Hi Vassilis

agreed with previous comments. PCA and RDA are very similar is what they do. Although, they differ as PCA is unconstrained (search for any variable that best explains spp composition), whereas RDA is constrained (search for the best explanatory variables). It depends on the gradient lengths (tested with a DCA or DCCA).

Hope this helps,

good luck, Cesar

State the questions your are asking, as a series of questions for example. And then see whether your RDA can answer them all, or if you need a model with different assumptions. Then also we can maybe help you a bit more. The PCA and RDA are very similar is what they do put they are presenting information about the variation in your factors in different ways.

I did RDA using both biotic data and env.variables and I have triplot graphs showing the species, the gradients of env.parameters and the samplingstations. PCA can be used to order-classify the sampling stations using only the env.variables or only the biotic data (species). I did all three cases and the sampling stations show similar ordinations. Should I include all or just RDA?

You have just described what you did, not why you are doing it.

What questions are you asking of your data?

Sorry for not giving my question well. I want to assess the response of functional groups (biotic data) to the environmental gradients (abiotic data) and to classify the sampling stations using both abiotic and biotic data. I think that for the above RDA gives me this answer fully. What I do not understand is that since I have both biotic and abiotic data I can not find a reason to use also PCA. In the CANOCO maual they say that PCA completes the RDA but I do not know in what way?

Dear Vassilis Aschonitis,

If you have both biotic/species and environmental/abiotic data, in such cases you can use the direct gradient analysis (i.e. either RDA or CCA). If you already tested the linearity by DCA (shorter gradient/axis length in DCA axes i.e. the environmental variables have linear response on species/functional groups), in such cases you can use the RDA instead of CCA. You should not confuse, if you have no environmental data and data show tendency of linear distribution in DCA, in such cases you can use the PCA to see distribution of your species/functional group/ or your testing variable. I think you will be clear now.

For more information, you can see: ter Braak and Smilauer (2002), Leps and Smilauer (2003), Prentice and Ter Braak (1986), Palmer (http://ordination.okstate.edu/overview.htm), etc.

On a side note, if Euclidean distances or chi-squared distances are not the best wat to describe similarities/dissimilarities among your data points, you can also consider distance-based equivalents of such analyses, such as distance-based RDA, metric Multidimensional scaling (Principal coordinates analysis), CAP (canonical analysis of principal coordinates), etc. With these you can use other measures, such as Bray-Curtis dissimilarities. You can use basically any measure, by assuming it accurately describes the concept of "similarity" or "distance" in your data.

Just to let you know there are also other approaches (and maybe confuse you even more). :p

First, PCA is NOT designed for this type of data you have. It presumes linear response function. It cannot reveal, from principle, relationships you want to discover. Even CCA and RDA produce distorted plots, as documented many times.

If you have a single (prevailing, dominating) environmental gradient, you might like to explore http://R-Forge.R-project.org/projects/seriation/, http://lyle.smu.edu/IDA/seriation/

If not, you might want to try out http://cran.r-project.org, http://vegan.r-forge.r-project.org/

or http://pbil.univ-lyon1.fr/ADE-4

HTH,

Peter.

All I would add to the new answers is to stop worrying about what others have done and just be happy you answer your question in the way you wish to. I like the ADE4 package too, but if you want to switch to R before you pick a package I suggest you look at "Intro to applied Multivariate analysis in R", it is a great text to have to hand and to wade through the different MV methods..... although it does not include ADE4 (they may have their own book soon, if not already)

thank you all for your kind contributions! I worked with the CANOCO because the papers that I read used it and it is quite easy, fast and comprehesible. The manuals are very clear. I would suggest you to use it.

Hi Vassilis

agreed with previous comments. PCA and RDA are very similar is what they do. Although, they differ as PCA is unconstrained (search for any variable that best explains spp composition), whereas RDA is constrained (search for the best explanatory variables). It depends on the gradient lengths (tested with a DCA or DCCA).

Hope this helps,

good luck, Cesar

Hi Vassilis,

Ter Braak and Smilauer also mentioned somewhere in their book the dual nature of RDA and CCA. I think even if your gradient length are short (

I did DCA for two cases a) for the taxa abundance and b) function groups composition. In the first case the greater gradient length was equal to 3.6 and in the second was less than 2. For being sure I did afterwards both CCA and RDA for both taxa and function groups and the RDA described shlightly better the % variance in both cases, so I finaly choose RDA.

