Monday, January 21, 2013

Synchrony in metacommunities

A new year, a new publication, this time with Shubha and Jurek: "Population synchrony decreases with richness and increases with environmental fluctuations in an experimental metacommunity" in Oecologia. We continued our work with specialization in metacommunities, but this time looked at the implications on population synchrony.

Key figure:
We found, as predicted, that:

  • the synchrony between populations of a specialist species within a metacommunity is more influenced by environmental fluctuations compared to a generalist species
  • that increasing species richness decreased individual population synchrony
While these results make perfect sense from an ecological point of view, getting this paper published was not so straightforward because of the unbalanced experimental design we had to use. This struggle is transcribed in the boring science language at the end of the M&M that completely glosses over the blood, sweat, and tears we experienced in our back-and-forth with reviewers:
"We performed an unbalanced factorial analysis using the 15 effective numbers of dimensions, since replication of the two-species treatment in our experiment resulted in enough error degrees of freedom to test for an interaction effect between environmental fluctuation (EF) and species richness. We decided to use this unbalanced but replicated design because of time constraints. We were unable to implement a completely balanced design due to sample processing demands, which, for the sake of consistency, meant that all of the samples had to be counted within a short time frame."
 Because we were interested in synchrony in a metacommunity, we (i.e., Shubha) had to count densities of 3 species within all 4 communities of each metacommunity, or approximately 96 x 3 in total. All this in a limited time, and the individuals had to be put back to the original community to not affect population dynamics. This meant that we only had 2 replicated metacommunities (the experimental unit) for the 2 species treatment, and for all other treatments only 1 metacommunity.

We thus had to be really careful in how we worded the Figure 2 caption (see above, last line), since we could in principle only compute a measure of spread (e.g., standard error or deviation) for the 2 species points. However, that was also based on only 2 replicates. We thus decided not to add these, because it would be even more confusing to have some points with and some without some measure of variation indicated in the figure.

This could, of course, raise the question whether not adding this information to the main figure of our article is not only removing potential confusion, but also removes visual clues to readers about the limitations of our study, i.e., the lack of replication for most of the treatments.

I do believe we are only guilty of the first (removing confusion), since to test the interaction term EF x S, we only had 3 degrees of freedom (see Table 1). Despite this low power, we fought hard to get these results in, because when looking at the Figure 2, I think nobody would question that the lines that connect EF treatments with the same species richness are basically parallel to each other, and thus provides visual confirmation of the non-significant interaction term associated with the low degrees of freedom due to lack of full-factorial design.

This is thus one of my publications that for a large part hinge on the visual interpretation of the results (Figure 2, above) and the statistics (and all the other figures and tables) only had a confirmational role. I was really glad I had read Analysis of messy data, to convince the reviewers.

Thursday, January 10, 2013

Variation decomposition is a zombie idea?

Another post as a response to something written by Jeremy Fox! I think it becomes time that I meet him in person so that I can address him by “Jeremy” instead of “Dr. Fox”, “Fox”, or “the author”. I tried to remove all the salesmanship from my response, though, because I wanted to make sure that 1) I summarized his blog post correctly, and 2) expressed my thoughts as clearly as possible. This will make it easier for others to point out my logical mistakes, and thus to correct my thinking, and my understanding of ecology.

Original argument by Dr. Fox

  • correlation does not equal causation
  • e.g. abundance in species over time
    • data
      • 2 species, abundances not correlated
      • species abundances correlated with weather
    • naive conclusion
      • no density dependence
      • weather important
    • Dr. Fox’s conclusion: this is wrong
      • correlation is not prediction of density dependence (example from economics)
  • implication
    • zombie idea: “misinterpreting correlations, and lack of correlations, among variables as evidence for or against causality when those variables are affected by density-dependent feedbacks”
    • in community ecology: “Given that there’s intra- and interspecific density dependence within sites, I doubt that you can reliably infer the causes of metacommunity structure just by looking for statistical associations between environmental and geographic variables, and species abundances.”

My initial response

  • after reading that last quote, I thought that Dr. Fox made this conclusion:
    • so “correlations, partial correlations, variance partitioning, multiple regression, structural equation modeling, or related statistical methods” become useless in ecology, since this is probably a density-dependent dynamical system
    • because of this zombie idea dismissing general statistical techniques as tools to “infer how causality works in a density-dependent dynamical system” seems a bridge too far. I think that this could become a zombie idea in itself if lots of ecologists would follow this advice.
  • but I doubt that this is the intent of the blog post. The key word being “reliably”; he probably added this for people like myself to the sentence. Without it, it would seem that Dr. Fox is negating the scientific method; with this word, you can talk about observational versus experimental approaches, weak and strong tests of certain hypotheses etc.
  • I thus read the whole blog post again in more detail, followed by coming up with a second response.

My second response

  • I do think that Dr. Fox pulls a sleight of hand:
    • Dr. Fox’s conclusion and example only deals with pointing out that correlation is not a good prediction for density dependence
    • but the data do provide evidence for the hypothesis that weather influences population abundances through food availability etc.
  • so my conclusion for the data from Dr. Fox outlined above
    • there is evidence that weather is important
    • not a good test for whether density dependence is important or not, so not possible to make any conclusion on this aspect
  • implications of this for variation decomposition as a method to detect metacommunity processes
    • any statistical method is only as good as the prediction and associated hypothesis it is testing
      • not good to test for density dependence
      • could be for testing e.g. effects of environment on species abundances
    • statistical methods have their limitations (see indeed Gilbert and Bennett 2010, as pointed out by Dr. Fox)
      • thus observational data should be accompanied by experimental studies
      • see e.g. my PhD work (from a long time ago, observational study using variation decomposition, and experimental study)
    • my 2005 publication should thus only be seen as a starting point
      • there is lots of evidence for species sorting in response to environmental condition in general in these potentially dynamic systems (in 73% of the data sets)
      • it is difficult, but not impossible, to include for instance competition or density dependence or time or evolution into these analyses.
      • we should explore more interesting or detailed predictions, and some of these could use variation decomposition (e.g. generalist versus specialist species or body size)
  • The real problem would be that the zombie idea of testing density dependence with correlation could somehow make the correlations between species abundances and weather invalid.
    • I maybe missed this explanation in the blog post, but I could not find how density dependence could make not significant external drivers significant (but if I am wrong, let me know)
    • maybe this is addressed in Ziebarth et al. 2010, and is implicitly suggested by Dr. Fox?

My conclusions

  • Dr. Fox convincingly illustrates that correlation is not a good prediction for the hypothesis of density dependence, and there exist better predictions and tests.
  • My 2005 article is flawed, but I do not think that this zombie idea of density dependence is the most important one (we are working on a follow article, though).
  • Writing blog posts are fun, especially if you have a skilled writer as Dr. Fox giving you starting points. I just have to take care that when somebody writes to get a reaction (which is part of the appeal of Fox’s writing, and which he suggests in this and this post), that I read and react to the actual points made, and try to “read around” the salesmanship aspect of the writing.