Skip to content

Evidence and Evolution

I have just finished doing a review (for Systematic Biology) that took me six months. It was not because I was slack. It was because I had to read it in three page segments. The book is


“Evidence and Evolution: The Logic Behind the Science” (Elliott Sober)

I won’t give the full review here. Suffice it to say if you go to it looking for the evidence for evolution, you will be sadly disappointed. This is an essay in epistemology, covering, among other things, a debate over likelihoodism in model selection (109 pages), intelligent design (80pp), natural selection (75pp), common ancestry (67pp) and a conclusion (15pp).

I think that the likelihoodism approach to theories and models is not quite right, and that Bayes is Better, but despite that, it’s a detailed book on argument and logic in science. If you agree with it, then it will be a source of much richness. If not, then the intelligent design chapter is still of interest. The natural selection chapter reprises and extends the arguments in his 1984 book The Nature of Selection: Evolutionary Theory in Philosophical Focus, and the common ancestry book does the same with his 1988 book Reconstructing the Past: Parsimony, Evolution, and Inference. The conclusion chapter can profitably be read by itself.

One point I do disagree with fairly strongly is his treatment of common ancestry as being based on similarities. Although he gestures in the direction of homology and homoplasy and the difference between them, he appears to think they are the same kind of relation – similarity – that differ solely in our knowledge of the evolutionary process. We know that homoplasy is based on natural selection causing convergent evolution, and so we can identify, maybe, when it is the similarity-causing process.

But homology is a relation of identity (under all forms and functions, to paraphrase Owen), and there can in fact be a lack of salient similarity between homologs (the human hand and the digits of bird forelimbs, for example), while mere similarity depends crucially on what you take as the metric. At is best, homoplasy is a signal for selection; at its worst it is subjective choice of criteria. Either way it must be eliminated for phylogeny, which Sober doesn’t seem to recognise clearly.

This is a book for the philosophically inclined, and not all of them. I must declare a conflict of interest here: Sober will be visiting my department next year, and I am his “sponsor” (read: I did the paperwork). I declare this not because it makes it easier for me to agree, but the reverse. I am going to have to defend my point of view when he comes by, damnit. But it is the hardest book I have read since Kierkegaard (a comparison I think he might not like, but it’s for entirely different reasons of difficulty, I assure you all).

2 Comments

  1. Chris Stephens Chris Stephens

    I look forward to reading your full review.

    One question about your comments here, though: I guess I read chapter 1 of E and E as not so much a defense of Likelihoodism, but rather has a defense of pluralism – that none of the three major approaches to statistical inference and evidence (Likelihoodism, Bayesianism or Frequentism) has got it all right, and that this is a bit of a change from some of his earlier work, where he more clearly favoured Likelihoodism (this is in part a result of his work on model-selection, which he discovered after working on his books N of S and RTP). Likelihoods such as Royall and Edwards never (according to Sober) properly appreciate or got into Akaike’s model selection stuff.

    But if your point is that he’s (still) not a full fledged Bayesian, then that is of course right. But I think its also clear that he thinks model selection (which he classifies as a kind of frequentism) has something important epistemically to offer beyond the likelihood approach.

    C

    • John Wilkins John Wilkins

      In the large review I make the same point, sort of, but the fact is Sober uses likelihoodism throughout the book, as he is focussing on hypothesis testing (in a very broad manner, if you can call intelligent design a hypothesis; Sarkar correctly notes there’s no “there” there).

      Part of the problem I seem to have here lies in the assumption that knowledge is a matter of quantifiable ratios or statistical measures. A Bayesian doesn’t have quite that trouble because the priors are relative at best. But it’s not my field so I cannot be sensible about it.

Comments are closed.

Optimized by Optimole