I'm conducting analysis of bird counts for my Master's thesis on effects of patch size and connectivity on birds of High Andean landscapes. My first goal is to use ordination analysis to figure out which bird species are associated to each of the different kinds of habitat (forest, transitional and open matrix). I have lots of environmental/spatial variables recorded, but I decided to begin with an unconstrained ordination, just labeling the sites with different colours according to habitat and checking which sites and which species seem to group together.
My data is not very good (for many reasons, one of them just not having had enough time in the field) but I'm trying to salvage it the best I can. I've ran a CA and a DCA on my species matrix, using vegan package in R, and the procrustes function shows me large (and quite chaotic) differences between the plots from one method and the other. Is this telling me that arch effects or compression of extreme scores is happening with the CA, and so I should opt for the DCA? Or is it just because the CA explains very little variation in the data (the first two axis amount to around 18% of total inertia), so sites and species will just float around with no real meaning when I do the DCA?
A little extra question - would it help me to get more variation explained if I remove from my dataset some of the rarest species or some of the ones that move around the most between the CA and the DCA?
First, your data is perfect, so don't say that! Second, ordination is really only a preliminary analysis, so don't worry too much if it doesn't tell you much.
Third, keep the rare species in because they are the ones of most interest. And fourth, you could correlate bird abundance with patch size as in the paper below :)
Article Effects of structural and functional connectivity and patch ...
Andrew Paul McKenzie Pegman, thank you for the encouragement! At this stage I'm sometimes panicking more than I should. You're right about ordinations being only preliminary analyses, and it all comes down in the end to actual statistical testing and I haven't even tried that yet. There's also other types of ordination that I can carry out and maybe will bring me better results - NMDS, for example. The ultimate goal will be, as you suggest, to correlate bird abundances to patch size. What I am trying now is, first of all, to separate my data in forest-associated communities and matrix-associated commuities, since I plan to test effect of patch size separately on each of them.
Hello Joao. For a multivariate analysis in this case I would recommend you to apply a PCA (Principal Component Analysis). It can show you several correlations at the same time. And I agree with Andrew, don´t take of the analysis the rare species, maybe they are the ones that can provide the most valuable information. But you always can take of outliers from a single group (or species in this case), which are simply data that, for any reason, is totally different from the normal distribution. I mean, if you have, let´s say 30 records for Species 1 corresponding to forest and 1 record of the same species in the transitional habitat, that would be clearly an outlier and, only for statistical purposes, you can remove that record from the analysis because it will influence in the resulting value, but of course you have to do this only when you are totally confident that you are dealing with an outlier. But, if you have just one record of a rare species, that maybe it´s behaving different from the rest of species into the same Family, you must use that data and try to find out the reasons for that difference. Maybe that will help you to explain why that species is so rare.
Freddy Cáceres, thank you for the feedback! Being able to discern what are valuable records of rare species and what are outliers is tricky but it might definitely be the key to get more informative results out of the analysis.
You have to consider there is no perfect method and you need to try several ones
with knowledge of the default and advantages of each (see literature for this)
PCA or CA are good if you know the main defaults like the sensibility of rare species -logic if you consider the principles of the methods, based on the weight of individuals and variables and similatities - or the artefact of the horseshoe. You can avoid this problem with the NSCA (non symmetric CA) and compare with the CA and PCA. This is data mining more than confirmatory statistics but it could be very helpful for having an overview of a complex dataset. ADE-4 R package and literature is perfect for this with a lot of ecological -and pedagogic-examples. This is a French school but with the main resources in English, and with a lot of very useful graphic tools to make very informative figures. Have a good exploration ! (so much interesting than any videogame)
João Manuel Cordeiro Vale Pereira
Starting with PCA is a good advise, you can reduce number of environmental variables either by choosing among strongly correlated ones or using a few PCA axes (which account for most variation in your system) as composite variables.
Then I would perform DCA and check correlations with the chosen variables.
- Badges
- Science topic
- Similar topics
- Biological Science
- Zoology
- Ornithology