Hello Paul,

I think you must be more specific regarding your question.

Please note that you need to take a decision regarding the measure of species diversity for a site. In the old days when I was dealing with this type of problems usually there 3 measures were used: species richness (just number of observed species), species richness in standardised (for size – number of individuals  – (I am thinking about rarefaction curves)) sample, dominance index, Shannon diversity index. Another thing is that you may use data pooled across all the sampling time or at specific time points.

I imagine that once you have the right measure you can correlate this with some environmental gradients (another important thing is whether the gradients are constant over time or changing).

Also the Zero Inflated Poisson Regression Model can be used in different contexts (I can think. I am not sure if I understand how you would like to use this model to build a matrix of similarities between sites.

Probably I could help you as in the past I was working as a statistician supporting a team dealing with biodiversity data but you must be more clear regarding what you need.

Bartosz Jenner

Hello Bartosz,

Thanks a lot. I did calculate Piloue's evenness index and now the job is to relate, how that index change with salinity and temperature. I did use glmms with binomial quasi family. However the part I'm not still sure is " Community data is collected from various sampling sites within estuary" and this is done for months and years.

I wonder if it would be a good/bad idea of using Sampling site as a random variable or not?

cheers

Sourav

So what I would start with is to follow the path taken by many authors and remove the rare species. However, you will need to work through the literature and figure out why they have taken this approach and make sure that it applies to your data.

I work with insects. My first thought in reading this is that most of the species in an ecosystem are rare. So if I sample macrolepidoptera for several years I might collect 500 species. 80% of the individuals are represented by 10 species. Well over 100 species are represented by 1 to 5 captures. With this in mind, we have to also realize that biodiversity estimates are greatly influenced by the shape of the tails in the distribution. So what one does by eliminating the rare individuals is to delete most of the information on the shape of the tail. So your biodiversity estimates will all be compromised by this method. One could argue that you are penalizing all sites equally, but the counter argument is that this assumes that all sites have the same proportion of rare individuals -- or that rarity at each site has the same causes. Conceptually we can all list the causes of rarity, but it is much harder to argue that the relative contribution of each cause is the same for each site.

1) was the sampling effort the same at all sites? I will hope so.

Maybe step 1 should be to understand how biodiversity is calculated. Anne E. Magurran has written a couple of good books on the topic. EstimateS is a good program for calculating many of the biodiversity estimates (http://viceroy.eeb.uconn.edu/estimates/), though it is not the only option. With a biodiversity estimate at each site, use principle component analysis to identify important environmental variables. Does your answer change depending on which biodiversity estimate you use? In this approach, I don't see why anyone would discard rare species data. Thus we end up back to a recommendation of reading the literature and understanding why other authors discard these data. I could see ignoring rare species if you are actually interested in abundance and not biodiversity.

Your options will be greatly constrained by the sampling effort. Multiple years of data? Multiple samples per study site? Multiple times each year?

Hi Sourav,

In Legendre's Numerical Ecology book it is suggested using Hellinger transformation for community data with many zeros for perfoming PCA, CCA, etc. But I am not sure if it is also suitable to GLMM.

All the best,

Vinícius

Sourav,

Always a difficult choice.  The answer partially depends on how many species you are considering.  If your species list has ~300 species, then rare species would be difficult to deal.  The usual methods for community analysis tend to be driven by trends in the dominants, although transformations can strengthen the influence of less common taxa.

It sounds like you have data for a repeated measures/longitudinal analysis so you might check out those methods.  Using a mixed model might be valid with such a design for summary statistics.  CCA might also be a good choice, depending on the strength of the assumption violations.  Lots of opinions about the various methods but either CCA or MDS should work well to uncover community-level patterns combined with detailed analysis of other species.  

You should check out the work by Dunstan et al. 2013. Finite mixture of regression modeling for high-dimensional count and biomass data in ecology.  Jour of Ag, Biol, and Env Statistics.  

There is another similar paper but I couldn't find it in my piles of papers.

You might also be interested in these papers.

Guisan et al. 2002 Generalized linear and generalized additive models in studies of species distributions: setting the scene. Ecological modeling 

Guisan et al. 1999 GLM versus CCA spatial modeling of plant species distribution. Plant Ecology

Thanks guys, much appreciated for your inputs. I feel much guided.

I'm personally not in favor of leaving any data points even if comes from rare species. However as I explained first I'm very new to community ecology and often coming across papers which have done so. For my case the biodiversity is low < 20 species. However the trouble is there are copepods counts even > 20000 compared to crab count of 1-2 from a sampling site. The sampling design is very even, well replicated and capture well spatial dimension and is for 21 month continuous sampling. Frequency monthly.  

cheers

Sourav

I also do not suggest to omit the data point with zero, which I believe could compromise ecological explanations.

Rather, I would suggest to transform your data set (using third to fifth root transformation). Afterwards you can try component analyses which could clearly give the variation between sites (you can apply parametric or nonparametric approach). Please check the percent variability between the first and second axis or third axis if you omit the first one. checking the loadings could also help to identify the main causes of variation.

Dear Paul, 

The package glmmADMB could be suitable to you analyses involving richness and abundance (count data) ~ environmental variables. The package includes zero-inflated models, which seems to be appropriate to your data.

It could be interesting to think about spatial or temporal pseudoreplication in your study design. If so, you can add site or 'date' as random variables.

Please, take a look at zuur et al 2009 - Mixed effects models and extensions in Ecology with R.

Instead of exploring your data via nMDS, CA or PCA you can check covariance matrices and then fall into glmms. 

I hope my comments are useful.

Best whishes,

Thanks Davi,

Yes, I'm exploring more of GLMM routes. Your insights will definitely help me.

Sourav