Hey Vassilis,

So PCA and NMDS and other 'unconstrained ordinations' examine the full extent of variation within a community data set. Maximising variation based on the dissimilarity measure used (and obviously how it does this varies by ordination). The above statement of PCA not being suitable I believe is 100% correct and is far better suited to decomposing physicochemical environmental data sets and examining them otherwise. I preferably use NMDS or non metric multidimensional scaling (aka MDS) for examination of community data. Although this issue with this is that gradients or axes obtained are nonlinear. You can then apply environmental fitting proceedures to this data for instance ??vegan::envfit if you use R at all. And I would personally recommend vegan for ordinations and the documentation that goes with.

Secondarily there are a range of 'constrained' methods. These constrain data by several environmental variables maximising variance based on these. Where by the axes represent linear combinations of these variables. So the resulting output is actually rather different and tells you something quite different about the data. As Leps and Smilaeur will suggest these methods a complimentary. You may also partial variables, in CANOCO and R Vegan package, where they have a statistical influence on the data, but are not of importance to your particular gradients being examined. RDA, dbRDA, CCA are examples of constrained ordinations and all have papers supporting their use. I've found dbRDA pretty handy, because as someone above mentioned, you can specify the dissimilarity measure used.

But also as was mentioned start out with a clear picture of what you wish to obtain and it'll make things a lot easier on you. Multivariate Analysis of Ecological

Communities in R: vegan tutorial Jari Oksanen (2011) is a good place to start if you wish to use R.

Cheers,

Jon

Vassilis

I think what you must understand first of all, are the variables you using in your analysis, this is what I mean for instance. If you have T°c, S°/oo and Density for example, you should not involve all of them because density depends on T and S, so you better only use density. Are macroinvertebrates affected by density? If not, then use temperature, because salinity has little variations and the sensibility of your analyses wuold not be as fine as you are looking for. An so you can make the same with all variables you have, redundancy starts with the variables.

Hi Vassilis George,

I guessed you already got an answer to your question. Here were several answers and also answers for so many answers too. Community analysis itself is very complicated and so many hypothesis have been borne each day as you know.

The descriptive analysis in ordination have so many times told and explained by earlier that we must not forget zero inflation in the sample by species data matrix. Here what I saw you having two different values of the first axis gradient lengths. Yes, I fully agree Mr Shanker's answer if the matrix follows normal norms.

Instead of running RDA or PCA I recommend DCA would be the best choice with over fitting of environmental variable. But you must not forget to check zero inflation at first.

Goodluck.

Chitra

Dear All

Seasonal greetings!!!

Could RDA be performed in SPSS or R software. I have to employ this stats for my thesis . Inform me please.

contact: [email protected]

Thanks

Rizwan

Hi Rizwan,

sure you can perform RDA in R as following :

library(vegan)

library(ggplot2)

require(XLConnect)

#load the data as an excel sheets

wb

Hi Rizwan and Abdulfattah,

I think syntax you have given above would be turned up to PCA not RDA. I suggest please follow this very nice link:

http://www.fromthebottomoftheheap.net/2012/04/11/customising-vegans-ordination-plots/

Goodluck.

Regards,

Chitra

Hi Chitra,

thanks, I guess the example you have giving can not be generalized with other type of data, as far as I knew that  dune and dune.env data are so limited for simple rda calculations , you can summarized the data then it become so simple to be plotted

by using  the summary of the rda func

smry

The Canoco software is free?

Dear All

I have applied some biofertilizers in the field soil to know their impact on soil variables (enzymes, ph, EC) and plant growth variables (plant height, fresh and dry weight branch etc.). What kind of ordination method can be used in this data set? For instance, PCA, PCoA, CCA, DCA, RDA etc. Also how to frame the data matrix? Should all the data in a single excel file? Data set of column and row is also important. Please suggest any link.

Regards

Rizwan

+919412819870

Email: [email protected]

Hi, Abdulfattah Alghurbani

i used your given code for RDA, but facing problem for running it. Especially in loadworkbook function. Could you guide me please?

Regards,

Tatoba