Here is my post about Tuesday at ESA 2012.
My first stop was the session run by Ben Bolker (one of the people I enjoyed meeting and chatting with on Monday) and Drew Tyre (another case of meeting up with an old friend; mmm… feeling old). The session was based around Leo Breiman’s (2001) comparison of two cultures in statistical modelling. One is based on analyzing simple models that aim for understanding. The other is based on complicated models that aim for predictive performance.
The talks in this session largely focused on analyzing large complicated datasets, but at the same time using those analyses to try to find understanding. The talks were focused on species distribution modelling, although the ideas and methods could be applied more widely. I think the one thing missing (or at least under-represented) was a clear link between explicit theory and the data analyses.
That said, the talks and analyses were impressive. I won’t do them justice by trying to summarize them here, but I will comment on two snippets that impressed me.
The first was Daniel Fink’s analyses of the distribution of bird species in the lower 48 states of the USA. One critique of correlative species distribution models is that different factors are likely to influence distributions in different parts of a species range. However, standard species distribution models essentially assume the same relationships hold everywhere, while this might not be true. For example, water availability might limit the range of species closer to the equator, while food availability or cold tolerance might limit the range of a species at its poleward extreme. See some of Mike Kearney’s work on modelling the physiological limits of species for examples of this.
The nice approach of Daniel Fink was that individual models were fit for different overlapping regions (defined by space and time), and then a prediction at a point was made by averaging over the predictions for regions that intersect with that point. The result was a species distribution model that varied in both space and time, illustrated by impressive dynamic maps of migrating passerines. If you are interested in species distribution modelling, this is something to keep an eye on.
The other point of note was an aside by Tom Dietterich. He described how in one of his classes, his students simulate some data with a model (say with multiple predictors and non-linear terms). Then, the students fit a range of models to those data and compare their predictive performance against an independent set of data. Which model makes the best predictions? One with the same functional form as the true model used to simulate the data, but with estimated parameters? Or a simpler model? Yes, you were right if you answered the latter. The simpler model made better predictions (unless of course there is enough data to reconstruct the true model almost perfectly). Ooo, I like simple models. Which is not to say I dislike complicated models.
I guess a third point of interest was the impressive collection of speakers.
Next I headed off to the workshop on “Halting attrition: mentoring and retention of women and minority students”. We had four panelists (Meg Lowman, Adelaide Johnson, Carmen Cid, and Brian Wee; I just edited the names after finding my notes), who spoke about their experiences and advice for overcoming discrimination on the basis of race and gender. There were some great stories (some were personal, so I’m not going to blog them without permission). Just as the discussion among the attendees was getting started, I am sad that I had to leave for a meeting. However, a report of the workshop is going to be written and published, so I look forward to that.
One observation: it seems to me that action to overcome discrimination is required at many levels, but particularly by institutions and individuals. The session focused a little on what people who are targets of discrimination could do, and also what institutions might do (or not do). But I had hoped that we would also discuss what individuals who are not targets of discrimination might be able to do. Perhaps that was discussed after I left. I think we need more contributions from non-targeted individuals, but I’m not 100% what those contributions should be. I am hoping for advice, but didn’t get a chance to ask.
After my meeting, I saw Laura Pollock’s talk on her PhD research about how traits and phylogenetic relatedness influence co-occurrence of eucalypts. Species with similar tree height and SLA co-occurred (reflecting a similar response to the environment), yet species with dissimilar relatedness were more likely to co-occur (perhaps reflecting the inbreeding of closely related eucalypts). For more on Laura’s research, see her website.
Next up, I moved to the modelling session. My favourite talk there was by Oedekoven on Bayesian distance sampling. The talk illustrated the approach with an experimental manipulation of plantings in field edges to enhance densities of indigo bunting. Now, they are fine looking birds! But, getting back to the point, an advantage of a Bayesian approach is that random effects, for example, can be incorporated easily into the analysis. However, it seems that Bayesian methods are used rarely in distance sampling.
I missed lots of talks that I would have liked to have seen. Check out some others via the posts of other ESA2012 bloggers: here, here and here. There are probably other posts – please add links in the comments.
But I did get to talk with Ilona Naujokaitis-Lewis about her model of hooded warbler dynamics in response to climate change. A nice aspect her model is that she used measurable relationships between observed demographic rates and climate to parameterize the model. This sets it apart from lots of models that link climate and population dynamics, where the link is parameterized via a relationship between climate and distribution. Of course, distribution is a function of many aspects of population dynamics, and the relationships are simply correlative. In Ilona’s model, the link between demographic rates and climate is explicit, which makes it particularly powerful.
The summary: another good day.