You have data which is effectively a short panel. It has cross- sectional ( multiple communities) variation and time series (each country over 5 years). So you can look at between community variation at a point in time and within communities across time at a descriptive level. Then, after you have explored whether the differences are greater across time within communities or at a single point in time across communities, you could formally look at changes to a key variable using panel data methods (fixed or random effects models) to isolate the key factors associated with changes to your key variable of interest. Or, depending on your technical ability, you could pool all your data and estimate basic models with community and time dummies on the right hand side. An alternative strategy would be to use nested models if you have multiple data points within a community. Hope this helps? Marc 

Hi Marc, thank you very much for your answer; This is the kind of approach I was thinking about.

However, I am not mostly interested in knowing if within community differences in time are larger than among communities at each time point (although it should be included as you said as a first descriptive step), but really to depict the dynamic of the communities over time and if the different communities covary. Panel data methods might be one option, but how to incorporate temporal dependence among data points?

Hi Valerie. I guess there are two ways forward. Firstly, if I knew more about your actual data, particularly your dependent and explanatory variables I could be more precise about what models were worth exploring. If you use Stata then I could write you some basic coding to run some models. Or, depending on the sensitivity of your data, you could send it through and I could have a look. Or I could send you some useful practical notes for analysing panel data in Stata . It's up to you. Broadly, if I had the data I can usually sort it out within a few hours . But you choose what's best for you. Cheers, Marc 

The short answer is simply to use a lag operator for your dependent variable and others. Your panel is short so you could not fully explore these features but it might tell you something. Or simply to take the 5 year annualised mean and look at deviations from this mean. This addresses mean reversion, the tendency of variables to return to their long-run trend level which is based on core underlying fundamentals. 

Hi Marc. Thank you very much for your willigness to help me. I am using R and I am unfortunately totally ignorant of Stata. 

I cannot send you any data, because I do not have it myself yet but I try to explain you a little more in details what it is about. It is a quite complex survey from the ecological field. Species from many different groups have been collected in different cropping systems each year for five years and several environmental variables have been collected at the same time (temperature, moisture, aso).

The aim of this study is to assess whereas the dynamics of the sampled species communities is different in the different cropping systems, i.e. what is the impact of the cropping system on the dynamics (and ideally interactions) of species communities.

I think looking at deviations might be a good way to start. Lag analysis is appealing, but as you say it will quickly appear limited.

You may read:

http://rspb.royalsocietypublishing.org/content/281/1778/20132728

http://www.researchgate.net/publication/236964847_Diversity_in_time_and_space_wanted_dead_and_alive/file/6a85e52d8fa4c24e92.pdf

Article Diversity in time and space: Wanted dead and alive

Thank you Thiago, very interesting reading. The papers deal mostly with larger temporal scales but I may pick up some ideas.

How about this attached file from Marine Ecology Progress Series 467:147-166. This nonlinear regression approach will take care of seasonal trends. So you could run the regression, get the best model, and then develop more traditional models using the residuals rather than the raw data. I would try comparing the results from this answer and some of the other suggested techniques. If they all give you the same answer it improves you confidence. If they give you a different answer then you have to work more at figuring out why one model is correct and the other less correct.

Hi Valérie (and all), I interpreted the original question a little differently, i.e. 'is it possible to make a statement about the dynamics of communities if you only have a few years worth of data'. In part this is a question about the techniques / technology available, but I think you might also want to look at time scales.

I mostly work in arid systems were annual variation in vegetation composition is dominated by species responding to rainfall – in a good year the species is/are there, in a bad year they are not. In such cases it is often useful to look at the longer lived species.

In your case you might have to look at the longevity of the ...longer lived species to determine if your time-frame is “short” or not.

You might thus consider looking at different subsets of your data, either including or excluding some annual species, and focusing on the perennials. This may permit you to identify drivers that affect your vegetation over the short or long term.

Hi Friedrich, thank you for this answer. You reformulated my question very good. It would be an interesting approach to look at a subset of perennial species, the problem is that I am not working with plants but with insects and thus expect higher turnover in species composition and possibly compare this turnover across space and time.

Hi Valerie,

:-) it seems that I misunderstood the topic - sorry about that. You might find a equivalent approach for insects? I assume that the variation in species composition within a single year would require more frequent sampling - spatially and temporally.

Is there a discussion of r-selection vs. K-selection - in relative terms - when only insects are considered, i.e. those that have a faster response to environmental conditions vs those that don't (predatory insects vs. herbivorous insects, body size?)?

Alternatively, would you link insect composition to plant composition (back to considering plant communities...)?

I can only ask questions as this is not my field. ... and sorry about the long sentences :-)

I think it will be difficult to look at selection strategies over such a short time scale. Unfortunately I have only one sampling occasion per year and thus no possibility to analyze within vs among year dynamics. Also the r- vs K-strategies might be more useful for larger-size animals than for insects; but it is a very good investigation point.

Insect occurrence is intimately linked to plant composition of course, but my data are coming from crop systems, thus mainly with only the main crop species.

Complications include:

1) When did each species colonize each crop island?

2) What separates each crop and how far apart are they? Are the data from all of France, or three 5m x 15m field plots of each crop per year separated by 5m grass buffers? If the data are from all of France, that would be good.

3) Are correlations between species abundance cause-and-effect, or are they artifacts of the agricultural environment?

4) I would separate generalists from specialists. You might get more information from specialists because people are more likely to believe a cause-effect relationship. Part of this is also separate herbivore/omnivore/predator/parasitoid. Even in cropping systems, Friedrich's point about ecological niche is valid. If an insect depends on a weed, then you need to know when that weed colonized your crop in sufficient abundance to result in an insect being captured.

5) Predator-prey cycles often have a lag, and in insects this lag is temperature dependent. However, all you have is a single snapshot of species abundance for each year. Calculating growing degree days will simply fragment your data.

How about this: For each year calculate a correlation between all pairs of species. Non-significant correlations in any year have a correlation coefficient of zero (you will need to enter the data in as a zero). Take these correlation coefficients by year and calculate a mean across years. Test if this mean is zero. For non-zero means, take the by-year data and test if there is a difference in correlation coefficient between crops.

I know this is not the fancy ecological approach that you were looking for, but I think it is the best that your data will support unless there is a significant geospatial component to your data. The one sample per year is a difficult problem to overcome.

Thanks Timothy for this thoughtful answer. The data are coming for a single region, because we wanted to reduce to a maximum the variation due to the geographical location of the fields. Of course if we would have the opportunity to gather data from such a large number of places that this variation could be controlled for would be even better.

The idea of the study was to start from scratch, having in all sampling cropps the same original state. The cropping history of all sites was similar in order to "control" for differences in cropping history. Then, different cropping systems were implemented and data collected at the same time each year in all system. We expect the species communities to become increasingly different over time. However, how to disentangle all the possible causes is a real challenge.

We tried to address your important question in a more general context, in the attached paper. Cross-sectional studies typically involve few observations on a large number of individuals while in longitudinal studies few individuals are typically measured many times. A practical compromise between a longitudinal study and a cross-sectional study is an interval study. Interval plots are measured at least twice with the objective of assessing the rate of change of a target variable between successive re-measurements. 

Thanks Klaus, I will have a look.