Today, I went to listen to Andrew Gelman’s views on the philosophy of Bayesian statistics and this gave me a good opportunity for a 22k bike ride!, as the talk took place in the south-eastern part of the city. (I had not been yet to the new campus of Université Paris Diderot called Paris Rive Gauche. It is brand new, in a renovated district around the Grands Moulins de Paris. The place is buzzing with construction work and the Rue Watt I wanted to visit for its association with Léo Mallet is surrounded by cranes and engines.)
Back to philosophy: Andrew unsurprisingly stated he was not one for conventional philosophical perspectives! He thus went on to demonstrate that Bayesian statistics was not an inductive method but truly an hypothetico-deductive meccanism in the right line of Popper and Lakatos. The main criticism about conventional Bayesian thinking was that Bayesian model choice, by using a discrete collection ot models is inappropriate: on the one hand, models (including priors) can be criticised from the inside. On the other hand, a continuous collective is preferable to the standard model averaging found in Bayesian statistics. Obviously, I do not agree with the ideas that you can test your prior based on the data nor with the fact that the requirement of Bayesian testing on alternatives is a drawback [as we also argued in the Molecular Ecology disputing paper]. But, thanks to all its provocative aspects, this was an enjoyable talk and I think that thru it I understood a bit better Popper’s opposition to induction…
Xi'an's Og 