Monday, January 18, 2016

Have humans altered community interactions?

A recent Nature paper argues that there is evidence for human impacts on communities starting at least six thousand years ago, which altered the interactions that structure communities. “Holocene shifts in the assembly of plant and animal communities implicate human impacts” from Lyons et al. (2016, Nature) analyses data spanning modern communities through to 300 million year old fossils, to measure how the co-occurrence structure of communities has changed. The analyses look at the co-occurrence of pairs of species, and identifies those that are are significantly more likely ('aggregation') or less likely ('segregation') than a null expectation. Once the authors identified the species pairs with non-random co-occurrences, they calculated the proportion of these that were aggregated (i.e. y-axis on Figure 1). Compared to the ancient past, the authors suggest that modern species had fewer aggregated species pairs than in the past, perhaps reflecting an increase in negative interactions or distinct habitat preferences. 
Main figure from Lyons et al. 2016.
The interpretation offered by the paper is “[o]ur results suggest that assemblage co-occurrence patterns remained relatively consistent for 300 Myr but have changed over the Holocene as the impact of humans has dramatically increased.” and "...that the rules governing the assembly of communities have recently been changed by human activity". 

There are many important and timely issues related to this – changing processes in natural systems, lasting human effects, the need to use all available data from across scales, the value of cross-disciplinary collaboration. But, in my view, the paper ignores a number of the assumptions and considerations that are essential to community ecology. There are a number of statistical issues that others have pointed out (e.g. temporal autocorrelation, use of loess regression, null model questions), but a few in particular are things I was warned about in graduate courses. Such as the peril of proportions as response data (Jackson 1997), and the collapsing of huge amounts of data into an analysis of a summary of the data ("the proportion of significant pairwise associations that are aggregated"). Beyond the potential issues with calculating correct error terms, interpretation is made much more difficult for the reader. 

Most importantly, in my view, the Nature paper commits the sin of ignoring the essential role of scale in community ecology. A good amount of time and writing has been spent reconciling issues of spatial and temporal scale in ecology. These concepts are essential even to the definition of a 'community'. And yet, scale is barely an afterthought for these analyses.  (Sorry, perhaps that's a bit over-dramatic....) Fossils—undeniably an incomplete and biased sample of the an assemblage—can't be described to more than a very broad spatial and temporal scale. E.g. a 2 million year old fossil and a 2.1 million year old fossil may or may not have interacted, habitats may have varied between those times, and populations of S1 and S2 may well have differed greatly over a few thousand years. Compare this to modern data, which represents species occurring at the exact same time and in relatively small areas. The differences in scale is huge, and so these data are not directly comparable.

Furthermore, because we know that scale matters, we might predict that co-occurrences should increase at larger spatial grains (you include more habitat, so more species with the same broad requirements will be routinely found in a large area). But the authors reported that they found no significant relationship between dataset scale and the degree of aggregation observed (their Figure 2, not replicated here): this might suggest the methodology or analyses needs further consideration. Co-occurrence data is also, unfortunately, a fairly weak source of inference for questions about community assembly, without other data. So while the questions remain fascinating to me - is community assembly changing fundamentally over time? is that a frequent occurrence or driven by humans? what did paleo-communities look like? - I think that the appropriate data and analyses to answer these questions are not so easy to find and apply.

Response from Brian McGill:
My comment I was trying to post was:

Interesting perspective Caroline! As a coauthor, I of course am bound to disagree. I'll keep it short, but 3 thoughts:

1) The authors definitely agonized over potential confounding effects. Indeed spent over a year on it. In my experience paleoecologists default to assuming everything is an artefact in their data until they can convince themselves otherwise, much more than neo-ecologists do.
2) They did analyze the effects of scale (both space and time) and found it didn't seem to have much effect at all on the variable of interest (% aggregations). You interpret this as "this might suggest the methodology or analyses needs further consideration". But to me, I hardly think we know enough about scaling of species interactions to reject empirical data when it contradicts our very limited theoretical knowledge (speculation might be a better word) of how species interactions scale.
3) To me (and I think most of the coauthors) by far the most convincing point is that the pattern (a transition around 8000 years ago plus or minus after 300,000,000 years of constancy) occurs WITHIN the two datasets that span it (pollen of North America and mammal bones of North America both span about 20,000 years ago to modern times) and they have consistent taphonomies, sampling methods, etc and yet both show the transition.

I agree that better data without these issues is difficult (impossible?) to find. The question is what you do with that. Do you not anwwer certain questions. Or do you do the best you can and put it it out for public assessment. Obviously I side with the latter.

Thanks for the provoking commentary.




Jon said...

Looking at their Extended Data Table, the issue of scale seems to be completely mitigated in the pollen-based subset of their data. The spatial and temporal scales are identical for the 4200+ pollen localities spanning 21,000 years (these are almost certainly core data, hence the detailed data on spatio-temporal resolution). Pollen is essentially indestructable, so it's basically a gold-standard since the issues of biased preservation don't really apply to it.

I agree that co-occurrence data is far from optimal, but the marked departure in co-occurrence patterns from the identically-scaled pollen data is definitely interesting. The fact that the same general shift is observed in the other datasets (which do have all of the problems you described) still seems impressive.

There is a lot more spread in the aggregation/segregation in the oldest pollen data, but those are also based on *way* smaller data sets (<100 sites per 1kyrs before 13Ka vs >200 sites per 1kyrs for 12Ka and younger).

Jon said...

Looking at their Figure 2, it strikes me that if the pollen data are removed, then there is (at least visually) a negative relationship between the temporal extent/grain and proportion aggregated. The pollen data definitely drive the shape of the line from 10Ka to the Recent. I wish they'd just plotted the pollen data separately on a linear axis, since it's sampled so evenly through time...

Caroline Tucker said...

Hi Jon - thanks for your thoughts. I wondered about the pollen data as possibly a more consistent source of data (although I didn't know whether biased preservation or difference in dispersal of different species were issues). It would have been great if they had provided a separate plot for the pollen data.

I think the Extended Data Figure 3 still makes me uncertain how strong the relationships the authors found were, since once the modern data are removed from the loess regression, the resulting plot shows a variable trend rather than a decreasing one. But again, perhaps the relationship for pollen only would be convincing - I just can't tell looking at the plot...

Jessica said...

I'll note that the pollen data are the subject of a separate paper entirely! We didn't address the issue of timing of the shift from mostly aggregated to mostly segregated pairs in that paper, but rather tried to attribute particular significant species pairs to being caused by differences in environmental niches, potentially dispersal limitation, or, if none of those seemed to be a factor, a potential species/taxa interaction. I tend to take the extreme shifts in % aggregation prior to 15 kyr BP with a grain of salt because there are fewer sites during those focusing on data from 14 kyr BP to the present you can see the slow but steady shift from dominance by aggregated pairs to dominance by segregated pairs.

Here's the link to that paper on Wiley, though if anyone needs a reprint they can email me at jblois ucmerced edu:

Caroline Tucker said...

Thanks Jessica - I will definitely read the pollen paper! I appreciate the discussion you and Brian have contributed here - lots to think about.