Hoisted from the comments: Mathew Leibold and I on variance partitioning in metacommunity ecology

As part of my #ESA100 reflections, I commented that variance partitioning was dead as a way to infer the drivers of metacommunity structure. Mathew Leibold didn’t get what I was on about at all. Understandably–my remarks were brief and in retrospect not as clear as they should’ve been.* He was kind enough to take the time to comment at length on what he sees as the main problems with variance partitioning and how it’s currently applied, which gave me the chance to clarify my own views. I agree with Mathew on many points.

Wanted to highlight this exchange because I think it addresses an important issue in ecology. Understanding metacommunities is a really important job for community ecology, and right now variance partitioning is probably the single most popular tool for the job. It’s vitally important that we use that tool in effective ways, improve it if we can, and have a good understanding of what it can’t and can’t do. Mathew and I agree that:

  • There’s been excessive enthusiasm for using variance partitioning as a diagnostic test for metacommunity theories. There are too many possible “kinds” of metacommunities–far more than the small number of “paradigm” special cases on which existing theory focuses–for variance partitioning to be used as a diagnostic tool.
  • Insufficient attention has been given to how to interpret variance partitioning even if various statistical issues with it are addressed. (I’d add–don’t know if Mathew would agree–that this is a common problem in ecology. When a new statistical tool is developed, subsequent work tends to focus on identifying and resolving technical statistical issues with that tool. Which is fine, but tends to have the unfortunate side effect that equally or even more important non-statistical issues of how to interpret the tool tend to get neglected. Or worse, get mistaken for technical statistical issues.)
  • Variance partitioning remains a potentially useful statistical tool, and future work needs to focus on how to interpret and use that tool most effectively. (Not as a standalone diagnostic tool, but as one line of evidence among others, I’d say.)

I also think this exchange of comments was a nice example of how blogging can contribute to scientific discussion. So I wanted to highlight it for that reason as well.

*Mathew’s a friend and knows how my brain works. So if he can’t tell what the hell I’m on about, probably lots of other people couldn’t tell either. Which is my bad.

5 thoughts on “Hoisted from the comments: Mathew Leibold and I on variance partitioning in metacommunity ecology

  1. Hi Jeremy,

    I must say, I tend to agree with you that variation partitioning can be a very misleading tool when trying to ‘categorise’ metacommunities. As you wrote, the proportion of spatially- and environmentally explained variation is sometimes completely unrelated to the underlying mechanisms responsible for those patterns.

    But Mathew does have a point too. If used properly, in a way that specifically tries to identify spurious patterns and underlying biases, variation partitioning can give unique insights into ecological and evolutionary mechanisms.

    On this point, I’d like to shamelessly plug on of my own papers. Although it was a biogeography study (as opposed to metacommunity ecology), we used variation partitioning to decompose the variation of African vertebrate species’ ranges into spatial and environmental components. We then compared these components to those from simulated data and found that very strong environmental patterns (as much as 40% variation in the E|S component!) could arise even in the absence of environmental mechanisms.

    However, the cool part is that once we corrected for these general bias, we found that the explanatory power of spatial and environmental predictor variables were closely associated with the boundaries between biogeographical regions in Africa. To us, this suggests that the predominant underlying mechanisms differ amongst biogeographical regions. Of course, it is damn difficult to explain exactly *how* they differ, but that is a question for another day…

    Those who might be interested can read a brief summary of the study here:

    Or the full paper here:

  2. Let me get this straight. I’m a collateral victim of a reviewer who referred heavily to this post (and the ones upstream) as if it was any scientific fact. Cynical scientists would probably say that’s business as usual. However, I truly believe one should be careful not throwing the baby with the bathwater for three main reasons:
    1) Spatial and temporal variation among communities (expressed as distances or dissimilarities on a multidimensional manifold) is a generic measure of beta diversity. To suggest that variation partitioning is useless in this context implies that processes structuring (partitioning) beta diversity cannot be adequately modeled.
    2) Gilbert & Bennett 2010 and Smith & Lundholm 2010 questioned the ability of relatively simple linear models to capture complex non-linear ecological processes. The same could be alleged of most approaches routinely used to model species’ distributions. To suggest that community ecologists do a poor job at capturing processes is an ongoing debate and an unproductive argument.
    3) Whether we like it or not, a quick literature search reveals that variance partitioning approaches (e.g., RDA, PCoA, other constrained ordination methods and the like) are in fact gaining ground, especially in recent years. Perhaps should we take advantage of their growing popularity.


    • Let me get this straight: are you suggesting that we’re responsible for something that a reviewer said? As I’m sure you recognize, scientists form their professional judgments based on all sorts of inputs, many of which aren’t peer reviewed. They read the literature. They talk to colleagues. They attend talks. Etc. And yes, they read blog posts. I’m sorry if you disagree with a reviewer. But professional disagreements are a normal part of science. And that a reviewer’s views on the topic in question were shaped by reading our blog doesn’t make those views any less (or more) correct.

      As the post and associated discussion makes clear, what’s dead is variance partitioning as a diagnostic tool for inferring the processes driving metacommunity structure, for instance diagnosing the relative importance of “niche” vs. “neutral” processes. That wasn’t sufficiently clear in my original post on this topic, hence this follow up post in which I drew on Mathew’s comments in order to clarify my views. I’m sorry if a reviewer of your paper misinterpreted this post as saying something else. I completely agree that variance partitioning has its uses, such as for describing beta diversity. With respect, I never said or implied that beta diversity cannot be statistically described!

      Re: 2, I respectfully disagree. If a purported diagnostic tool doesn’t actually work, then it doesn’t work, full stop. Even if no other more effective diagnostic tool is available. Pointing this out does not make the perfect the enemy of the good, or the good enough. In fact, being honest about the limitations and flaws of existing diagnostic tools is the strongest possible motivation for development of better tools. Recognizing the flaws in existing diagnostic tools also can motivate a switch to other research approaches that don’t rely on statistical diagnostic tools–for instance, experimental approaches. See here for further discussion of this point: https://dynamicecology.wordpress.com/2012/10/24/no-i-dont-need-to-propose-an-alternative-approach-in-order-to-criticize-yours/. And I also disagree that community ecologists should stop worrying about their ability to diagnose process from pattern just because we’ve long been arguing about whether there’s any generally-applicable and effective way to do that. Some arguments are long-standing for good reason, and are too important to just stop having. See here for discussion of this point in a different but related context: https://dynamicecology.wordpress.com/2011/06/01/why-ecologists-should-refight-the-null-model-wars/

      Re: 3, I’m afraid I don’t follow. Are you suggesting that science is a popularity contest? Are you saying that popular approaches necessarily are correct? And by “take advantage of their growing popularity”, do you mean that scientists ought to jump on bandwagons (https://dynamicecology.wordpress.com/2011/09/14/bandwagons-in-ecology/)? Personally, I prefer to judge approaches on their merits rather than by their popularity. Those are two different things.

      • Dear Jeremy,

        To the question: are you suggesting that we’re responsible for something that a reviewer said? Of course not. It’s just an honest way of saying how I ended up here.

        I mentioned points 2 and 3 to place the discussion in a broader context. These are facts, not opinions. I’m neither a supporter nor a doubter of variance partitioning approaches,

        Otherwise, we do not disagree much.



Leave a Comment

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s