Realism in species distribution models

Buckley LB, Urban MC, Angilletta MJ, Crozier LG, Rissler LJ, & Sears MW (2010). Can mechanism inform species' distribution models? Ecology letters PMID: 20482574

Predicting species distributions is not easy. Current approaches can be broken into two broad categories: "correlative" or "mechanistic" models. Buckley et al. (2010) do something very unique by comparing the relative accuracy of these two approaches (a total of 5 models) for two species (a butterfly and a lizard, see image). Their findings are interesting and very informative, but their conclusions lack some potential insight and they miss some important opportunities to advance our ability to predict.

To begin, what do "correlative" and "mechanistic" mean? Correlative models are based on current species distribution and environmental conditions. Mechanistic models are based on organismal responses to the environment, and can use factors like energy gain (model #3 below...), reproduction/survival based on climate observations (#4), or reproduction/survival calculated by energy gain (#5).

THEIR MODELS
Correlative - require species location and env. data:
1. Maximum Entropy (maxent)
2. Generalized Linear Models (GLM) 
Mechanistic:
3. Biophysical threshold - requires trait thresholds and env. heat variables.
4. Life history - requires demographic data and env. data.
5. Foraging energetic - requires trait data and env. heat variables.
(Nice work on the models, Buckley et al.!)

What was my prediction for the study? That the mechanistic models would outperform the correlative models because of increased "biological realism" (seemed reasonable). And, what did Buckley et al. (2010) find? The correlative models performed just as well as the mechanistic models for predicting current species distribution... In fact, my rough calculation from their results table has the correlative models correct 77% of the time, but the mechanistic models correct only 70% of the time! (bummer)

Why did that happen?! Buckley et al (2010) point out two reasons for why the mechanistic models might not do as well as I expected:
1. The constraints in the mechanistic model might not be the most important variables, and
2. parameter estimates for the mechanistic models might have errors.
I think that the first part is relatively easy to get handle on with more background knowledge for the species in question. The second part is a little bit more tricky, though. Buckley et al.'s (2010) solution for the second issue is to extend "sensitivity analyses" for parameter values, since we know that there is uncertainty in the parameters. There is a great (as yet unexplored) opportunity here to develop species distribution models that consider some of this uncertainty!

The bottom line:
My rough, correlative prediction for the current and future distributions of Acer saccharum (shown in green and transparent pink above, with observations as filled points) might not be as bad as I thought (data from GBIF and LifeMapper - using the BIOCLIM correlative model). But, it would be nice to incorporate some mechanism into it. This is especially important considering that the correlative models in Buckley et al. (2010) produced VERY DIFFERENT forecasts than the mechanistic models.

As Buckley et al. (2010) point out, we are on our way to making these models really useful. All that is needed is a better accounting of what mechanisms are important, and how uncertain we are about those mechanisms. (Easy, right?!)