Hi Valérie, I perfectly agree that it would be unwise not to treat data for some answer as byproduct of the original scope of the research, but in this case you should say "once data have been collected and treated statistically for the original purpose, then...". This is part of the knowledge process, every experiment will lead to other new questions that, in turn, need to be answered. But the problem is now when will you stop exploiting old data and decide to conduct a new survey to better target the new questions? The risk is being captured in an uncontrolled loop moving you away from generalizable knowlege and from valid inference.

Hi Patricia, I see your philosophical point, and I can agree on that in "absolute" terms, however the "problem" is slightly different if seen within the scientific knowledge process and within the current practice of research in the real geographic world. See my answer to Valérie.

Best

Giovanni, I think we are discussing things at slightly different levels. I was taking Einstein's comment at face value. Of course, one can make deeper interpretation of the simple statement, but I was making a simple challenge to the actual words.

Valérie, the problem you present requires the scaling-up of trait data from the species level to the community level in order to analyze the linking between traits and the site isolation factor. The most common solution is to find community weighted mean traits, but there other methods involving fuzzy weighting of the species abundance matrix, which you will find in my publications on this topic.

Thanks Valério, I'll have a look at your publications

Would say that many traits like juvenility (generation time), mating strategy and perenniality are acting on genetic drift. Probably Remy Petit is the right reference for this and his http://www4.bordeaux-aquitaine.inra.fr/biogeco_eng/People/M-P/Petit-Remy deserves attention. For computational analysis the ADE4 library for R-project gives hints and procedures

What sort of species?? my very late answer was meant for plants. As far Einstein and good practices there are the obvious limits of a posteriori procedures. Paleontology would not exist following Patricia and Giovanni strictu sensu. Impact analysis or metanalysis do not follow a scheme shared as the good one.

Hi Guido, I did not sampled plants but insects. However, ADE4 library provides very useful examples of possible analyses. These are I noticed not very broadly applied, because quite restricted to research carried at the University of Lyon. Sadly.

It occurred to me that the meta-community concept and nestedness analysis might produce something of interest

Inside ADE4 which is Lyonnaise try to find the analysis of lizard distribution of islands in Cortez sea; apart peculiarities non simmetric correspondences and Mantel test should be of general application