Science cakes – catch-up post
I have been seriously negligent in keeping the website up to date, thanks in large part to a ridiculously busy semester of teaching. But Adam Hayward has shamed me into catching up on our science discussions by promptly writing up a summary of the last meeting (which I missed). So before I publish his summary, we have some catching up to do. The notes below are brief but hopefully give some of the flavour of the sessions this autumn, as well as some potentially useful links to the original science that we have discussed.
Let’s start with the first session of last semester. Expect me to fill in the subsequent sessions in the next few days as I sit with a warm drink and watch the snow fall from the in-laws home in Norway…
Oct 9, 2014
Attendees: Andy, Clare, Doctor Paine, Lilly, Luc, Matt, Robbie, Rosalind
Robbie began proceedings by introducing himself and presenting a stats problem related to a recent Nature paper by Wilson et al., 2014:
The authors evaluated the relative importance of human impacts as opposed to aspects of the chimp populations (such as the density of the population and the number of male chimps) when predicting the incidence of conspecific killings. They concluded that the best models did not include any predictor related to human impact, whereas the population parameters provided better models.
The authors then used model averaging to report the coefficients, and Robbie’s question concerned the suitability of model averaging for this kind of analysis. This stemmed in part from a comment he had received during a recent statistics workshop in which the instructor recommended averaging predictions rather than coefficients.
We discussed this for some time, and although there was broad agreement about the contours, there were still some differences of opinion at the end of our discussion. It seems to me that the question is partly about emphasis (are you more interested in hypothesis testing or in prediction?) and partly about the causes of an inability to distinguish between models (is there collinearity between predictors, and if so, what is the causal relationship between these predictors?). Andy pointed out that for most data sets, the difference in approach is probably trivial, and I agree. In fact, if the difference is not trivial, one might question why this is true and wish to think about the covariance structure of your dataset more carefully before using model averaging in the first place…
Andy provided the next contribution, stimulated by a paper that was given to him by his undergraduate tutees, by Amici et al., 2014.
The experiment in question tested for prosocial behaviour (the tendency to behave in a way that benefits another) in some monkeys and apes. For example, the subjects were given the choice of pulling a lever that gave food only to the lever puller, or to both the lever puller and his neighbor. The authors failed to find any evidence for prosociality (the subjects seemed indifferent to what was happening to neighbours so long as they were getting food themselves), having controlled for some potentially confounding variables that may have been the cause of inconsistent positive findings in earlier studies of the same phenomenon.
We spent some time wrestling with the implications of these negative results. Matt noted that testing these kinds of phenomena empirically is difficult, while Robbie wondered if the expectation of prosociality is prefaced on social structure which may or may not have been absent in the current study. Rosalind asked if a different version of the experiment involving related individuals would be more conclusive?
My own comment concerned whether the entire question is relevant – is there ever a free lunch in nature (i.e., is there ever a situation where what an animal does can help another at no cost at all to itself)?
Andy mentioned the court case involving the non-human personhood of an ape Does our provision of “human” rights to other apes depend on our assessment of their cognitive and altruistic tendencies? If our Science Cake session has just shown that chimps are murderous even without our intervention, and that apes and monkeys don’t give a damn about whether their neighbours get a biscuit, should that make us more or less inclined to attribute rights to other primates?
(Ed: I note that the court case has just recently been resolved as I write this, and the court did not consider the imbalance of evidence for chimpicide versus prosociality:
– does that excuse the long delay in transcribing Science Cake notes?)
We closed the first session of Science Cake with a discussion of some dance fly data that Rosalind presented involving associations between characteristics of male and female mating partners. This was mostly motivated by her lousy supervisor (me), who received her initial pitch about how to sell the work skeptically. We batted around some suggestions for a while, and while there was no firm resolution, I think Rosalind came away with some good ideas and left a promise that the data would resurface later for another discussion once she has a better sense of the sales pitch. Perhaps the winning bit of motivating was performed by Matt, and echoed by Tim (paraphrased here because my notes are imperfect and this was so long ago): “It’s not so much a question of what you do, but what you say, in other words, make sure that your analysis addresses a story that you can tell convincingly.”
To be continued….